# ADVANCES OF NEUROIMAGING AND DATA ANALYSIS

EDITED BY : Jue Zhang, Brad Manor, Hongyu An and Xiaoying Wang PUBLISHED IN : Frontiers in Neurology

### Frontiers eBook Copyright Statement

The copyright in the text of individual articles in this eBook is the property of their respective authors or their respective institutions or funders. The copyright in graphics and images within each article may be subject to copyright of other parties. In both cases this is subject to a license granted to Frontiers. The compilation of articles constituting this eBook is the property of Frontiers.

Each article within this eBook, and the eBook itself, are published under the most recent version of the Creative Commons CC-BY licence. The version current at the date of publication of this eBook is CC-BY 4.0. If the CC-BY licence is updated, the licence granted by Frontiers is automatically updated to the new version.

When exercising any right under the CC-BY licence, Frontiers must be attributed as the original publisher of the article or eBook, as applicable.

Authors have the responsibility of ensuring that any graphics or other materials which are the property of others may be included in the CC-BY licence, but this should be checked before relying on the CC-BY licence to reproduce those materials. Any copyright notices relating to those materials must be complied with.

Copyright and source acknowledgement notices may not be removed and must be displayed in any copy, derivative work or partial copy which includes the elements in question.

All copyright, and all rights therein, are protected by national and international copyright laws. The above represents a summary only. For further information please read Frontiers' Conditions for Website Use and Copyright Statement, and the applicable CC-BY licence.

ISSN 1664-8714 ISBN 978-2-88963-750-8 DOI 10.3389/978-2-88963-750-8

### About Frontiers

Frontiers is more than just an open-access publisher of scholarly articles: it is a pioneering approach to the world of academia, radically improving the way scholarly research is managed. The grand vision of Frontiers is a world where all people have an equal opportunity to seek, share and generate knowledge. Frontiers provides immediate and permanent online open access to all its publications, but this alone is not enough to realize our grand goals.

### Frontiers Journal Series

The Frontiers Journal Series is a multi-tier and interdisciplinary set of open-access, online journals, promising a paradigm shift from the current review, selection and dissemination processes in academic publishing. All Frontiers journals are driven by researchers for researchers; therefore, they constitute a service to the scholarly community. At the same time, the Frontiers Journal Series operates on a revolutionary invention, the tiered publishing system, initially addressing specific communities of scholars, and gradually climbing up to broader public understanding, thus serving the interests of the lay society, too.

### Dedication to Quality

Each Frontiers article is a landmark of the highest quality, thanks to genuinely collaborative interactions between authors and review editors, who include some of the world's best academicians. Research must be certified by peers before entering a stream of knowledge that may eventually reach the public - and shape society; therefore, Frontiers only applies the most rigorous and unbiased reviews.

Frontiers revolutionizes research publishing by freely delivering the most outstanding research, evaluated with no bias from both the academic and social point of view. By applying the most advanced information technologies, Frontiers is catapulting scholarly publishing into a new generation.

### What are Frontiers Research Topics?

Frontiers Research Topics are very popular trademarks of the Frontiers Journals Series: they are collections of at least ten articles, all centered on a particular subject. With their unique mix of varied contributions from Original Research to Review Articles, Frontiers Research Topics unify the most influential researchers, the latest key findings and historical advances in a hot research area! Find out more on how to host your own Frontiers Research Topic or contribute to one as an author by contacting the Frontiers Editorial Office: researchtopics@frontiersin.org

# ADVANCES OF NEUROIMAGING AND DATA ANALYSIS

Topic Editors: Jue Zhang, Peking University, China Brad Manor, Institute for Aging Research, United States Hongyu An, Washington University in St. Louis, United States Xiaoying Wang, Peking University First Hospital, China

Citation: Zhang, J., Manor, B., An, H., Wang, X., eds. (2020). Advances of Neuroimaging and Data Analysis. Lausanne: Frontiers Media SA. doi: 10.3389/978-2-88963-750-8

# Table of Contents


Giulio Ruffini, David Ibañez, Marta Castellano, Laura Dubreuil-Vall, Aureli Soria-Frisch, Ron Postuma, Jean-François Gagnon and Jacques Montplaisir

*103 Lenticulostriate Arteries and Basal Ganglia Changes in Cerebral Autosomal Dominant Arteriopathy With Subcortical Infarcts and Leukoencephalopathy, a High-Field MRI Study*

Chen Ling, Xiaojing Fang, Qingle Kong, Yunchuang Sun, Bo Wang, Yan Zhuo, Jing An, Wei Zhang, Zhaoxia Wang, Zihao Zhang and Yun Yuan


Yue Wang, Huazheng Liang, Yu Luo, Yuan Zhou, Lingjing Jin, Shaoshi Wang and Yong Bi

*143 Ten Years of* BrainAGE *as a Neuroimaging Biomarker of Brain Aging: What Insights Have We Gained?*

Katja Franke and Christian Gaser

*169 Hemodynamic Surveillance of Unilateral Carotid Artery Stenting in Patients With or Without Contralateral Carotid Occlusion by TCD/TCCD in the Early Stage Following Procedure*

Ziguang Yan, Min Yang, Guochen Niu, Bihui Zhang, Xiaoqiang Tong, Hongjie Guo and Yinghua Zou

# Editorial: Advances of Neuroimaging and Data Analysis

Jue Zhang\*, Kun Chen, Di Wang, Fei Gao, Yijia Zheng and Mei Yang

*Academy for Advanced Interdisciplinary Studies, Peking University, Beijing, China*

Keywords: neuroimaging, neural mechanisms, fMRI, EEG, deep learning

**Editorial on the Research Topic**

### **Advances of Neuroimaging and Data Analysis**

Neuroimaging is a discipline that studies the structure and function of the nervous system by means of imaging technology, and where the images of the brain can be obtained in a non-invasive way. It explores a series of mechanisms such as cognition, information processing, and brain changes in the pathological state. In recent years, the neuroimaging has developed rapidly and become a powerful tool for medical research and diagnosis. With the increasing prevalence of neurological diseases, higher requirements have been put forward for neuroimaging technology and subsequent data analysis, and many advances have been made in this field.

Now, we will briefly summarize the cutting-edge progress in the theme of "neuroimaging and data analysis." A total of 16 papers have been published on this topic. They were presented from different countries, including China, USA, Germany, Italy, Brazil, and so on, and they involve novel neuroimaging technology, neuroimaging analysis, clinical diagnosis, and mechanism research. Accordingly, we divide these studies into three sub-topics.

The four papers in the first part of this special issue focused on the practice and development of cerebral hemodynamics in healthy elite athletes and patients, including the exploration of the clinical mechanisms and the discovery of new markers for clinical diagnosis and treatment. Bao et al. reported that by utilizing functional magnetic resonance (fMRI) technology, they realized that fatiguing aerobic exercise changed the cerebral blood supply in the brain and had no significant effect on the ability of the brain to extract oxygenation. Their study provides essential values for the evaluation of anaerobic exercise in sports science and clinics, suggesting that it is meant to establish the CBF and OEF as novel markers for physical and physiological function. Yan et al. assessed the cerebral hemodynamic variations, including bilateral middle cerebral artery (MCA) peak systolic velocity (PSV), pulsatility index (PI), and blood pressure (BP), in unilateral carotid artery stenosis patients with or without Contralateral Carotid Occlusion (CCO) in hours following carotid artery stenting (CAS) using transcranial doppler (TCD) and transcranial doppler color code (TCCD). In particular, they suggested that CCO was a factor of the increased blood flow velocity in ipsilateral MCA after unilateral CAS. Early identification of high-risk patients with transient ischemic attack (TIA) using imaging techniques is essential for administering the proper medications to treat or prevent TIA and the consequent stroke, which will improve the clinical diagnosis of TIA. Thus, Wang et al. explored the probability and related influencing factors of MR Hypoperfusion abnormalities in Chinese patients with transient ischemic attack and normal diffusion-weighted imaging (DWI) findings. Sheng et al. characterized the quantitative DTI-derived diffusion, and DSC-derived perfusion parameter changes underlying different Susceptibility-weighted imaging (SWI) signal intensities of multiple sclerosis (MS) lesions. Moreover, the creatively idea of the work was that the signal intensities detected on SWI in MS lesions might be a non-invasive biomarker that represented a specific stage of lesion evolution or a particular pathological substrate associated with iron deposition, demyelination/axonal injury, or inflammatory activity.

Edited and reviewed by: *Jan Kassubek, University of Ulm, Germany*

> \*Correspondence: *Jue Zhang zhangjue@pku.edu.cn*

### Specialty section:

*This article was submitted to Applied Neuroimaging, a section of the journal Frontiers in Neurology*

Received: *31 December 2019* Accepted: *18 March 2020* Published: *08 April 2020*

### Citation:

*Zhang J, Chen K, Wang D, Gao F, Zheng Y and Yang M (2020) Editorial: Advances of Neuroimaging and Data Analysis. Front. Neurol. 11:257. doi: 10.3389/fneur.2020.00257*

The second subtopic is about the exploration of neural mechanisms of different clinical pathologies and investigations of promising diagnostic methods, including the following eight papers. By using high-field magnetic resonance imaging (7.0- T MRI), Ling et al. analyzed the changes in the lenticulostriate arteries (LSAs) measurements, such as the number of LSA branches and the proportion of discontinuous LSAs in patients with CADASIL. They showed that patients with CADASIL exhibit fewer LSA branches and a higher proportion of discontinuous LSAs than healthy individuals. This suggested that 7.0-T MRI provides a promising and non-invasive method for the study of small artery damage in CADASIL. Song et al. demonstrated the feasibility of performing Functional Ultrasound Imaging (fUS) on two animal models during spinal cord stimulation (SCS). This study could pave the way for future systematic studies to investigate spinal cord functional organization and the mechanisms of spinal cord neuromodulation in vivo. Klietz et al. attempted to study the altered brain metabolism in Parkinson's disease (PD) patients systematically with the aid of the whole-brain MR spectroscopic imaging (wbMRSI). They demonstrated that wbMRSI-detectable brain metabolic alterations revealed the potential to serve as biomarkers for early PD.

Generally, how to make better use of image analysis methods to improve the efficiency and accuracy of clinical diagnosis is always an essential issue in neuroimaging. Sun et al. investigated topological organization of the brain structural connectome and demonstrated more severe disruptions of structural connectivity in amnestic mild cognitive impairment (aMCI) converters compared with non-converters. This work may provide potential structural connectome/connectivity-based biomarkers for predicting disease progression in aMCI, which is of great importance for the early diagnosis of Alzheimer's disease (AD). Ruffini et al. proposed a deep learning model for diagnosis/prognosis of Parkinson's disease (PD) derived from only a few minutes of eyes-closed resting electroencephalography data (EEG) and obtained excellent predicting performance, which perhaps contributes a useful tool for the analysis of EEG dynamics. Wuschek et al. aimed to reduce the variance of CSF protein concentrations and, hence, to increase their diagnostic value by considering brain volumes derived from magnetic resonance imaging (MRI). This work can still be considered as a meaningful attempt despite the conclusion that accounting for individual brain volumes is unlikely to decrease the variability of CSF protein concentrations considerably. Moreover, there are also several investigations about the experiment design and pathological characteristics. Schäfer et al. conducted a study to find optimized design paradigms for presenting baby body odors in the fMRI. The paradigms they recommend may transfer to general body odor perception. Jama-António et al. evaluated the frequency of hippocampal atrophy (HA), and the imaging findings and clinical evolution in patients with calcified neurocysticercotic lesions (CNLs), which promotes to identify parenchymal alterations associated with the occurrence of epileptic seizures.

The third subtopic mainly includes four literature reviews, including image analysis methods and challenges of new neuroimaging technology in clinical application. It comprehensively summarizes the functional magnetic resonance imaging, brain structure and aging, and other fields. Franke and Gaserfocused on establishing biomarkers of the neuroanatomical aging processes exemplifies for predicting age-associated neurodegenerative diseases. They summarized recent studies that utilize the innovative BrainAGE biomarker to evaluate the effects of interaction of genes, environment, life burden, diseases, or lifetime on individual neuroanatomical aging. Furthermore, they concluded that predictive analysis method could provide a personalized biomarker of brain structure, which helps to clarify and further study the patterns and mechanisms of individual differences in brain structure and disorder stages. Besides the BrainAGE biomarker, simultaneous EEG-fMRI technology could offer the possibility to characterize the relationship between EEG spectrum and regional brain activation, providing new insights on neurological and psychiatric diseases and, hopefully, new treatment targets. Mele et al. paid attention to simultaneous EEG-fMRI technology and related early studies, dealing with issues related to the acquisition and processing of simultaneous signals. They realized that despite this technique appear essential to investigate physiological brain networks in healthy subjects, which introduce new evidence about the electrical neural activity and the neurovascular coupling underpinning the BOLD signal, the optimal integrated and standardized analysis is still open, representing the real challenge that follows the technological development. Moreover, there are many innovative applications based on deep learning in various technical aspects of Neuro-Imaging, particularly applied to image acquisition, risk assessment, segmentation tasks. Zhu et al. addressed this topic and presented an overview. They pointed out that although deep learning techniques in medical imaging have been enthusiastically applied to imaging techniques with many enlightening advances, they are still in the initial stage and face challenges such as overfitting and difficult interpretation of models, lack of high-quality data sets, etc. It is worth mentioning that there is also a very interesting work in this part, O'Connor and Zeffiro summarized the difficulties of resting fMRI (rsfMRI) in clinical diagnosis thorough investigation, such as availability of robust denoising procedures, and single-subject analysis techniques. The survey results showed that despite some perceived impediments to expanding clinical rs-fMRI use, neuroradiologists were generally confident in the clinical research and application of rs-fMRI.

To sum up, this special issue covers three topics of neuroimaging and data analysis: (1) exploring the physiological mechanism and diagnostic methods of clinical diseases; (2) investigating how the new technology can be effectively applied in clinical practice; (3) tracking the development of cuttingedge technologies. These researches not only contribute to understanding the impact of the development of neuroimaging on the perception of the nerve system, especially in the influence on structure-function and brain-behavior relationships, but also provide new insight into the role of neuroimaging in clinical application. Using imaging techniques to advance the understanding of pathology, abnormal development, and the use of biomarkers or other questions of clinical utility will be an essential part of neuroimaging. However, it is also a problem worthy of attention to objectively view the development

of new technology and its proper use in clinical practice. In particular, deep learning, as an excellent and widely used image analysis method, has much work to do to increase its internal interpretability and use limited medical data for practical analysis.

# AUTHOR CONTRIBUTIONS

JZ organized and proofread the writing of the editorial. KC, YZ, DW, FG, and MY wrote the manuscript draft.

**Conflict of Interest:** The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

Copyright © 2020 Zhang, Chen, Wang, Gao, Zheng and Yang. This is an open-access article distributed under the terms of the Creative Commons Attribution License (CC BY). The use, distribution or reproduction in other forums is permitted, provided the original author(s) and the copyright owner(s) are credited and that the original publication in this journal is cited, in accordance with accepted academic practice. No use, distribution or reproduction is permitted which does not comply with these terms.

# Prediction of Conversion From Amnestic Mild Cognitive Impairment to Alzheimer's Disease Based on the Brain Structural Connectome

Yu Sun1†, Qiuhui Bi 2,3,4†, Xiaoni Wang1†, Xiaochen Hu<sup>5</sup> , Huijie Li 6,7, Xiaobo Li <sup>8</sup> , Ting Ma<sup>9</sup> , Jie Lu<sup>10</sup>, Piu Chan1,11,12, Ni Shu2,3,4 \* and Ying Han1,11,12,13 \*

<sup>1</sup> Department of Neurology, XuanWu Hospital of Capital Medical University, Beijing, China, <sup>2</sup> State Key Laboratory of Cognitive Neuroscience and Learning & IDG/McGovern Institute for Brain Research, Beijing Normal University, Beijing, China, <sup>3</sup> Center for Collaboration and Innovation in Brain and Learning Sciences, Beijing Normal University, Beijing, China, <sup>4</sup> Beijing Key Laboratory of Brain Imaging and Connectomics, Beijing Normal University, Beijing, China, <sup>5</sup> Department of Psychiatry and Psychotherapy, Medical Faculty, University of Cologne, Cologne, Germany, <sup>6</sup> Department of Psychology, University of Chinese Academy of Sciences, Beijing, China, <sup>7</sup> CAS Key Laboratory of Behavioral Science, Institute of Psychology, Beijing, China, <sup>8</sup> Department of Biomedical Engineering, New Jersey Institute of Technology, Newark, NJ, United States, <sup>9</sup> Department of Electronic and Information Engineering, Harbin Institute of Technology Shenzhen Graduate School, Shenzhen, China, <sup>10</sup> Department of Radiology, XuanWu Hospital of Capital Medical University, Beijing, China, <sup>11</sup> Beijing Institute of Geriatrics, XuanWu Hospital of Capital Medical University, Beijing, China, <sup>12</sup> National Clinical Research Center for Geriatric Disorders, Beijing, China, <sup>13</sup> Center of Alzheimer's Disease, Beijing Institute for Brain Disorders, Beijing, China

### Edited by:

Hongyu An, Washington University in St. Louis, United States

### Reviewed by:

Nicola Amoroso, Università degli Studi di Bari, Italy Jordi A. Matias-Guiu, Hospital Clínico San Carlos, Spain

### \*Correspondence:

Ying Han hanying@xwh.ccmu.edu.cn Ni Shu nshu@bnu.edu.cn

†These authors have contributed equally to this work

### Specialty section:

This article was submitted to Applied Neuroimaging, a section of the journal Frontiers in Neurology

Received: 10 October 2018 Accepted: 20 December 2018 Published: 10 January 2019

### Citation:

Sun Y, Bi Q, Wang X, Hu X, Li H, Li X, Ma T, Lu J, Chan P, Shu N and Han Y (2019) Prediction of Conversion From Amnestic Mild Cognitive Impairment to Alzheimer's Disease Based on the Brain Structural Connectome. Front. Neurol. 9:1178. doi: 10.3389/fneur.2018.01178 Background: Early prediction of disease progression in patients with amnestic mild cognitive impairment (aMCI) is important for early diagnosis and intervention of Alzheimer's disease (AD). Previous brain network studies have suggested topological disruptions of the brain connectome in aMCI patients. However, whether brain connectome markers at baseline can predict longitudinal conversion from aMCI to AD remains largely unknown.

Methods: In this study, 52 patients with aMCI and 26 demographically matched healthy controls from a longitudinal cohort were evaluated. During 2 years of follow-up, 26 patients with aMCI were retrospectively classified as aMCI converters and 26 patients remained stable as aMCI non-converters based on whether they were subsequently diagnosed with AD. For each participant, diffusion tensor imaging at baseline and deterministic tractography were used to map the whole-brain white matter structural connectome. Graph theoretical analysis was applied to investigate the convergent and divergent connectivity patterns of structural connectome between aMCI converters and non-converters.

Results: Disrupted topological organization of the brain structural connectome were identified in both aMCI converters and non-converters. More severe disruptions of structural connectivity in aMCI converters compared with non-converters were found, especially in the default-mode network regions and connections. Finally, a support vector machine-based classification demonstrated the good discriminative ability of structural connectivity in differentiating aMCI patients from controls with an accuracy of 98%, and in discriminating converters from non-converters with an accuracy of 81%.

**8**

Conclusion: Our study provides potential structural connectome/connectivity-based biomarkers for predicting disease progression in aMCI, which is important for the early diagnosis of AD.

Keywords: brain network, conversion, diffusion tensor imaging, graph theory, mild cognitive impairment, machine learning

## INTRODUCTION

Mild cognitive impairment (MCI) is generally associated with a higher risk of dementia and is considered as an intermediate stage between normal aging and Alzheimer's disease (AD) (1– 3). A prospective population-based study in elders showed that the incidence of dementia was highest for patients with amnestic MCI (aMCI) (4). However, not all patients with aMCI progress to dementia (5). Early prediction and identification of individuals with aMCI who are at high risk for conversion to AD aids timely detection of dementia, which is essential for early intervention strategies.

Previous studies have shown the potential of imaging markers to predict conversion from MCI to AD dementia. Among multiple neuroimaging modalities, MRI has attracted significant interests due to its completely non-invasive nature, high availability in mild symptomatic patients and high spatial resolution. Structural MRI biomarkers such as gray matter atrophy in the medial temporal lobe (6) and hippocampal/entorhinal cortex (7) have been identified as efficacious AD-specific biomarkers for the early diagnosis and prediction of disease progression. With diffusion MRI techniques, promising markers of microstructural white matter (WM) damage in AD and MCI patients have been proposed (5, 8, 9). Specifically, regional diffusion metrics of limbic WM in the fornix, posterior cingulum, and parahippocampal gyrus have shown better performance than volumetric measurements of gray matter in predicting MCI conversion (10–14).

However, compared with local or regional imaging markers, the network model has provided a new perspective to investigate the neuropathological progression of AD from a system level (15–18). The whole-brain WM structural network at macroscopic level can be constructed with diffusion MRI and tractography approaches. The topological organization of brain network can be further characterized with graph theoretical analysis (for reviews, see (19, 20). Several non-trivial topological properties, such as small-worldness, modular structure, and rich-club organization of WM networks have been consistently demonstrated in healthy population (21, 22). For AD and aMCI, previous WM network studies have suggested that AD patients exhibit decreased topological efficiency than healthy controls, which is associated with cognitive decline (23, 24). Similarly, our previous work has also found decreased network efficiency in patients with aMCI (25–27) and in those at an earlier stage (28). Importantly, hub regions are preferentially disrupted in AD and aMCI patients, especially the default mode network (DMN) regions, which concentrated most of the pathology of Aβ deposition (29–32). These findings suggest potential, sensitive connectome-based markers for the early detection of structural alterations due to pathological or/and neurodegenerative processes in the early stages of AD. Recently, machine learning, deep learning and complex brain networks have been recently applied to the early diagnosis of neurodegenerative diseases with interesting results (33–36). Specifically, functional MRI network studies have found more severe disruptions in MCI converters, which may distinguish converters from non-converters with high accuracy (37–39). Structural MRI studies have also found topological differences of brain connectome between the two groups (40–42). However, whether the structural brain connectome can provide sensitive markers to predict longitudinal conversion from aMCI to AD has remained largely unknown.

Thus, in our study, we focused on aMCI patients who progressed to probable AD in 2 years after their baseline scan (referred to as "aMCI converters") and compared them with aMCI patients who were clinically stable (i.e., did not develop AD) during 2 years follow-up (referred to as "aMCI non-converters"). Diffusion MRI tractography and graph theory approaches were performed to investigate baseline differences in the topological organization of the WM structural networks between aMCI converters and non-converters. We sought to determine (1) whether the WM networks would show progressive alterations in aMCI converters compared with nonconverters, (2) how network disruptions would predict disease progression in aMCI patients, and (3) the potential utility of brain structural connectome for individual prediction and diagnosis in the early stage of AD.

### MATERIALS AND METHODS

### Participants

This retrospective study involved 78 elderly subjects, including 52 aMCI patients, who were recruited from the Memory Clinic of the Neurology Department, XuanWu Hospital, Capital Medical University, Beijing, China and 26 demographically matched healthy controls (HCs) who were recruited from local communities. The inclusive criteria of aMCI patients were proposed by Petersen (43, 44) and described as follows: (1) definite complaints of memory declined, preferably confirmed by an informant; (2) objective cognitive performances in single or multiple domains including memory documented by neuropsychological tests scores were below or equal to 1.5 SD of age- and education-adjusted norms; (3) a Clinical Dementia Rating (CDR) score of 0.5; (4) preservation of independence in activities of daily living; and (5) not sufficient to meet the criteria for dementia based on DSM-IV-R (Diagnostic and Statistical Manual of Mental Disorders, 4th edition, revised). Subjects who had no complaints of cognition and normal objective cognitive performances as well as a CDR score of 0 were referred as HCs. The exclusive criteria of participants were as follows: (1) a history of stroke, traumatic brain injury, neurological/psychiatric diseases, and other central nervous system diseases that may lead to cognitive impairment; (2) major depression (Hamilton Depression Rating Scale score >24 points); (3) other systemic diseases including thyroid dysfunction, syphilis, severe anemia, or HIV that may cause cognitive impairment; (4) addictions or treatments that would influence cognitive ability; (5) vessel disease included cortical and/or subcortical infarcts, or WM hyperintensity and lesions; (6) severe visual or auditory disabilities.

These participants were selected from a larger cohort (n = 205) and consisted of those who had completed MRI scanning at baseline and undergone a 2 years longitudinal follow-up at least once. During follow-up, patients with aMCI were reclassified as aMCI converters (aMCI-c) or aMCI nonconverters (aMCI-nc) based on whether they were subsequently diagnosed with dementia. The diagnosis of dementia was triggered by a change in the CDR score from 0.5 to 1.0 and confirmed by neuropsychological tests and physician evaluations. This study included 26 aMCI-c who converted to AD within 2 years and 26 demographically matched aMCI-nc who remained stable during the follow-up.

All participants underwent regular neuropsychological assessments, including the Mini-Mental State Examination (MMSE) (45), Montreal Cognitive Assessment (MoCA) (46), Auditory Verbal Learning Test (AVLT), CDR (47), Hamilton Depression Rating Scale (48), and Activities of Daily Living scale. The study was registered on ClinicalTrials.gov (Identifier: NCT02225964) and study protocol was approved by XuanWu Hospital of Capital Medical University institutional review board, and all participants completed a written informed consent process before any study procedures. **Table 1** summarized the main demographic and clinical information of all participants.

### Data Acquisition

All participants were scanned using a Siemens Trio 3.0 T MRI scanner at XuanWu Hospital of Capital Medical University. Participants lay still with their heads fixed by straps and foam to minimize movement. The T1-weighted images were acquired using a magnetization prepared rapid gradient echo (MPRAGE) sequence with the following parameters: repetition time (TR) = 1,900 ms; echo time (TE) = 2.2 ms; flip angle = 9 ◦ ; acquisition matrix = 256 × 224; field of view (FOV) = 256 × 224 mm<sup>2</sup> ; slice thickness = 1 mm; no gap; 176 sagittal slices; and average = 1. The diffusion tensor imaging (DTI) data were acquired using a single-shot EPI sequence with the following parameters: TR = 11,000 ms; TE = 98 ms; flip angle = 90◦ ; acquisition matrix = 128 × 116; FOV = 256 × 232 mm<sup>2</sup> ; slice thickness = 2 mm; no gap; 60 axial slices; and average = 3. Thirty non-linear diffusion weighting directions with b = 1,000 s/mm<sup>2</sup> and one b0 image were obtained. All images were reviewed and the leukoencephalopathy and vascular comorbidity was evaluated by an experienced neuroradiologist.

### Data Preprocessing

First, the DTI data was preprocessed to remove the effect of eddy current distortion and motion artifact by applying an affine alignment of the diffusion-weighted images to the reference b0 image. Then the transformation was applied to reorient the b-matrix. Second, the diffusion tensor was calculated and diagonalized to obtain 3 eigenvalues (λ1, λ2, λ3) and their corresponding eigenvectors. Finally, the FA image was calculated. The preprocessing procedure was performed with the FMRIB Diffusion Toolbox (FDT) in FSL (version 5.0, http://fsl.fmrib.ox. ac.uk/fsl/fslwiki/FDT).

### Brain Network Construction

For each participant, the individual WM structural network was constructed with the following procedures.

### Network Node Definition

To define the network node, we used the Automated Anatomical Labeling (AAL) atlas to parcellate the brain into 90 regions (49). Briefly, T1-weighted image was coregistered to the b0 image in DTI space. Then the transformed T1 images were normalized to the ICBM152 T1 template in the Montreal Neurological Institute (MNI) space. Finally, inverse transformations were applied to AAL atlas to obtain an individual parcellation of 90 ROIs (45 for each hemisphere, **Table S1**), each representing a node of the network (**Figure 1**). All procedures were performed using the SPM8 software (https://www.fil.ion.ucl.ac.uk/spm/software/ SPM8/).

### WM Tractography

Deterministic tractography was performed to reconstruct the whole-brain fiber streamlines, by seeding each voxel with an FA >0.2. The tractography was terminated if it reached a voxel with an FA <0.2 or turned an angle >45 degrees (50). The tractography was performed using Diffusion Toolkit (http://www.trackvis.org/dtk/) based on the "fiber assignment by continuous tracking" method (50).

### Network Edge Definition

Between each pair of ROIs, the weight of the edge was defined as the number of fiber streamlines (FN) with two end points located in these two regions. Therefore, an FN-weighted 90 × 90 structural connectivity (SC) network was constructed for each participant (**Figure 1**).

# Network Analysis

### Small-World Properties

Several graph metrics were calculated to quantify the topological organization of WM structural networks, including network strength (Sp), global efficiency (Eglob), local efficiency (Eloc), shortest path length (Lp), clustering coefficient (Cp), and smallworld parameters (λ, γ, and σ) (51). For regional characteristics, we calculated the nodal global and local efficiency (52). The detailed definitions of these network measures can refer to (51) and **Supplement 1**.

### TABLE 1 | Demographics and neuropsychological testing.


Values are represented as the mean ± SD (range). All of the scores are raw values.

HC, healthy control; aMCI, amnestic mild cognitive impairment; aMCI-c, aMCI converters; aMCI-nc, aMCI non-converters; MMSE, Mini-Mental State Examination (Chinese Version); MoCA, Montreal Cognitive Assessment (Beijing Version); AVLT, Auditory Verbal Learning Test.

\*The P-values were obtained using one-way analysis of variance (ANOVA). Post-hoc pairwise comparisons were performed using a t-test. P < 0.05 was considered significant. #The P-values were obtained using the Kruskal-Wallis one-way ANOVA.

<sup>a</sup>post-hoc paired comparisons showed a significant group difference between aMCI-c vs. aMCI-nc.

<sup>b</sup>post-hoc paired comparisons showed a significant group difference between aMCI-c vs. HC.

<sup>c</sup>post-hoc paired comparisons showed a significant group difference between aMCI-nc vs. HC.

FIGURE 1 | Flowchart for construction of the WM structural network by DTI. (1) Coregistration from an individual T1-weighted image (A) to a DTI b0 image (B). (2) Nonlinear registration from the T1-weighted image in the native DTI space to the ICBM152 T1 template in the MNI space (D). (3) Application of the inverse transformation (T−<sup>1</sup> ) to the AAL atlas in the MNI space (E), which results in individual-specific parcellation in the native DTI space (F). (4) The reconstruction of the whole-brain WM fibers (C) was performed using deterministic tractography in Diffusion Toolkit. (5) The weighted networks of each subject (G) were created by computing the number of the streamlines that connected each pair of brain regions. The connection matrix and 3D representation (axial view) of the WM structural network of a representative healthy subject are shown in the right panel. The nodes are located according to their centroid stereotaxic coordinates and the edges are sized according to their connection weights.

### Hub Distribution

To identify the hub distributions of WM networks in each group, we constructed the backbone network with consistent edges which exist in over 80% subjects for each group. Based on the backbone network, we identified the hub regions by sorting the nodal degree [K(i) > mean + std]. According to the categorization of the nodes into hub and non-hub regions, the edges were classified into rich-club, feeder and local connections (21, 22). Finally, the connection strength of each type of connections were calculated for each participant.

The graph analyses of brain networks were performed using the in-house software, GRETNA (http://www.nitrc.org/projects/ gretna/) (53) and were visualized using BrainNet Viewer software (http://www.nitrc.org/projects/bnv/) (54).

### Statistical Analysis Group Differences

Demographic factors and clinical scores including age, years of education, and neuropsychological scores among the three groups were compared using one-way analysis of variance (ANOVA). Post-hoc pairwise comparisons were then performed using t-tests. Gender distribution was compared with the Kruskal-Wallis one-way ANOVA. To determine the group difference in network metrics, comparisons were performed with univariate analysis of covariance (ANCOVA). Post-hoc pairwise comparisons were then performed using a general linear model. The effects of age, gender and years of education were adjusted for all of these analyses. For regional properties, multiple comparisons were corrected by using the false discovery rate (FDR) correction.

### Network-Based Statistic (NBS)

To identify the specific connected components with significant different structural connections between each pair of groups, we used a NBS approach (55). Briefly, a primary clusterdefining threshold was first applied to identify connections above threshold, and the size (i.e., number of edges) of all connected components was determined. For each component, a corrected p-value was calculated using the null distribution of maximal connected component size, which was derived using the permutation approach (5,000 permutations). Notably, multiple linear regressions were performed to remove the effects of age, gender and years of education before the permutation tests. The detailed descriptions of the NBS analyses can refer to (55) and **Supplement 1**.

### ROC Analysis

To determine the power of the connection strength of the NBS components to serve as potential biomarkers for clinical diagnosis of aMCI patients and differentiation between converters and non-converters, we performed a receiver operating characteristic (ROC) curve analysis for the strength of NBS components, which showed significant group differences.

### Relationships Between Network Metrics and Clinical Scores

For the network metrics showing significant group differences, partial correlation analyses were performed between the network metrics and clinical scores for aMCI converters and nonconverters separately, while removing the effects of age, gender and years of education. All the statistical analyses were performed using the MATLAB program (The MathWorks, Inc.).

### Support Vector Machine-Based Classification

To determine the discriminative ability of structural connectivity in separating aMCI patients from controls and separating converters from non-converters, we used the connection strength of the edges as the features for individual classification. For each pair of groups, we performed a support vector machine (SVM) classification, with a Gauss kernel function and the default settings of C = 1, coef = 0 and gamma as the reciprocal of the number of features in the LIVSVM Toolbox (http://www.csie. ntu.edu.tw/~cjlin/libsvm/) (56). Leave-one-out cross-validation (LOOCV) was used to evaluate the SVM model. Each subject was designated the test subject in turns while the remaining ones were used to train the SVM predictor. The hyperplane derived from the training subjects was then used to make a prediction about the group label of the test subject. Sensitivity, specificity, accuracy, and area under the curve (AUC) value were calculated to assess the performance of the classifier.

To avoid overfitting and reduce the redundant information, the F-score was calculated for each feature (connection), and the features with higher F-scores were used to train the model. The number of selected features (1%−20% with an interval of 1%) was decided by a grid search. The F-score was defined as (57):

$$F(\mathbf{i}) = \frac{\left(\overline{\mathbf{x}}\_i^{(+)} - \overline{\mathbf{x}}\_i\right)^2 + \left(\overline{\mathbf{x}}\_i^{(-)} - \overline{\mathbf{x}}\_i\right)^2}{\frac{1}{n\_+ - 1} \sum\_{k=1}^{n\_+} \left(\mathbf{x}\_{k,i}^{(+)} - \overline{\mathbf{x}}\_i^{(+)}\right)^2 + \frac{1}{n\_- - 1} \sum\_{k=1}^{n\_-} \left(\mathbf{x}\_{k,i}^{(-)} - \overline{\mathbf{x}}\_i^{(-)}\right)^2} \tag{1}$$

where x i , x (+) i , x (−) i are the average of the i-th feature of the whole, positive, and negative data sets, respectively; x (+) k, i is the ith feature of the k-th positive instance; and x (−) k, i is the i-th feature of the k-th negative instance.

The radial basis kernel function was defined as:

$$\mathbf{K}\left(\mathbf{x},\mathbf{z}\right) = e^{\left(\frac{\|\mathbf{x}-\mathbf{z}\|^2}{2\rho^2}\right)}\tag{2}$$

where x, z is the feature vector of a different instance; e is the Euler number, and γ is the hyper-parameter.

# Reproducibility Analysis

### Effects of Different Thresholds

To test the stability of the results, we constructed individual WM networks with five different thresholds of fiber number (wij = 1, 2, 3, 4, 5). If the streamline number of an edge was less than the threshold, the edge weight was set to zero. For each threshold, the global network metrics were computed, and the group differences were assessed.

### Effects of Different Parcellation Schemes

To evaluate the effects of different parcellation schemes on the network metrics, we further subdivided the low-resolution AAL (L-AAL) template into 1024 ROIs of equal size [i.e., high-resolution (H-1024)] (58). A high-resolution network was constructed for each participant and followed that with the same network analysis.

## RESULTS

### Demographics and Neuropsychological Testing

No group differences were found in age, gender and years of education among the three groups. For clinical scores, aMCI patients showed a lower MMSE [F(2,75) = 20.29, p < 0.001], MoCA [F(2,75) = 58.54, p < 0.001], and AVLT scores [AVLT-immediate recall: F(2,75) = 53.07, p < 0.001; AVLTdelayed recall: F(2,75) = 59.08, p < 0.001; AVLT-recognition: F(2,75) = 20.22, p < 0.001] than controls. Between the two aMCI groups, lower MMSE and MoCA scores were observed in aMCI converters relative to non-converters (all p < 0.05; **Table 1**).

### Global Topology of the WM Structural Networks

Characteristic small-world organization of the WM networks (λ ≈ 1, γ > 1) were observed for both aMCI patients and control subjects. Among the three groups, ANCOVAs on the global network properties showed significant group effects in network strength [F(2,75) = 10.18, p = 0.0001], global efficiency [F(2,75) = 6.51, p = 0.0025], local efficiency

### TABLE 2 | Group differences in global network metrics.


Values are represented as the mean ± SD. Abbreviations: HC, healthy control; aMCI, amnestic mild cognitive impairment; aMCI-c, aMCI converters; aMCI-nc, aMCI non-converters. Lp, shortest path length; Cp, clustering coefficient.

The P-values were obtained with a univariate analysis of covariance (ANCOVA). Post-hoc pairwise comparisons were then performed using a general linear model. The effects of age, gender and years of education were adjusted for all of these analyses. P < 0.05 was considered significant.

apost-hoc paired comparisons showed a significant group difference between aMCI-c vs. aMCI-nc.

<sup>b</sup>post-hoc paired comparisons showed a significant group difference between aMCI-c vs. HC.

<sup>c</sup>post-hoc paired comparisons showed a significant group difference between aMCI-nc vs. HC.

[F(2,75) = 8.05, p = 0.0007], shortest path length [F(2,75) = 6.40, p = 0.0028] and clustering coefficient [F(2,75) = 5.20, p = 0.0078; **Table 2**] (**Figure 2**). In addition, post-hoc comparisons showed significantly reduced network strength, global efficiency and local efficiency in both aMCI converters and non-converters relative to the controls. Increased shortest path length and decreased clustering coefficient were found only in aMCI converters relative to controls. Between aMCI converters and nonconverters, significant group differences were found in network strength [t(47) =2.28, p = 0.027], local efficiency [t(47) = 2.19, p = 0.034), shortest path length [t(47) = −2.12, p = 0.039], and clustering coefficient [t(47) = 2.20, p = 0.033; **Table 2**; **Figure 2**].

### Node-Based Analysis

Following the discovery of a disrupted global network organization, we further localized the regions with altered nodal global and local efficiency. For nodal global efficiency, regions with significant group effects were mainly distributed in the frontal and parietal cortices, including 7 frontal regions (right dorsolateral superior frontal gyrus, right middle frontal gyrus, right opercular part of the inferior frontal gyrus, right triangular part of the inferior frontal gyrus, left anterior cingulate gyrus, bilateral supplementary motor area) and 3 parietal regions (left posterior cingulate gyrus, bilateral precuneus) (p < 0.05, corrected) (**Figure 3**). Post-hoc tests showed that all of these regions showed reduced global efficiency in both aMCI converters and non-converters relative to controls. In particular, several brain regions showed more severe disruptions in aMCI converters compared with non-converters, including the bilateral precuneus, left anterior cingulate gyrus, right middle frontal gyrus, and right triangular part of the inferior frontal gyrus (all p < 0.05).

For nodal local efficiency, regions with significant group effects were mainly distributed in the limbic cortices (bilateral median cingulate and paracingulate gyri and posterior cingulate gyrus), temporal cortices (left superior temporal gyrus, right temporal pole, and bilateral hippocampus), subcortical regions (left caudate nucleus and bilateral putamen) and right superior occipital gyrus (p < 0.05, corrected) (**Figure 4**). All of these regions had a reduced local efficiency in aMCI converters compared with controls. In seven of these regions, including the bilateral putamen, bilateral median cingulate and paracingulate gyri, left posterior cingulate gyrus, left hippocampus and left caudate nucleus, reduced local efficiency was observed in aMCI non-converters compared with controls. Between the two aMCI groups, four regions (left superior temporal gyrus, right superior occipital gyrus, right posterior cingulate gyrus, and right hippocampus) showed a more severe disruption of local efficiency in the aMCI converters relative to non-converters (all p < 0.05).

### Connectivity-Based Analysis

NBS analyses were carried out to identify the disrupted connected components in patients. Compared to healthy controls, a single component with 83 nodes and 177 connections was altered in aMCI converters (p < 0.001, corrected) and a component with 73 nodes and 122 connections was detected in aMCI non-converters (p < 0.001, corrected) (**Figure 5A**). The involved regions had a widespread distribution across the frontal, temporal, parietal, occipital, and subcortical regions. The comparison between aMCI converters and non-converters revealed a component with decreased strength in converters, which was composed of 70 nodes and 81 connections (p < 0.05, corrected), mainly involving the bilateral precuneus, bilateral putamen, left anterior cingulate gyrus, right superior parietal gyrus, left middle temporal gyrus, left paracentral lobule, and left superior occipital gyrus (**Figure 5A**).

ROC analyses were performed to evaluate the discriminative ability of the disrupted component identified by NBS. The NBS component exhibited good performance for the discrimination between aMCI converters and healthy controls (with an AUC value of 0.96), between aMCI non-converters and healthy controls (with an AUC value of 0.91) and between aMCI

FIGURE 2 | Global measures of the WM structural network were quantified in the aMCI converters, non-converters, and controls. The bars and error bars represent the fitted values and the standard deviations, respectively. The fitted values indicates the residuals of the original values of the network metrics after removing the effects of age, gender and years of education. The asterisk indicated a significant difference between groups. (\*) represents a significant group difference at p < 0.05; (\*\*) represents a significant group difference at p < 0.01; and (\*\*\*) represents a significant group difference at p < 0.001.

converters and non-converters (with an AUC value of 0.89) (**Figure 5B**).

### Rich-Club Organization

Similar hub distributions were found across three groups (**Figure 6A**), mainly located in bilateral precuneus, bilateral putamen, right dorsolateral superior frontal gyrus, left middle temporal gyrus and several occipital regions. Several brain regions were identified as hubs only in the control group, such as bilateral orbital part of superior frontal gyrus. Among the three groups, significant group effects were identified in the strength of rich-club [F(2,75) = 6.67, p = 0.0022], feeder [F(2,75) = 7.25, p = 0.0013] and local [F(2,75) = 9.44, p = 0.0002] connections (**Figure 6B**). Compared with healthy controls, aMCI converters showed significant decreases in all three types of connections (all p < 0.005) and aMCI non-converters showed decreases in richclub [t(47) = 2.06, p = 0.045] and local [t(47) = 2.82, p = 0.007] connections. Only feeder connections decreased significantly in aMCI converters compared with non-converters [t(47) = 2.26, p = 0.028].

## Correlations Between Network Metrics and Neuropsychological Tests

The relationship between network metrics and clinical scores were examined for aMCI converters and non-converters, respectively. In aMCI converters: MoCA was positively correlated with global efficiency (r = 0.41; p = 0.049), and negatively correlated with shortest path length (r = −0.53; p = 0.010); MMSE was negatively correlated with shortest path length (r = −0.43; p = 0.041) (**Figure 7A**). In aMCI non-converters: MMSE was positively correlated with network strength (r = 0.44; p = 0.034) and global efficiency (r = 0.47; p = 0.022), and negatively correlated with shortest path

FIGURE 4 | The distribution of brain regions with significant group effects in the nodal local efficiency among the three groups (p < 0.05, corrected). The node sizes indicate the significance of group differences in the nodal local efficiency. For each node, the bar and error bar represent the fitted values and the standard deviations, respectively, of the nodal local efficiency in each group. post-hoc tests revealed that all of these regions had a reduced nodal local efficiency in aMCI converters compared with controls. In seven of these regions, including the bilateral putamen, bilateral median cingulate, and paracingulate gyri, left posterior cingulate gyrus, left hippocampus and left caudate nucleus, reduced local efficiency was observed in aMCI non-converters compared with controls. Between the two aMCI groups, four regions (left superior temporal gyrus, right superior occipital gyrus, right posterior cingulate gyrus, and right hippocampus) showed a more severe disruption of local efficiency in the aMCI converters relative to non-converters. (\*) represents a significant group difference at p < 0.05; (\*\*) represents a significant group difference at p < 0.01; and (\*\*\*) represents a significant group difference at p < 0.001.

FIGURE 5 | Altered structural connectivity between each pair of groups identified using NBS. (A) Compared to healthy controls, a single component with 83 nodes and 177 connections was altered in aMCI converters (p < 0.001, corrected) and a component with 73 nodes and 122 connections was detected in aMCI non-converters (p < 0.001, corrected). The comparison between aMCI converters and non-converters revealed a component with decreased strength in converters, which was composed of 70 nodes and 81 connections (p < 0.05, corrected). The edge sizes indicate the significance of the between-group differences. (B) ROC curve of the NBS component between aMCI converters and healthy controls (AUC = 0.96); between aMCI non-converters and healthy controls (AUC = 0.91); and between aMCI converters and non-converters (AUC = 0.89). (AUC, area under the curve).

length (r = −0.47; p = 0.025); AVLT-Immediate Recall was positively correlated with global efficiency (r = 0.43; p = 0.038) (**Figure 7B**).

### Individual Classification of aMCI Converters and Non-converters

The results of SVM classification demonstrated good discriminative ability of structural connectivity in the differentiation between aMCI patients and controls and between aMCI converters and non-converters. The ROC curves for the classification between each pair of groups are shown in **Figure 8A**. For the discrimination between aMCI converters and controls, an AUC value of 1.00 was obtained, with an accuracy of 98.08%, sensitivity of 100% and specificity of 96.15%. Between aMCI non-converters and controls, an AUC value of 0.99 was obtained, with an accuracy of 98.08%, sensitivity of 100% and specificity of 96.15%. Between aMCI converters and non-converters, an AUC value of 0.89 was obtained, with an accuracy of 80.77%, sensitivity of 92.31%, and specificity of 69.23%. The effects of number of selected features on the classification accuracy were also evaluated (**Figure S1**).

The discriminative features for the classification were mapped onto the regions, which were rendered with the total number of connections from this region selected as features in the SVM classification (**Figure 8B**). For the classification between aMCI and controls, the most selected features were connections of the bilateral precuneus, bilateral posterior cingulate gyrus, right putamen, right thalamus, right dorsolateral superior frontal gyrus, left orbital part of the inferior frontal gyrus, and left caudate nucleus. For the classification between aMCI converters and non-converters, the most contributed features were connections of the bilateral precuneus, bilateral middle temporal gyrus, bilateral putamen, right medial superior frontal gyrus and left triangular part of the inferior frontal gyrus.

### Reproducibility of the Findings Effects of Different Thresholds

For the different thresholds of network construction (wij = 1,2,3,4,5), similar group differences were found for network strength, global efficiency, local efficiency, and shortest path length (all p < 0.05) (**Figure S2A**).

### Effects of Different Parcellation

For the high-resolution (H-1024) network analysis, significant group effects in network strength [F(2,75) = 10.14, p = 0.0001], global efficiency [F(2,75) = 9.40, p = 0.0002], local efficiency [F(2,75) = 6.41, p = 0.0027], and shortest path length [F(2,75) = 9.30, p = 0.0003] were observed (**Figure S2B**). Post-hoc analysis revealed significantly reduced network strength, global

efficiency, local efficiency and increased shortest path length in both aMCI converters and non-converters relative to the controls (all p < 0.05). Between aMCI converters and nonconverters, significant group differences were found in network

strength [t(47) = 2.06, p = 0.044], global efficiency [t(47) = 2.06, p = 0.044], and shortest path length [t(47) = −2.27, p = 0.028]. The group differences of global network metrics were comparable with those from low-resolution networks.

# DISCUSSION

By combining DTI tractography and graph theoretical analyses, we demonstrated convergent and divergent topological alterations of the brain structural connectome in aMCI converters and non-converters. More severe disruptions of the structural connectome were identified in aMCI converters, especially in the DMN regions and connections. Importantly, the structural connectivity showed good discriminative ability in the differentiation of aMCI converters and nonconverters, providing potential connectome-based markers for the early prediction of disease progression in aMCI patients.

### Global Network Disruption Between MCI Converters And Non-converters

First, we found similar patterns of global network alterations in both aMCI converters and non-converters. Compared with healthy controls, aMCI patients showed reduced network strength, global efficiency and local efficiency, but remained similar with respect to small-world parameters. These findings are consistent with our previous graph analysis of brain structural networks in aMCI patients (26, 27). As a disconnection disease, lower network global, and local efficiency were related to the widespread disruption of both long-range and shortrange structural connectivity in aMCI patients, which indicated the pathological or degenerative alterations of WM in the early stage of AD. The possible mechanisms of structural disconnection may be due to cortical amyloid deposition, neural dysfunction, vascular damage, demyelination and so on (59–61).

Importantly, compared to aMCI non-converters, converters demonstrated lower network strength, local efficiency, and increased shortest path length even at baseline. Lower network strength was associated with sparse connectivity of brain networks, which indicated reduced WM integrity in the early phase of aMCI converters. This finding is in line with the evidence of more severe disruption of WM connectivity in MCI converters than non-converters with conventional DTI analyses (10, 12, 62). In addition, decreased local efficiency is mainly due to the loss of short-range connections among the neighborhood regions, and an increased shortest path length may be attributable to the disrupted long-range connections between remote regions, which is important for interregional effective integrity or prompt transfer of information in brain networks and constitutes the basis of cognitive processes (63). The alteration pattern of WM networks between converters and non-converters was similar to that in prior cross-sectional studies, which have identified network alterations with disease progression in AD and MCI patients (64–66). More severe disruptions of network properties in AD patients relative to MCI patients were found. Our study confirmed these cross-sectional reports of network dysfunction in AD and MCI and extended those with additional new findings.

Before disease transition, more severe structural or functional connectivity alterations already existed in the aMCI converters compared with non-converters (37–41). From the current study, we found that the network measures from DTI data are sensitive enough to detect the topological differences even at baseline, and correlated with the disease severity evaluated by clinical scores (MMSE, MoCA, and AVLT-Immediate Recall) in aMCI patients. Compared with the traditional regional or local brain measures, brain network studies provide a systematic perspective to investigate the disease progression and new insights into understanding the neuropathological mechanisms of disease conversion. Our results suggest the pivotal role of WM network disruption in the genesis of dementia and highlight the potential of a disease marker to identify patients at risk for dementia at an early stage.

## Regional/Connectivity Differences Between MCI Converters and Non-converters

Between aMCI patients and controls, significant differences in nodal global efficiency were mainly located in the bilateral precuneus, prefrontal cortex, and posterior cingulate gyrus, consistent with our previous network findings of aMCI patients (25–27). The reduced nodal global efficiency reflected a disrupted global integration of the structural connectivity in these regions, which may be due to more severe disconnection in ADrelated hub brain regions concentrating most of the amyloid deposition (30, 31, 67–69). Furthermore, relative to aMCI non-converters, aMCI converters showed reduced nodal global efficiency in the bilateral precuneus, left anterior cingulate gyrus and right middle frontal gyrus, the regions that belong to the default mode network (DMN), which overlap with brain regions in distribution of early accumulation of cortical Aβ fibril (70), as well as to the pattern of hypometabolism found on FDG-PET studies (71) and of hypoperfusion on resting MR perfusion studies of AD patients (72). A functional MRI study has suggested the significant predictive value of DMN connectivity in predicting the disease progression to AD in MCI patients (73). Amyloid accumulation started from the DMN and was correlated with hypoconnectivity of the DMN (70). The association between amyloid accumulation and cognition was found to be influenced by functional connectivity of the DMN (74). Moreover, a prior DTI study has suggested that an increased amyloid burden is related to changes in topology of WM network architecture in MCI and AD patients (60), suggesting that pathological propagation affects largescale functional and structural brain networks with disease progression. Notably, the most significant differences between converters and non-converters were located in the bilateral precuneus; as one of the most important regions of the DMN, the precuneus plays a critical role in memory processing and AD progression. A previous structural MRI-based network study has found that betweenness centrality of the precuneus is associated with cognitive decline (75), which may suggest a key role of the precuneus in the disease conversion of aMCI patients.

Meanwhile, group differences in nodal local efficiency were mainly located in the bilateral hippocampus, middle and posterior cingulate gyri, superior and middle temporal gyri, which were characteristic AD-signature regions (9). Previous neuroimaging studies have also reported the structural or functional alterations in these brain regions in AD and MCI patients (27, 76–78). Reduced local efficiency of these regions may reflect local impairment in functional segregation of episodic memory, which may be related to the structural disruptions of short-range connections within the memory network which centered on the hippocampus (79– 81). Relative to non-converter, reduced nodal local efficiency in the left superior temporal gyrus, right hippocampus and right superior occipital gyrus were found in aMCI converters. These regions were also identified as features for predicting progression to AD in MCI patients based on amyloid-PET (82).

Similar hub distributions were found across three groups, which were consistent with previous findings (27, 83). Hubs play a pivotal role in global information transfer and seem to be vulnerable and preferentially affected in AD patients (29). In our study, both hub and non-hub regions showed decreased efficiency and all categories of edges showed lower strength in aMCI patients. Between aMCI converter and non-converters, only feeder connections showed progressive disruption. We speculate that aMCI initiates with a widespread disruption of WM connectivity, and alterations in feeder connections may be with important predictive value for the disease progression.

Machine learning approaches for the individual prediction of disease progression Identifying sensitive and early biomarkers for the individual prediction of disease progression is important for early disease diagnosis and precise medicine. Machine learning approaches with big multimodality data provide a promising area for future intelligent computer-aided-diagnosis (84). For AD and MCI, a number of previous studies have tested different imaging, CSF or neuropsychological measures for the early prediction of disease conversion (9, 85–87). Based on the brain structure connectome and SVM classification, we obtained a high classification accuracy of 98% between aMCI patients and controls. Even between converters and non-converters, the accuracy can reach 81%, which is comparable and even higher than previous results (10, 12, 14, 38, 42, 88). This finding suggested the potential utility of brain structural connectivity/connectome-based markers for the individual prediction of disease conversion, which may provide biologically relevant information not present in other imaging markers.

### Methodological Issues

Several methodological issues should be addressed. First, the results were limited by the small sample size. In the future studies, several large publicly available datasets, such as ADNI, should be used as an independent cohort for validating the reproducibility of our findings. Second, we only identified abnormalities in patients with aMCI converters and non-converters at baseline, and longitudinal follow-up studies of the same study population are needed to verify the effects of early imaging markers for disease prediction. Third, we only studied WM structural networks. In future studies, the combination of the multimodal imaging and conventional pathological biomarkers would contribute to a more comprehensive prediction of the progression from aMCI to AD dementia. Finally, some newly developed network analysis approaches, such as multiplex networks, can help early AD classification (33). These approaches deserve a further investigation in future studies.

# CONCLUSIONS

By using DTI tractography combined with graph analysis, our study demonstrated more severe disrupted topological organization of brain structural connectome in aMCI converters compared with non-converters, providing potential connectivity/connectome-based biomarkers for the early prediction of disease progression in aMCI patients.

# AUTHOR CONTRIBUTIONS

YS, QB, and XW: manuscript preparation and drafting. YS, XW, YH, and JL: clinical assessments and data acquisition. PC and YH: clinical diagnosis. QB, XH, HL, XL, TM, and NS: data analysis and interpretation. NS and YH: study conception and design. YS, QB, and XW contributed equally to this work. All authors have contributed to the manuscript revising and editing critically for important intellectual content and given final approval of the version, agreed to be accountable for all aspects of the work in ensuring that questions related to the accuracy or integrity of any part of the work are appropriately investigated and resolved.

# FUNDING

This work was supported by the National Key Research and Development Program of China (2016YFC1306300, 2016YFC0103000), National Natural Science Foundation of China (Grant 61633018, 81522021, 81430037, 81471731, 31371007, 81471732, 81671761, 81871425), National Basic Research Program (973 Program) (2015CB351702), Beijing Municipal Commission of Health and Family Planning (PXM2018\_026283\_000002), Beijing Nature Science Foundation (7161009, 7132147), the Youth Innovation Promotion Association CAS (2016084), Basic Research Foundation Key Project Track of Shenzhen Science and Technology Program (JCYJ20160509162237418, JCYJ20170413110656460) and Fundamental Research Funds for the Central Universities (2017XTCX04).

### SUPPLEMENTARY MATERIAL

The Supplementary Material for this article can be found online at: https://www.frontiersin.org/articles/10.3389/fneur. 2018.01178/full#supplementary-material

## REFERENCES


impairment patients: a longitudinal whole-brain voxel-wise degree analysis. Front Aging Neurosci. (2016) 8:195. doi: 10.3389/fnagi.2016.00195


connectivity patterns in mild cognitive impairment. Neurobiol Aging (2017) 55:177–89. doi: 10.1016/j.neurobiolaging.2017.03.023


disease and mild cognitive impairment. Alzheimers Dement. (2008) 4:265–70. doi: 10.1016/j.jalz.2008.04.006


**Conflict of Interest Statement:** The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

Copyright © 2019 Sun, Bi, Wang, Hu, Li, Li, Ma, Lu, Chan, Shu and Han. This is an open-access article distributed under the terms of the Creative Commons Attribution License (CC BY). The use, distribution or reproduction in other forums is permitted, provided the original author(s) and the copyright owner(s) are credited and that the original publication in this journal is cited, in accordance with accepted academic practice. No use, distribution or reproduction is permitted which does not comply with these terms.

# The Design Matters: How to Detect Neural Correlates of Baby Body Odors

Laura Schäfer <sup>1</sup> \*, Thomas Hummel <sup>2</sup> and Ilona Croy <sup>1</sup>

<sup>1</sup> Department of Psychotherapy and Psychosomatic Medicine, Technische Universität Dresden, Dresden, Germany, <sup>2</sup> Smell and Taste Clinic, Department of Otorhinolaryngology, Technische Universität Dresden, Dresden, Germany

Functional magnetic resonance imaging of body odors is challenging due to methodological obstacles of odor presentation in the scanner and low intensity of body odors. Hence, few imaging studies investigated neural responses to body odors. Those differ in design characteristics and have shown varying results. Evidence on central processing of baby body odors has been scarce but might be important in order to detect neural correlates of bonding in mothers. A suitable paradigm for investigating perception of baby body odors has still to be established. We compared neural responses to baby body odors in a new to a conventional block design in a sample of ten normosmic mothers. For the new short design, 6 s of continuous odor presentation were followed by 19 s baseline and 13 repetitions were performed. For the conventional long design, 15 s of pulsed odor presentation were followed by 30 s of baseline and eight repetitions were performed. Neural responses were observed in brain structures related to basal and higher-order olfactory processing, such as insula, orbitofrontal cortex, and amygdala. Neural responses following the short design were significantly higher in comparison to the long design. This effect was based on higher number of repetitions but affected olfactory areas differently. The BOLD signal in the primary olfactory structures was enhanced by short and continuous stimulation, secondary structures did profit from longer stimulations with many repetitions. The short design is recommended as a suitable paradigm in order to detect neuronal correlates of baby body odors.

### Edited by:

Hongyu An, Washington University in St. Louis, United States

### Reviewed by:

Bruno J. Weder, University of Bern, Switzerland Flavia Di Pietro, University of Sydney, Australia

### \*Correspondence:

Laura Schäfer laura.schaefer@ uniklinikum-dresden.de

### Specialty section:

This article was submitted to Applied Neuroimaging, a section of the journal Frontiers in Neurology

Received: 10 October 2018 Accepted: 21 December 2018 Published: 16 January 2019

### Citation:

Schäfer L, Hummel T and Croy I (2019) The Design Matters: How to Detect Neural Correlates of Baby Body Odors. Front. Neurol. 9:1182. doi: 10.3389/fneur.2018.01182

Keywords: fMRI design, olfaction, olfactory fMRI, body odor, baby odor, body odor perception

# INTRODUCTION

Neural processing of social stimuli has been well studied for the senses of vision and audition, but examination of interpersonal human chemosensation is just in the beginning due to challenges related to the olfactory system.

The detection of reliable neural activations to odors is complicated due to the anatomical structures of the olfactory system and methodological obstacles related to the presentation of olfactory stimuli (1). We briefly outline those challenges.

Central olfactory processing occurs in several stages [compare (1)]. Olfactory signals coming from the olfactory bulb (OB) pass on to the basal frontal and medial temporal lobe. Thereby, the piriform cortex, the amygdala, the perirhinal and entorhinal cortices receive parts of the incoming information from the OB (2). Those areas are commonly considered as primary olfactory areas (3).

**23**

From there, olfactory information is further processed in secondary structures, such as the anterior insula, hippocampus, hypothalamus, and orbitofrontal cortex (OFC). In contrast to other modalities, olfactory processing is characterized by direct pathways projecting into primary and secondary structures without passing through the thalamus first. Due to the subcortical structures involved in olfactory processing, the detection of olfactory signals in functional magnetic resonance imaging (fMRI) is challenging. The olfactory system is surrounded by the frontal and the paranasal sinus, and the acoustic meatus containing various tissues (bones, vessels, air) with different magnetic field homogeneity characteristics (1). Those make this system especially sensitive to susceptibility artifacts and limit signal detection in the mediobasal parts of the brain. Although well-adjusted fMRI sequences can reduce those artifacts, a systematic overview of the most suitable procedures is still missing.

Another difficulty in olfactory fMRI is the odor presentation: stimulus concentration and duration are typically operated by computer-controlled olfactometers, which are stationed outside the scanner and deliver odors via several meters of tubing to the participants' nose. Thus, presenting precise stimulus onsets is challenging. Particular devices, e.g., portable olfactometers, facilitate stimulus presentation, as they allow the odors to be placed close to the MRI scanner or within the scanning room [e.g., (4)].

Besides that, the rapid adaptation to olfactory stimuli needs to be considered (1) and the length of the olfactory stimulation period as neural oscillations occurring after a longer stimulation time may affect the signal (5).

In addition to those general challenges of olfactory fMRI, the stimulation with body odors has particular demands: Body odors are generally weak and not easily producible or storable in high concentration as compared to other, e.g., liquid odorants. Typically, clothes worn by the subject serve as body odor stimuli, but the amount of odor molecules within such a piece of clothing is limited. This weaker concentration of molecules may explain the weaker neural activation compared to other olfactory stimuli.

Further, the field of studies investigating neuronal processing of body odors is small and lacks conventions about optimal designs. To our knowledge, only four original fMRI investigations on body odor perception exist (compare **Table 1**). Two used a block design with about 20 s of pulsed odor presentation (6, 7); the other two used an event-related design with about 3 s of continuous odor presentation (8, 9). All four studies report weak activations in general and in some studies the expected olfactory areas were not observed at all. Further studies based on positron emission tomography [PET, (10, 11)], or near infrared spectroscopy [NIRS; (12)] report similar, and again, weak effects (see **Table 1**).

Besides olfactory areas, both the anterior and the posterior cingulate cortex (ACC, PCC) have been associated with body odor perception (6, 10) and it was supposed that the processing of endogenous (body-) odors differs from exogenous odors and activates other brain areas apart from the olfactory system (10).

To our knowledge, only two imaging studies have investigated baby body odor perception in mothers [fMRI: (7); NIRS: (12)]. Baby body odors are subtle which implicate that investigations and the detection of strong neural effects are especially challenging. The present study was conducted in order to investigate which design characteristics are particularly suitable for imaging neural responses to baby body odors.

We designed a new, short block presentation paradigm aimed to account for rapid adaptation (by shortening odor presentation time to 6 s) and for weak neural responses following body odors (by increasing the number of stimulus repetitions). We compared this to a long block design, which follows recent recommendations (1) with 15 s of odor presentation; hereby the odor presentation was performed in a pulsed way to overcome adaptation. Our targeted outcome was the strength of neural activation in olfactory relevant brain areas depending on the design. According to previous results, we focused our analysis on the anterior insula, the OFC, the piriform cortex and the thalamus, as well as on the ACC and PCC. We furthermore included the amygdala and the hippocampus as regions of interest (ROI) which are frequently activated in response to odor presentation.

# MATERIALS AND METHODS

The ethics committee of the University of Dresden (Code: EK 104032015) approved the conduction of the study according to the "World Medical Association's Declaration of Helsinki." Written, informed consent was obtained from all participants.

## Participants

Our sample consisted of 10 healthy, normosmic mothers (aged 27 to 39 years, M = 32.2; SD = 4.7) having a child under the age of 2 years (aged 10 to 15 months, M= 10.30, SD = 4.22). Normosmic functioning was ensured with a Sniffin' Sticks identification screening (13). This study was done as a pilot measurement for a larger project.

# Magnetic Resonance Imaging Procedures

Functional magnetic resonance data were acquired on a Siemens 3T scanner SONATA with an 8-channel head coil using a protocol with a T2<sup>∗</sup> -weighted gradient-echo, echo-planar imaging sequence (TR = 2.5 s, TE 51 ms, flip angle 90◦ , 25 mm × 6 mm axial slices, 3.6 × 3.6 mm in-plane resolution). In order to receive a precise anatomical mapping of the functional data, a high resolution T1 sequence (TR = 2.5 s, 0.7 × 1mm inplane resolution) was added. The scanning planes were oriented parallel to the anterior-posterior commissure line and covered olfactory relevant regions from the cerebellum up to the dorsal end of the cingulate cortex. As all areas dorsal to the cingulate cortex were no regions of interest in the present study, we decided to limit the scanned area of axial sections from the brain stem up to the cingulate cortex (z = 45 at y = −80 to z = 20 at y = 60) in order to enhance the repetition time and to allow for more scans during the session.

TABLE 1 | Comparison of olfactory imaging studies investigating neural correlates of body odor perception.


volume corrected; ACC, anterior cingulate cortex; PCC, posterior cingulate cortex; OFC, orbitofrontal

 cortex; PFC, prefrontal cortex.


the baby odor stimuli were observed.

TABLE 2 | Mean values and standard deviations for ratings of pleasantness,

 intensity, and wanting for each design for the single (own, unfamiliar) and merged across own and unfamiliar baby body odor, T-Test

### Body Odor Sampling and Presentation Procedure

Body odor samples were collected with onesies worn for one night by the babies after a standardized procedure (see **Supplementary Material**). The armpit of the onesie was stored in a glass bottle connected with teflon tubes (5 m length) to the air-dilution computer controlled olfactometer ( 4).

Two different designs of odor presentation were used. Both lasted for the same time of 6 min, but differed in the duration, mode, and number of repetitions of odor presentation within (**Supplementary Figure 1**). This was done in order to match previous design characteristics which used either block designs with long pulsed stimulus presentation ( 6 , 7) or event-related designs with short continuous presentation ( 8 , 9).

Hence, we used a long pulsed block design and compared this to a short odor presentation. The short was similar to previous event-related designs in terms of a short continuous presentation but differed as we did not jitter and randomize the olfactory stimuli within the run. We refrained from that in order to not over complicate the comparison with additional variables as study power was limited.

In the long design, 8 on-blocks of 15-s each in which the odor was delivered were followed each by 8 off-blocks of 30 s each. Due to the long on-blocks, a pulsed odor presentation, where 2 s of air followed every 1 s of odor presentation, was used in order to minimize adaptation and habituation to the odors. In the short design, 13 on-blocks in which the odor was continuously delivered for 6 s were followed each by 13 offblocks of 19 s each. Each paradigm was tested with two different stimuli in randomized order: the body odor of the own baby and an unfamiliar sex- and age-matched child, resulting in four runs in total. During baseline, clean air was presented. As the main focus of the present study was to compare the design paradigms, the effect of baby body odor was merged across own and unfamiliar baby for statistical analysis. Single results of own and unfamiliar child are provided in **Supplementary Tables 3, 4** . Before the experiment, participants were instructed to breathe regularly through the nose as follows: "You are presented to baby body odors, one of which is your child. Please, breathe regularly and smoothly as normally through the nose." After each run, participants rated pleasantness, intensity, and wanting of the odor stimuli on a Likert-scale ranging from 1 = "not pleasant/intense/not at all" to 10 "very pleasant/intense/very much." Pleasantness and wanting reflect different characteristics of reward (14). Wanting thereby indicates the incentive value of the stimulus and was assessed with the item asking "How much would you like to smell the odor again?," whereas pleasantness displays the hedonic aspect and was assessed by the question "How pleasant is this odor?" In addition, the mothers were asked if the presented odor belonged to their own child ("yes/no/I don't know)." Answers of the behavioral ratings are provided in **Table 2** .

### Data Analysis

Data was analyzed with SPM 12 (Wellcome Trust Center for Neuroimaging, London, UK, implemented in Matlab R2014b; MathWorks, Inc., Natick, MA, USA). The preprocessing was done identically for both designs with the default settings used in SPM 12 and comprised realignment with 2nd degree Bspline, unwarping with 4th degree B-spline, and co-registration by segmentation fitting to the individual T1 volume. The images used for analyses were spatially normalized (stereotactically transformed into MNI ICBM 152-space) and smoothed with a Gaussian kernel of 6 mm FWHM.

For the first level analyses, we started with the two sessions performed with the short design: The full 13 stimulation periods were contrasted to the full (13 × 6 s = 78 s) subsequent offperiod (13 × 19 s = 247 s, compare **Supplementary Figure 1**). We named this contrast "shortfull."

For both sessions performed with the long design, the whole on-period (8 × 15 s = 120 s) was contrasted to the whole subsequent off-period (8 × 30 s = 240 s). We named this contrast "longfull."

As the short design comprised more repetitions than the long design, we performed an additional analysis. In order to match the number of repetitions between both designs, we analyzed only the first 8 on- and off–blocks from both sessions. We named this contrast "shortreduced."

As the long design was characterized by longer stimulus delivery than the short design, an additional analysis was performed. In order to match the stimulation duration, only the first half of the on-period was used (8 × 7.5 s = 60 s) and compared to the whole subsequent off-period (compare **Figure 1**). We named this contrast "longreduced."

For the second level analyses, four t-contrasts with onesample t-tests were computed for the overall effect of the baby odor (on-period merged across own and unfamiliar baby vs. off period, clean air) for each design and analysis approach (shortfull, longfull, shortreduced, longreduced) in order to detect general activations related to the odor presentation across all subjects.

As the main aim of this study was not the determination of neural activations, but the exploration of the best suitable design characteristics, the comparison between both designs was based on the signal strength within a given ROI. ROI analyses were performed for the following regions: Anterior insula, OFC, amygdala, hippocampus, ACC, PCC, piriform cortex, and thalamus. ROIs were built with WFU Pick Atlas 3.0.3 (15) toolbox for SPM (for details, see **Supplementary Material**). ROI analyses were performed contrasting the effect of the baby body odor to the baseline condition. For each ROI in each design (apart from the ACC and the piriform cortex where no supra threshold activations were observed), the mean beta signal across all subjects was extracted for a 4 mm sphere around the peak voxel using MarsBar (16).

Subsequently, a generalized linear mixed model (GLM) was performed (IBM SPSS Statistics 25) in order to test the effect of the design on the signal strength. Each participant (n = 10) served as an individual, each stimulus (own and other baby) and each ROI (anterior insula, OFC, amygdala, hippocampus, PCC, thalamus) served as repeated measurement. The extracted mean beta signal was used as target for the main effect of the design across all ROIs.

We contrasted the new to the conventional design (shortfull vs. longfull). Afterwards, we systematically compared the different analysis approaches to each other in order to specify whether this effect was based on the number of repetitions, duration (length of stimulation period) or mode (continuous or pulsed stimulation) of the presentation. For effect sizes, we calculated Cohen's d. Results within the ROIs are descriptively reported.

In order to explore additional activations following baby odor stimulation, a whole-brain analysis was performed for the strongest design (shortfull). The effect of baby body odor (merged across own and unfamiliar baby) was contrasted to the baseline with a threshold of p < 0.001 (uncorrected) and a cluster extent threshold of k > 20 (**Supplementary Table 2**, **Supplementary Figure 1**). Analyses of the single effects of each baby body odor (own baby vs. baseline; unfamiliar baby vs. baseline) are presented in the **Supplementary Material** (compare **Supplementary Tables 3, 4**, **Supplementary Figure 2**).

# RESULTS

### ROI Analyses

There were superior BOLD signal activations in the short design compared to the long design across all ROIs [shortfull vs. longfull: F(1,22) = 8.67, p = 0.007, d = 0.34, see **Figure 1**]. We aimed to systematically compare whether this effect was based on number of repetitions, duration, or mode of presentation.

The comparison between the longfull to the shortreduced design indicated an effect of the number of repetitions: When both designs had the same number of repetitions, the short was not superior to the long design anymore [F(1, 13) = 1.74, p = 0.220). The comparison of the shortfull to the longreduced design indicated no effect of stimulation duration: When both designs had the same duration, the short was still superior to the long design [shortfull vs. longreduced: F(1, 159) = 15.61, p < 0.001, d = 0.24).

Thus, the observed superiority of the short design could be either due to number of repetitions or to the mode of presentation. In order to explore this further, we statistically compared the designs changing the parameter of interest (number of repetitions, mode, duration) and keeping the other two elements constant:

The direct comparison of number of repetitions, when keeping duration and mode constant, did not show a significant effect across the ROIs [shortfull vs. shortreduced: F(1, 18) = 0.01, p = 0.922]. Visual inspection revealed a differential effect: A high number of repetitions led to lower BOLD signal in amygdala and hippocampus, but to higher signal in secondary structures, namely the OFC and PCC (**Figure 1**).

The direct comparison of mode when keeping number of repetitions and duration constant, did not show a significant effect across all ROIs [shortreduced vs. longreduced: F(1, 21) = 1.59, p = 0.221]. Visual inspections showed a more differential effect, so that continuous presentation led to a higher signal in all ROIs except for the PCC and the anterior insula (**Figure 1**).

The direct comparison of duration when keeping number of repetitions and mode constant, did not show a significant effect across the ROIs [longfull vs. longreduced: F(1, 41) = 0.67, p = 0.419].

value (RH, right hemisphere; LH, left hemisphere, and MNI coordinates are displayed in square brackets) of each ROI and for each design across all subjects (n = 10). Please note in the anatomical visualization, that peak activations may have occurred on different hemispheres. Error bars display 95% CI.

Visual inspection showed—again—a more differential effect: a reduced duration of stimulation led to higher signal in amygdala, hippocampus and anterior insula, but to lower signal in the OFC (**Figure 1**).

### Whole Brain Analyses

Whole brain analysis was performed for the paradigm with the strongest neural activation (shortfull) and revealed rather weak responses in a total of four significantly activated areas, namely the superior temporal gyrus (STG), the OFC, the brain stem, and the anterior insula (compare **Supplementary Table 1**).

### DISCUSSION

Our results demonstrated superior activations in the short compared to the long design across ROIs. Systematic analyses revealed differential effects on olfactory areas depending on number of repetitions, duration, and mode of the stimulation. The clearest results were observed for the amygdala: for this structure, considered as part of the primary olfactory cortex (3), it seems beneficial to design body odor stimulation with fewer repetitions per run, shorter duration, and continuous presentation. We assume that this effect is due to the rapid habituation and adaptation in primary olfactory areas (17). To overcome the early habituation and preserve power, we suggest a higher number of short runs. Alternatively, stimulation with long and jittered inter-stimulus intervals can be recommended, though this will increase the total duration of the design.

For subsequent and later habituating structures, namely the OFC, many repetitions and long stimulation seem to be beneficial. Such an approach was implemented in the long design. However, great care has to be exerted in order to achieve a sufficient number of repetitions with this design. An optimal combination of long stimulation and high number of repetitions should be weighed. Based on our data we suggest 15 s of stimulation and at least 13 repetitions.

Taken together, our study showed diverse effects on different brain areas. A reduced stimulation duration for instance led to stronger signal in amygdala, hippocampus, and anterior insula, but to weaker signal in the OFC. This matches previous research showing that BOLD signal of hippocampus and anterior insula have similar time courses, while the BOLD signal time course of the OFC is delayed (17). The authors attributed this to the high interconnections, which result in similar patterns between the former structures. The OFC receives likewise direct input from primary olfactory areas (3). Additional incoming information via the thalamic pathway may explain its prolonged response (17). Hence, particular design characteristics should be considered with regard to the areas of interest.

A recent study (5) suggested a benefit of a high number of repetitions and short stimulation duration due to oscillations in the neural signal, which only occur after longer duration. Our study partly supports this assumption, as the combination of short and continuous stimulation with higher number of repetitions showed strongest activations. Yet, this effect could not be linked to the short duration, but rather to differential effects on primary or secondary structures depending on the respective combination of design characteristics.

The comparison in our study refers only to a block design. The short design was in fact similar to an event-related design in terms of short continuous stimulation alternating with rather long off-periods (8, 9). However, the stimuli were not randomized within a run; the stimulation was longer than in conventional event-related designs and on-off-periods alternated in the same interval. Further research comparing the short with a randomized and jittered design might be informative.

We are aware that the explanatory power of the study is limited due to the small sample size. However, we like to briefly review the additional results. Beyond the olfactory regions, presentation of baby body odors activated the PCC, as well as the STG. The PCC has been related to social chemosignaling (10), which matches our findings. As the STG is important for social cognition (18), the observed activation in our study might be referred to the social relevance of the baby odor stimuli.

The smell of the own baby is crucial for mother-child interactions and facilitates kin recognition and bonding in many species. In humans, higher reward-associated neural responses to baby body odors were observed in mothers compared to non-mothers (7) and it was suggested that maternal bonding is moderated by olfactory cues. The present study aimed to work out a suitable design for the detection of neural correlates to baby body odors. It provides the ground to examine the differences of neural processing of body odors from the own vs. other children.

# CONCLUSION

There is no common paradigm for the detection of neural correlates to body odor perception and the few studies performed in this area showed diverse results. The present study was conducted in order to find optimized design paradigms for presenting baby body odors in the fMRI and results may transfer to general body odor perception. As the short design revealed superior activations, we recommend this as a time-efficient and effective paradigm.

# AUTHOR CONTRIBUTIONS

IC, TH, and LS contributed to conception and design of the study. LS acquired the data. LS and IC performed the statistical analysis. LS wrote the first draft of the manuscript. IC wrote sections of the manuscript and TH critically revised the manuscript. All authors contributed to manuscript revision, read and approved the submitted version.

# FUNDING

This research was funded by the Deutsche Forschungs gemeinschaft (DFG) for the project (CR479/4-1). The impact of body odor on bonding and incest avoidance over the course of life: A developmental and neuropsychological approach.

# SUPPLEMENTARY MATERIAL

The Supplementary Material for this article can be found online at: https://www.frontiersin.org/articles/10.3389/fneur. 2018.01182/full#supplementary-material

# REFERENCES


**Conflict of Interest Statement:** The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

Copyright © 2019 Schäfer, Hummel and Croy. This is an open-access article distributed under the terms of the Creative Commons Attribution License (CC BY). The use, distribution or reproduction in other forums is permitted, provided the original author(s) and the copyright owner(s) are credited and that the original publication in this journal is cited, in accordance with accepted academic practice. No use, distribution or reproduction is permitted which does not comply with these terms.

# Functional Ultrasound Imaging of Spinal Cord Hemodynamic Responses to Epidural Electrical Stimulation: A Feasibility Study

Pengfei Song1†, Carlos A. Cuellar 2†, Shanshan Tang<sup>1</sup> , Riazul Islam<sup>2</sup> , Hai Wen<sup>2</sup> , Chengwu Huang<sup>1</sup> , Armando Manduca<sup>3</sup> , Joshua D. Trzasko<sup>1</sup> , Bruce E. Knudsen<sup>2</sup> , Kendall H. Lee2,3,4, Shigao Chen1,3 \* and Igor A. Lavrov 2,5,6 \*

<sup>1</sup> Department of Radiology, Mayo Clinic, Rochester, MN, United States, <sup>2</sup> Department of Neurologic Surgery, Mayo Clinic, Rochester, MN, United States, <sup>3</sup> Department of Physiology and Biomedical Engineering, Mayo Clinic, Rochester, MN, United States, <sup>4</sup> Department of Physical Medicine and Rehabilitation, Mayo Clinic, Rochester, MN, United States, <sup>5</sup> Department of Neurology, Mayo Clinic, Rochester, MN, United States, <sup>6</sup> Institute of Fundamental Medicine and Biology, Kazan Federal University, Kazan, Russia

### Edited by:

Jue Zhang, Peking University, China

### Reviewed by:

Jordi A. Matias-Guiu, Hospital Clínico San Carlos, Spain Gabriel Gonzalez-Escamilla, Johannes Gutenberg University Mainz, Germany Yukun Luo, PLA General Hospital, China

### \*Correspondence:

Igor A. Lavrov Lavrov.Igor@mayo.edu Shigao Chen Chen.Shigao@mayo.edu

†These authors share first authorship

### Specialty section:

This article was submitted to Applied Neuroimaging, a section of the journal Frontiers in Neurology

Received: 24 October 2018 Accepted: 04 March 2019 Published: 26 March 2019

### Citation:

Song P, Cuellar CA, Tang S, Islam R, Wen H, Huang C, Manduca A, Trzasko JD, Knudsen BE, Lee KH, Chen S and Lavrov IA (2019) Functional Ultrasound Imaging of Spinal Cord Hemodynamic Responses to Epidural Electrical Stimulation: A Feasibility Study. Front. Neurol. 10:279. doi: 10.3389/fneur.2019.00279 This study presents the first implementation of functional ultrasound (fUS) imaging of the spinal cord to monitor local hemodynamic response to epidural electrical spinal cord stimulation (SCS) on two small and large animal models. SCS has been successfully applied to control chronic refractory pain and recently was evolved to alleviate motor impairment in Parkinson's disease and after spinal cord injury. At present, however, the mechanisms underlying SCS remain unclear, and current methods for monitoring SCS are limited in their capacity to provide the required sensitivity and spatiotemporal resolutions to evaluate functional changes in response to SCS. fUS is an emerging technology that has recently shown promising results in monitoring a variety of neural activities associated with the brain. Here we demonstrated the feasibility of performing fUS on two animal models during SCS. We showed in vivo spinal cord hemodynamic responses measured by fUS evoked by different SCS parameters. We also demonstrated that fUS has a higher sensitivity in monitoring spinal cord response than electromyography. The high spatial and temporal resolutions of fUS were demonstrated by localized measurements of hemodynamic responses at different spinal cord segments, and by reliable tracking of spinal cord responses to patterned electrical stimulations, respectively. Finally, we proposed optimized fUS imaging and post-processing methods for spinal cord. These results support feasibility of fUS imaging of the spinal cord and could pave the way for future systematic studies to investigate spinal cord functional organization and the mechanisms of spinal cord neuromodulation in vivo.

Keywords: functional ultrasound, spinal cord, hemodynamic responses, spinal cord injury, ultrafast imaging, electrical stimulation

# INTRODUCTION

Over the last decades, epidural electrical spinal cord stimulation (SCS) was successfully implemented to help patients with chronic intractable pain (1–3). Meanwhile, SCS was reported as a promising alternative strategy to alleviate symptoms of motor impairments for multiple sclerosis (4, 5) and Parkinson's disease (6–9), and to improve motor (10–14) and

**31**

autonomic functions (15) in patients with spinal cord injury. The therapeutic effects of SCS rely on the stimulation parameters used (intensity, frequency, pulse width, burst vs. continuous stimulation, electrode configuration, etc.). At the same time, the mechanisms and neural structures through which SCS inhibits chronic pain and enables motor control remain unclear, although several hypotheses were supported by computational simulations (16–18) and data, primarily obtained from electrophysiological recordings (19, 20). Electromyography (EMG) is widely used as a diagnostic tool for neuromuscular disease and a research tool for disorders of motor control. However, the EMG signal is limited and can provide one-dimensional information concerning the activation of spinal cord neurons. In this context, a combination of emerging, innovative techniques providing high spatial and temporal resolution, and electrophysiology techniques could provide critical information on mechanisms of SCS and further facilitate optimizations of SCS protocols. Spatial and/or temporal resolution of available functional imaging tools, such as PET and MEG, are far below what is required for evaluation of the spinal cord functional changes during SCS. Although the spatial resolution of functional magnetic resonance imaging (fMRI) reaches submillimeter with ultra-high magnetic field (21, 22), the size of MR machine can be prohibitive for an intraoperative monitoring.

Functional ultrasound (fUS) imaging has the potential to complement these techniques at low cost. fUS is an emerging method that leverages the novel ultrafast plane wave imaging technique and the neurovascular coupling effect to monitor hemodynamic responses of tissue associated with neural activities (23). Ultrafast plane wave imaging allows acquisition and accumulation of ultrasound data at 10–20 kHz frame rate, significantly boosting the Doppler sensitivity to small vessels for fUS imaging (24–26). The rich spatiotemporal information of ultrafast plane wave data also allows implementation of more robust and intelligent tissue clutter filters based on singular value decomposition (SVD) (27–29), further improving the sensitivity of monitoring small vessel hemodynamic responses for fUS. In contrast to fMRI which responds to both hemodynamic and metabolic variations, fUS is only sensitive to hemodynamic effects (23, 30). Therefore, interpretations of fUS results are not confounded by the complex interactions between the hemodynamic and metabolic effects (31). As compared to other imaging techniques, fUS has higher spatial and temporal resolutions and also potentially can be performed on freely moving animals with miniaturized transducer size for long-term and real-time monitoring (32, 33). This opens new directions for potential applications of fUS, since currently there is no available technique that could evaluate functional changes in spinal cord in real-time in vivo. fUS could help in evaluation of hemodynamic response during electrode placement in order to optimize leads location for neuromodulation therapies and for intraoperative monitoring of spinal cord hemodynamics during surgical procedures. Finally, fUS may help to generate important information about spinal cord functional organization, and particularly, could help to trace circuitry response during pharmacological interventions and neuromodulation.

One disadvantage of fUS is that ultrasound cannot effectively penetrate through the bone. Therefore, fUS typically requires removal or thinning of the skull to access the targeted tissue such as brain (23, 31). Nevertheless, fUS has demonstrated promising results in monitoring a wide range of brain activities involved with visual, auditory, olfactory, and motor functions (23, 34–36), imaging brain intrinsic connectivity (37), and measuring brain activities of humans including neonates (38) and during surgery (39). A comprehensive review of current preclinical and clinical applications of fUS was recently published in (40).

To the best of our knowledge, this is the first attempt of implementing fUS to study the effect of spinal cord stimulation in animal models. Here we present a methodology and work flow, including the optimized subpixel motion registration, SVD-based clutter filtering, and hemodynamic response quantification, to validate the feasibility of using fUS to examine the SCS response. The capability of the proposed work flow was tested on two species (rat and swine). Specific spinal cord hemodynamic responses associated with different SCS parameters were evaluated, including different voltages, and stimulation patterns.

### MATERIALS AND METHODS

Experimental procedures were approved by the Mayo Clinic Institutional Animal Care and Use Committee. The National Institutes of Health Guidelines for Animal Research (Guide for the Care and Use of Laboratory Animals) were observed rigorously. Animals were kept in controlled environment (21◦C, 45% humidity) on a 12-h light/dark cycle.

### Rat Study Procedure

Sprague-Dawley rats (3 males, 325–350 gr, ad libitum access to water and food) were anesthetized with isoflurane (1.5–3%). Laminectomies were performed at T13-L2 and the spinal cord was exposed. Two Teflon coated stainless steel wires were placed at T13 and L2 and sutured on dura (corresponding approximately to L2 and S1 segments of the spinal cord). Small windows were opened between T11-L12 and L3-L4 allowing wires to be passed under the T12 and L3 vertebrae. A small notch (0.5 mm) facing the spinal cord was made on the Teflon coating, serving as the stimulating electrode. Breathing motion was minimized by fixing the spine using a custom-made frame composed of a clamp holding the Th12 spinous process and two pieces retracting back muscles on both sides. Additionally, two rods were secured over the coxal bones in order to hold up the pelvic girdle. Dorsal skin flaps were attached around the frame to form a pool facilitating transducer positioning (**Figure 1**). SCS consisted of 0.5 ms squared pulses delivered at 40 Hz in monopolar or bipolar configurations. Two reference electrodes were inserted bilaterally in back muscles. EMG signals were recorded using dual needle electrodes (Medtronic, Memphis, TN) inserted bilaterally in tibialis anterior (TA) and gastrocnemius (GAS) hind limb muscles. Warm saline solution (1.5 ml) was administered S.C. every 2 h. At the end of the experiment, animals were euthanized using pentobarbital (150 mg/kg I.P.).

# Swine Study Procedure

A domestic white swine (male, 8 weeks old, 25 kg, ad libitum access to water, fed once daily) was initially anesthetized using a mixture of telazol (5 mg/kg) and xylazine (2 mg/kg I.V.). Anesthesia was maintained using isoflurane (1.5–3%). For analgesia, fentanyl (2–5 mg/kg/h) was administered throughout the experiment. Similar surgical procedures as described in the previous section were performed in swine (41). Two Teflon stainless steel wires were placed onto L4 and L5-L6 and sutured on dura after laminectomies were performed at L1-L6. Back muscles were retracted and the spine stabilized using 4 blunt tip rods that attached the spine to a custom-made frame. SCS was delivered at 40 Hz, 0.5 ms pulse width in bipolar configuration. A reference electrode was inserted in the back muscles. Needle electrodes (Medtronic, Memphis, TN) were inserted bilaterally in TA and GAS hind limb muscles to monitor EMG responses during SCS. At the end of the experiment, the subject was euthanized (sodium pentobarbital 100 mg/kg I.V.).

### fUS Imaging Setup

A Verasonics Vantage ultrasound system (Verasonics Inc., Kirkland, WA) and a Verasonics high frequency linear array transducer L22-14v (Verasonics Inc., Kirkland, WA) with center frequency of 15 MHz were used in this study. **Figure 1** shows the fUS imaging setup. The fUS transducer was positioned on the spinal cord between the rostral and caudal electrodes. An imaging field-of-view (FOV) was carefully selected to align with the longitudinal dimension of the spinal cord and intersect with the central canal (**Figure 1B**). The position of the fUS transducer was fixed throughout the study. A thin layer of mineral oil was added between the fUS transducer and the spinal cord for acoustic coupling.

An ultrafast compounding plane wave imaging-based fUS imaging sequence was developed for the study. As shown in **Figure 2A**, five steered plane waves (−4 to 4◦ , with 2◦ of step angle) were transmitted with each steering angle repeatedly transmitted three times to boost signal-to-noise-ratio (SNR). This compounding scheme has an equivalent SNR performance to a conventional 15-angle compounding sequence, but reduces the beamforming computational cost by a factor of 3 (32). The pulse repetition interval was 35 µs (corresponding to a pulse repetition frequency (PRF) of 28.6 kHz), and the total time cost for transmitting and receiving all 15 transmissions was 525 µs. To satisfy a post-compounding PRF of 500 Hz, a 1,475 µs no-op time was added to each group of compounding transmissions (**Figure 2A**). After coherent compounding (24), high quality ultrasound data was obtained (**Figure 2B**) and used as Doppler ensembles for future processing. A total of 200 Doppler ensembles (400 ms duration) were collected within each second to produce one power Doppler (PD) image per second (**Figure 2C**). For the rat experiment, a total of 120 s of fUS data was collected (corresponding to 24,000 frames of high framerate ultrasound data) for each trial of SCS, including 30 s of baseline measurement, 20 s of ES measurement, and 70 s of recovery measurement. Five trials were repeated for each SCS configuration. For the swine experiment, a total of 30 s of fUS data was collected (6,000 frames), including 5 s of baseline, 15 s of stimulation, and 10 s of recovery. Five trials were repeated for each SCS configuration.

For data synchronization with the SCS and EMG measurements, the Verasonics system was programmed to send a trigger-out signal at the beginning of each second when the first steered plane wave was transmitted. The trigger-out signal was recorded together with the SCS and EMG signals for post-processing.

# fUS Post-processing Steps

### Motion Correction

To facilitate accurate fUS measurements of hemodynamic responses, we developed a robust and fast sub-pixel motion correction algorithm to remove tissue motion induced by breathing and SCS. Motion correction was applied both on the original high frame-rate ultrasound data before clutter filtering (e.g., **Figure 2B**), and on the PD images after clutter filtering (e.g., **Figure 2C**). The motion correction method was based on the principles of phase correlation-based sub-pixel registration introduced in (42). Briefly, the method by Foroosh et al. (42) derived an analytical solution of the phase correlation function between images that are shifted by non-integer number of pixels

(1x, 1z), and presented a method of using the main peak and side peaks of the inverse Fourier transform of the phase correlation function (C) to calculate the sub-pixel displacement:

$$
\Delta x = \frac{\text{C(1,0)}}{\text{C(1,0)} \pm \text{C(0,0)}}
$$

$$
\Delta z = \frac{\text{C(0,1)}}{\text{C(0,1)} \pm \text{C(0,0)}}\tag{1}
$$

where C(0,0) indicates the main peak (i.e., location of the pixel with highest positive pixel value) and C(1,0) and C(0,1) indicates the side peaks (i.e., location of the pixel with second highest positive pixel value) along x-dimension and z-dimension, respectively. To improve the robustness of Equation (1) for ultrasound applications, we added additional measurements of 1x ′ and 1z ′ using the main peak and side peaks with highest negative pixel value:

$$
\Delta x' = \frac{\text{C}(-1, 0)}{-\text{C}(-1, 0) \pm \text{C}(0, 0)}
$$

$$
\Delta z' = \frac{\text{C}(0, -1)}{-\text{C}(0, -1) \pm \text{C}(0, 0)}\tag{2}
$$

Then an average sub-pixel displacement was calculated using the results from Equations (1) and (2). Other available sub-pixel motion estimation algorithm, such as the one presented in (43) and the normxcorr2.m function in MATLAB, require heavy upsampling of ultrasound signals in order to measure the subpixel motion between frames. In fUS imaging, this up-sampling procedure is extremely computationally expensive due to the large amount of ultrasound data acquired in temporal dimension. In contrast, the sub-pixel motion estimation algorithm used in this study does not require up-sampling and involves Fourier transform, which can be executed at extremely fast speed. Therefore, the computational cost can be greatly reduced with the method used in this study.

To further improve the robustness of sub-pixel displacement estimation and suppress false calculations, as shown in **Figure 3**, a tissue velocity curve (**Figure 3B**) was first derived by taking a derivative of the original displacement curve (**Figure 3A**). Then a tissue velocity thresholding (cutoff was determined empirically as 2 mm/s for this study) was applied to the velocity curve to reject high speed values, followed by an integral calculation to recover the displacement curve (**Figure 3C**). False displacement could be effectively removed by this process. This additional step was only

FIGURE 3 | (A) Original displacement curve with false displacement calculations. (B) Taking the derivative (i.e., velocity) of the displacement curve, and applying a tissue velocity threshold. (C) Integral of the velocity curve after rejection of large tissue velocities to remove the false displacement calculations.

applied to the original high frame-rate ultrasound data, not to the PD images.

Finally, to avoid creating the streaking artifacts associated with applying a phase-shift to the Fourier spectrum (due to bandlimited data), the gridded data interpolation (e.g., "griddedInterpolant.m" function in Matlab) was used to register the moved ultrasound frames.

### Tissue Clutter Filtering

The spatiotemporal SVD-based ultrasound clutter filter was used in this study to suppress tissue clutter and extract micro-vessel signals (27–29). Here we used the combination of an accelerated SVD method (44) and a noise equalization technique (45) for tissue clutter filtering. For the first 200 ultrasound ensembles in each trial, a full SVD was calculated to determine a low-cutoff singular value threshold for tissue rejection (28) and derive a noise field for noise equalization (45). The same low-cutoff value and noise field were used for the rest of the ultrasound data in the trial. **Figure 4** shows the PD images after the motion correction and the clutter filtering process for the rat spinal cord (**Figure 4A**) and the swine spinal cord (**Figure 4B**).

### Spinal Cord Hemodynamic Response Calculation and Measurement

Ultrasound Power Doppler signal measures the backscattering power of the moving blood, which reflects the blood volume at the interrogated location (e.g., each imaging pixel) (46). Here we define the spinal cord blood volume change (1SCBV) as the percentage of power Doppler (PD) signal variation compared to the baseline:

$$
\Delta SCBV = \frac{PD\_{stim} - PD\_{baseline}}{PD\_{baseline}} \times 100\%
$$

A Savitzky-Golay smoothing filter (47) (window length = 11, order = 1) was applied to the 1SCBV measurement for each imaging pixel along the temporal direction to remove noise. 1SCBV measurements with amplitude smaller than twice the standard deviation of the baseline fluctuations were rejected. The remaining 1SCBV measurements were

color-coded and superimposed on the PD images (**Figure 5A**, **Supplemental Videos 1**, **2** for spinal cord hemodynamic response with and without SCS).

For quantitative local 1SCBV measurements, four regionsof-interest (ROIs) were selected for the rostral-dorsal, rostralventral, caudal-dorsal, and caudal-ventral sections of the spinal cord (**Figure 5B**). For each section, the average 1SCBV was calculated using all pixels inside the ROI for each time point. Then the five 1SCBV curves from the five repeated SCS trials were averaged and smoothed (by Savitzky-Golay filter with 5th order and 21-sample window length) for quantitative measurements, as indicated by the blue and the orange curve in **Figure 5C**, respectively. Four parameters including the peak response, ascending slope of the response curve (i.e., response rate), area under the response curve (AURC), and the recovery time were derived from the 1SCBV curve. For the response rate, a linear fitting was performed on the ascending portion of the 1SCBV curve to calculate the slope (indicated by the yellow curve in **Figure 5C**). To determine the end point of the SCS response and spinal cord recovery, a linear fitting was performed on the descending portion of the 1SCBV curve, and the point where the fitted line intersects with the zero 1SCBV axis was

used as the end recovery point (indicated by the cross sign in **Figure 5C**). The time interval between peak response and end recovery point was calculated as the recovery time. Finally, the total area under the curve between the onset of SCS and the end recovery point was calculated as AURC, which reflects the total spinal blood volume variations within the imaging FOV in response to SCS.

# RESULTS

# Effect of SCS on Spinal Cord Hemodynamic Change vs. Muscle Neuro Electrophysiological Change

**Figure 6** shows the spinal cord hemodynamic responses to SCS on a rat model (rat #1) with different stimulation voltages (1.8 and 1.0 V) at 40 Hz SCS frequency. SCS at 1.8 V produced a clear EMG response reflected in the hemodynamic response maps and response curve (**Figures 6A,C,D**, and **Supplemental Video 3**). On the other hand, 1.0 V SCS did not produce a visible EMG response and only a weak response curve was observed primarily in dorsal part of the spinal cord (**Figures 6B–D**, and **Supplemental Video 4**). From these results, one can clearly see that higher SCS voltages produced stronger spinal cord hemodynamic responses. **Figure 7** shows that all quantitative spinal cord response measurements at different sections were decreased with stimulation at lower voltage. At the same time, for both 1.8 and 1.0 V of stimulation hemodynamic changes were higher at the dorsal compared to the ventral part of the spinal cord. Increasing SCS voltage also increased hemodynamic responses in the ventral parts of the spinal cord across

different segments, which correlates with the EMG observations in **Figure 6C**.

A gradually increased SCS voltage, from 0.4 to 1.2 V, was applied to another rat (rat #3). **Supplemental Figure 1** shows the monotonic and linear relationship between the measured 1SCBV and 1EMG at different SCS voltages. 1EMG denotes the increase in root-mean-square (RMS) of EMG signal during stimulation compared to its baseline. In our experiments we observed that different rats had different tolerance and reaction threshold to electrical stimulation. Even for the same rat, the reaction threshold could also vary with different stimulation frequency and electrode configuration. Results presented in **Supplemental Figure 1** was collected from a different rat to the results in **Figure 6**, therefore distinct voltages were used.

# Spatial Analysis of SCS Evoked Spinal Cord Hemodynamic Response

**Figure 8** shows the quantitative spinal cord hemodynamic responses to SCS categorized by different sections of the spinal cord. The main difference in hemodynamic changes with SCS was found between activation of the dorsal and ventral part of the spinal cord with higher activity in the dorsal part across all tested segments. The difference between rostral and caudal hemodynamics was less prominent, with higher hemodynamic response on rostral segments (where the electrode was placed). These results are in agreement with observations in **Figures 6A,B**, where the rostral-dorsal section of the spinal cord had the highest blood volume increase during the stimulation.

# Spinal Cord Hemodynamic Response to Patterned SCS

**Figure 9** shows the results of fUS monitoring of spinal cord response under a patterned SCS (rat #2). The patterned SCS consists of three ON-OFF SCS cycles, with each cycle containing a 20-s ON period and a 10-s OFF period with the SCS frequency 40 Hz and amplitude 0.6 V in bipolar configuration (**Figure 9A**). Compared to the result in **Figure 6**, a lowered stimulation voltage was used here, as the motor response threshold was different among animals and with varied SCS parameters and electrode configurations. From **Figure 9B**, one can clearly see the variations of spinal cord blood volume following the ON-OFF pattern of SCS. Inadequate recovery time was given between consecutive SCS periods, and consequently the spinal cord blood volume could not return to baseline value until the patterned SCS was OFF. Simultaneous EMG response is shown in **Figure 9C**. **Supplemental Video 5** shows one representative movie of the patterned SCS response in a rat model.

# Feasibility Study on Swine Model

**Figure 10** shows the results of the effect of SCS on hemodynamic changes in the swine spinal cord. A 40 Hz bipolar stimulation was used with a stimulation voltage of 10 V. Higher stimulation voltage was used in the swine model compared to the rat model due to differences in SCS thresholds for these two species. **Supplemental Video 6** shows the movie of the swine spinal cord response. Similar to the results observed in the rat study, the swine spinal cord showed well-correlated hemodynamic responses to the SCS. As shown in **Figure 10** and **Supplemental Video 6**, similar to the rat study, the dorsal spinal cord had significantly higher blood volume increase than the ventral spinal cord.

# DISCUSSION

An optimized work flow of using fUS to map local spinal cord hemodynamic response during epidural electrical stimulation was presented in this article. The proposed methodology was applied on two animal species for feasibility and capability validation. Although not a systematic study, the preliminary results presented here demonstrated great potential of fUS

in monitoring and evaluating the spinal cord's hemodynamic response during epidural electrical stimulation in vivo.

In order to save the computational cost associated with motion correction, the sub-pixel motion registration algorithm was used in this study. This fast algorithm cannot correct for non-rigid tissue motion which may occur in in vivo studies. This may result in residual tissue motion that may cause false spinal cord response measurements which produces fluctuations of fUS-measured spinal cord response.

In this study, we investigated the spinal cord hemodynamic response which was compared with electrophysiological measurements during spinal cord epidural stimulation. Compared to other functional imaging techniques, fUS provides superior spatiotemporal resolutions that allow investigation of local spinal cord responses even in small models like rat and monitoring the time-varying spinal cord responses evoked by SCS. Our data also suggest that fUS is a more sensitive technique than commonly used electrophysiological assessment such as EMG and can evaluate subthreshold to motor response level of SCS.

The main objective of this study was to test the feasibility and capability of using fUS to examine the epidural stimulation evoked specific changes in spinal cord hemodynamics, measured in the lumbosacral spinal cord segments. During in vivo experiments in small (rat) and large (swine) animal models, epidural stimulation produced significant blood volume changes in spinal cord with clear specificity to the different areas of the spinal cord. Specific anatomical organization of the spinal cord vasculature with anterior and posterior spinal arteries divides the spinal cord into two areas, providing relatively independent blood supply for ventral and dorsal parts of

the spinal cord (48–51). This difference between dorsal and ventral parts, although evident from anatomical studies, to our knowledge has not been correlated with the functional organization of the spinal cord until now. Comparison between right and left side of the spinal cord (rostral vs. caudal regions) was also important to assess the level of asymmetry in activation of spinal cord afferents, which could be functional or related to anatomical position of the electrode on the spinal cord.

In order to provide good control over the position of the fUS transducer and to reduce motion artifacts, this study was conducted on anesthetized animals. Accordingly, our current findings cannot reflect the full spectrum of spinal cord responses that can be observed in awake animals. For example, isoflurane anesthesia, used in this study, could affect vascular response by causing vasodilation (52).

One limitation of fUS is the motion artifacts induced by physiologic activities such as breathing and movement, which could affect data collection and may require sophisticated stabilization of the vertebral column and mechanical isolation from the muscles. Another limitation is direct placement of the fUS transducer on the spinal cord, since ultrasound cannot penetrate the vertebra, which is an obstacle for this technique in clinical translation. However, non-invasive fUS with microbubble-enhanced Power Doppler technique has been reported recently (40, 53), where fUS could be performed with intact skull bone. This non-invasive form of fUS imaging can be adopted and evaluated for spinal cord imaging in the future. Also, this limitation of removing vertebra could be potentially solved with miniaturization of the devices and development of implantable transducers.

Current information on spinal cord functional organization is primarily comes from electrophysiology experiments with intracellular or extracellular recordings or based on activity recorded in selected nerves or muscles. Using these approaches previous studies showed that spinal circuitry is highly sensitive to different modalities of afferent information, which determines immediate and long-term changes and complex mechanisms such as plasticity and neuroregeneration (54–56). Studies performed on acute decerebrated cats (57) suggest that epidural vs. intraspinal stimulation can activate different spinal cord networks with important role of sensory information in their modulation. The extensive convergence of afferent information on different types of neurons produces significant limitations in understanding of spinal circuitry organization with available electrophysiological tools in real-time (58, 59). Evaluation of spinal cord hemodynamic changes with fUS is a novel and highly sensitive tool that could help to provide information about realtime spinal cord activity across multiple segments and improve our understanding the spinal cord functional organization in vivo. As a proof-of-concept work, this study was only performed on a small and a large animal model. Massive and thorough investigations will be conducted in the future to explore the potentials of clinical translation.

### CONCLUSIONS

The importance of understanding the physiological and pathological mechanisms of the spinal cord hemodynamic regulation is critical for diagnostics, for clinical monitoring, and for developing novel therapies and new rehabilitation protocols. The results of the present study indicate that epidural stimulation can cause spinal hemodynamic changes related to complex neuronal activity of spinal circuitry in both small and large animal models. This study presents the first implementation of fUS to explore functional organization of the spinal cord hemodynamics and provides results on correlations between SCS induced neural activities and local hemodynamics changes. The fUS measurements indicate temporal and spatial resolutions not achievable by other electrophysiological methods. Future studies on modulation of neuronal activity and hemodynamic response with spinal cord stimulation will help to address critical questions about spinal cord functional organization in intact spinal cord and its acute and chronic changes related to different pathological conditions.

### DATA AVAILABILITY

All datasets generated for this study are included in the manuscript and/or the **Supplementary Files**.

### AUTHOR CONTRIBUTIONS

PS, CC, SC, RI, KL, and IL designed the experiment. PS, CC, ST, RI, and IL drafted the manuscript. PS, CC, RI, and CH collected experiment data. PS, ST, AM, and JT wrote the

### REFERENCES


algorithms for data processing. CC, RI, HW, and BK conducted the animal surgeries. All authors reviewed and participated in editing the manuscript.

# FUNDING

Research reported in this publication was supported in part by the Minnesota State Office for Higher Education Spinal Cord Injury and Traumatic Brain Injury Research Grant Program (FP00098975 and FP00093993), by the subsidy allocated to Kazan Federal University for the state assignment in the sphere of scientific activities (no. 17.9783.2017/8.9), and the National Cancer Institute (NCI) of the National Institutes of Health (NIH) under Award Number K99CA214523. The content is solely the responsibility of the authors and does not necessarily represent the official views of the National Institutes of Health.

### SUPPLEMENTARY MATERIAL

The Supplementary Material for this article can be found online at: https://www.frontiersin.org/articles/10.3389/fneur. 2019.00279/full#supplementary-material

Supplementary Figure 1 | Spinal cord hemodynamic response and EMG to a gradient voltage.

Supplementary Video 1 | Five trials of spinal cord hemodynamic response to electrical stimulation on a rat model.

Supplementary Video 2 | Silent spinal cord hemodynamic response to an OFF electrical stimulation on a rat model.

Supplementary Video 3 | Spinal cord hemodynamic response to a 40 Hz, 1.8 V, Monopolar electrical stimulation on a rat model.

Supplementary Video 4 | Spinal cord hemodynamic response to a 40 Hz, 1.0 V, Monopolar electrical stimulation on a rat model.

Supplementary Video 5 | Spinal cord hemodynamic response to a patterned electrical stimulation on a rat model.

Supplementary Video 6 | Spinal cord hemodynamic response to a bipolar electrical simulation on a swine model.


study of compound muscle action potentials. Spinal Cord. (2004) 42:401–16. doi: 10.1038/sj.sc.3101615


data highly increases doppler and fultrasound sensitivity. IEEE Trans Med Imaging. (2015) 34:2271–85. doi: 10.1109/TMI.2015. 2428634


Ferroelectr Freq Control. (2017) 64:1776–81. doi: 10.1109/TUFFC.2017. 2748387


adaptive states after complete loss of supraspinal input. Nat Neurosci. (2009) 12:1333–42. doi: 10.1038/nn.2401


**Conflict of Interest Statement:** The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

Copyright © 2019 Song, Cuellar, Tang, Islam, Wen, Huang, Manduca, Trzasko, Knudsen, Lee, Chen and Lavrov. This is an open-access article distributed under the terms of the Creative Commons Attribution License (CC BY). The use, distribution or reproduction in other forums is permitted, provided the original author(s) and the copyright owner(s) are credited and that the original publication in this journal is cited, in accordance with accepted academic practice. No use, distribution or reproduction is permitted which does not comply with these terms.

# Why is Clinical fMRI in a Resting State?

### Erin E. O'Connor\* and Thomas A. Zeffiro\*

Department of Diagnostic Radiology and Nuclear Medicine, University of Maryland Medical Center, Baltimore, MD, United States

While resting state fMRI (rs-fMRI) has gained widespread application in neuroimaging clinical research, its penetration into clinical medicine has been more limited. We surveyed a neuroradiology professional group to ascertain their experience with rs-fMRI, identify perceived barriers to using rs-fMRI clinically and elicit suggestions about ways to facilitate its use in clinical practice. The electronic survey also collected information about demographics and work environment using Likert scales. We found that 90% of the respondents had adequate equipment to conduct rs-fMRI and 82% found rs-fMRI data easy to collect. Fifty-nine percent have used rs-fMRI in their past research and 72% reported plans to use rs-fMRI for research in the next year. Nevertheless, only 40% plan to use rs-fMRI in clinical practice in the next year and 82% agreed that their clinical fMRI use is largely confined to pre-surgical planning applications. To explore the reasons for the persistent low utilization of rs-fMRI in clinical applications, we identified barriers to clinical rs-fMRI use related to the availability of robust denoising procedures, single-subject analysis techniques, demonstration of functional connectivity map reliability, regulatory clearance, reimbursement, and neuroradiologist training opportunities. In conclusion, while rs-fMRI use in clinical neuroradiology practice is limited, enthusiasm appears to be quite high and there are several possible avenues in which further research and development may facilitate its penetration into clinical practice.

### Keywords: rs-fMRI, network, individuals, FDA, CPT code, ASFNR, survey

# INTRODUCTION

Techniques for quantifying spatial and temporal brain activity have developed rapidly since the first demonstrations that MRI could be used to measure modulations in blood oxygen level dependent (BOLD) tissue contrast (1). The observation that MRI could be used to monitor temporally correlated low-frequency activity fluctuations in spatially remote brain areas led to widespread use of resting state functional magnetic resonance imaging (rs-fMRI) to evaluate resting state network (RSN) properties. While BOLD-contrast is an indirect measure of neural activity, similar inter-regional coherent spontaneous neural activity correlations have been observed with electrophysiological techniques (2), suggesting that rs-fMRI networks can provide useful information about the macroscopic organization of neural processing systems. The methods and possible uses of rs-fMRI have recently been reviewed (3, 4).

Establishing that rs-fMRI can identify spontaneous brain activity patterns resembling those seen with tasks (5) has led to its widespread acceptance, and a rapid expansion in rs-fMRI publications. Nearly 10,000 rs-fMRI papers are currently listed in PubMed. The most rapidly developing type of functional connectivity research involves investigations of disease-related group differences

### Edited by:

Hongyu An, Washington University in St. Louis, United States

### Reviewed by:

Haris Iqbal Sair, Johns Hopkins University, United States Alessandro Tessitore, Università degli Studi della Campania Luigi Vanvitelli Caserta, Italy

### \*Correspondence:

Erin E. O'Connor erin.oconnor@umm.edu Thomas A. Zeffiro zeffiro@neurometrika.org

### Specialty section:

This article was submitted to Applied Neuroimaging, a section of the journal Frontiers in Neurology

Received: 24 January 2019 Accepted: 05 April 2019 Published: 24 April 2019

### Citation:

O'Connor EE and Zeffiro TA (2019) Why is Clinical fMRI in a Resting State? Front. Neurol. 10:420. doi: 10.3389/fneur.2019.00420

**44**

in brain network structure, enabled by the relative simplicity of data collection from large samples. As a result, atypical restingstate connectivity has been demonstrated in a wide range of neuropsychiatric disorders, including epilepsy, schizophrenia, attention deficit hyperactivity disorder, Alzheimer's disease, stroke, and traumatic brain injury.

rs-fMRI has several advantages over task-fMRI in clinical contexts. First, data acquisition is less complex. Second, if mapping multiple neural systems is needed, rs-fMRI can identify them simultaneously, saving time. Finally, rs-fMRI can be performed in individuals unable to cooperate for fMRI tasks, such as young, sedated, paralyzed, comatose, aphasic, or cognitively impaired patients. In addition to its utility in detecting changes in group network properties, rs-fMRI can also be used to detect individual differences (6–10).

Although the first reports of rs-fMRI clinical applications appeared 10 years ago (11), rs-fMRI use in clinical neuroradiology practice remains in a nascent stage, limited mainly to pre-surgical planning (4, 12) and is typically performed in conjunction with task-fMRI. Given the rapid rise and widespread use of rs-fMRI in neuroimaging clinical research, it might be expected that rs-fMRI would already be widely used in clinical practice, particularly in academic centers. Nevertheless, this is not the case and the reasons for the relatively weak penetration of rs-fMRI methods into neuroradiology practice are not entirely clear. To determine attitudes toward the use of rs-fMRI use in neuroradiology research and practice, we recently queried the American Society for Functional Neuroradiology (ASFNR) membership. In this article we will discuss the results of this survey, covering opinions about the current state of rs-fMRI acquisition, analysis, and interpretation methods. We then address existing barriers to using rs-fMRI in clinical practice and propose possible solutions, presenting examples of typical group and individual subject rs-fMRI analyses using public domain data.

### METHODS

After obtaining a human subjects research exemption, an invitation to participate in a 20 item electronic survey was sent to ASFNR members to collect information concerning their use of rs-fMRI in clinical research and practice, demographics, and work environment. Responses were collected using 5-point Likert items and deidentified prior to analysis.

Because a majority of respondents expressed concerns that substantial analysis and interpretation problems need to be solved before rs-fMRI can be widely used in clinical practice, we next explored examples of typical rs-fMRI analysis variations using the publicly available NYU CSC TRT dataset (www. nitrc.org/projects/nyu\_trt), processed using the CONN Toolbox (13), a popular open-source rs-fMRI analysis program (www. nitrc.org/projects/conn). In one example, we explored the serial influence of time series preprocessing algorithms on language network detection using an inferior frontal gyrus ROI. Effects of applying global signal regression, incorporating head motion estimates, using anatomical CompCorr, and outlier elimination were examined in a group level analysis of 25 healthy participants. Next, we explored the effects of denoising on single participant data. The exercise revealed large effects that processing variations can have on the detection of domain-specific maps at the group or single-subject level. These results are presented in the discussion of existing barriers related to increasing rs-fMRI use in clinical practice.

## RESULTS

The response rate was 24% (71/294). Of these, the majority were involved in both clinical and research activities. Twentyone percent were female. Eighty-seven percent held MD, MBBS, or MD PhD degrees; the others were PhDs. Only two of the respondents were exclusively involved in research. The median time since training was 12 years.

Ninety-two percent of the ASFNR respondents reported having adequate MRI equipment to conduct rs-fMRI and 82% indicated that rs-fMRI was relatively easy to collect.

Eighty percent reported using task-fMRI and 59% reported using rs-fMRI in their past research. Seventy-two percent reported plans to use rs-fMRI for research in the next year. Yet, only 40% agreed, or strongly agreed, that they would use rs-fMRI in clinical practice in the next year. Eighty-two percent of respondents agreed or strongly agreed that task-fMRI and rs-fMRI clinical use are largely confined to pre-surgical planning, mentioning seizure focus detection as other promising application. Thirty-two percent agreed, or strongly agreed, that rs-fMRI is currently useful in pre-surgical planning and 68% agreed, or strongly agreed, that it will be useful in future surgical planning (**Supplement Table 1**).

While respondents expressed strong interest in rs-fMRI clinical applications, they expressed concerns that may explain its lack of penetration into clinical practice. For example, 66% agreed, or strongly agreed, that rs-fMRI data are difficult to analyze. Twenty-four percent expressed concern about the reliability and reproducibility of rs-fMRI in identifying canonical brain networks. Seventy-seven percent agreed, or strongly agreed, that there are substantial analysis problems to be solved before rs-fMRI can be widely used in clinical practice. In addition, 77% agreed, or strongly agreed, that there are substantial interpretation problems to be solved before rs-fMRI can be widely used in clinical practice (**Figure 1**).

### DISCUSSION

In summary, while most respondents had experience with fMRI in both clinical and research contexts, have adequate MRI systems at their institutions and are relatively enthusiastic about incorporating rs-fMRI into clinical protocols, a number of concerns appear to be slowing the translation of rs-fMRI from research to practice. Some barriers to rs-fMRI implementation

**Abbreviations:** BOLD, blood oxygen level dependence; rs-fMRI, resting state functional magnetic resonance imaging; ASFNR, American Society of Functional Neuroradiology; RSNs, resting state networks; ICA, independent component analysis; ROI, region of interest.

in clinical practice, and possible ways to circumvent them, are addressed below.

### Barrier 1: Precision Medicine Agenda

Functional MRI used in research settings typically averages participant data in order to detect differences in regional task effects between clinical and healthy groups. In clinical medicine, however, diagnostic inferences and treatment recommendations are made for single cases.

As most publications describe acquisition and analysis methods optimized to detect between-group effects, better methods to characterize rs-fMRI maps in individuals are needed. Acquisition technology advances, such as higher magnetic field strength, multi-channel coils, and faster image acquisition have led to substantial sensitivity improvements, making the study of individual resting state networks possible (14).

One simple way to improve network detection sensitivity is to lengthen scan time. While some canonical RSNs, such as the default mode or sensorimotor networks, can be reliably detected at the group level using 5–6 min scans, longer sampling times, on the order of 12–30 min, can substantially improve detection of networks exhibiting lower average connectivity (15, 16). Since rs-fMRI data is dominated by physiological noise, longer sampling times with short TRs allow more effective physiological denoising and more sensitive neural signal detection. While most analysis techniques assume static connectivity effects between pairs of network nodes, dynamic connectivity estimates can benefit even more from longer sampling times. Dynamic connectivity analysis, while relatively new to rs-fMRI, holds promise in providing quantitative estimates of timevarying connection phenomena that may be altered in brain disease (17).

Variance in intrinsic connectivity contributed by cognitive state and mood, rather than disease effects, may be responsible for individual network structure variation (18). Nevertheless, moderate-to-high test-retest reliability of rs-fMRI indices challenges these concerns (19). In addition, longer sampling times, as discussed above, can facilitate detection of individual static network structure in the face of moderate dynamic variations in connectivity.

While rs-fMRI is currently being used for preoperative planning in a few centers (20), other clinical applications are not as common. High within-subject reproducibility of RSNs suggests that they might serve as biomarkers for monitoring disease progression in individual patients (21).

Finally, tools comparing individual to group maps are needed. Structural templates based on normative data sets that take into account age, sex, magnetic field strength, and data quality have been developed (22). Standardizing rs-fMRI acquisition protocols, then collecting normative comparative data, would greatly facilitate rs-fMRI clinical use by allowing comparison of individuals to age, sex, and IQ adjusted norms. For example, a clinically relevant target, the left hemisphere language network, when identified using a left inferior gyrus ROI, exhibits substantial between-subject variability, even when averaging across three collection sessions (**Supplement Figure 1**). Of greater concern is the fact that the majority of patients referred for pre-surgical mapping have space occupying lesions that distort both local and global anatomy, making mapping to standard anatomical spaces difficult or impossible using conventional spatial normalization techniques. Moreover, slowly growing tumors may dynamically alter inter-regional connectivity, making comparisons to functional group maps derived from healthy participants difficult to interpret. In pre-surgical planning, precisely determining the details of how an individual patient's functional anatomy differs from a typical spatial distribution may be important in determining treatment recommendations.

### Barrier 2: Diversity of Measures

Numerous methods can characterize regional intrinsic connectivity, including ROI->ROI correlations, ROI->voxel correlations, independent component analysis (ICA) of canonical networks, dynamic functional connectivity analysis, and graph theory analysis [see (3, 23) for recent reviews]. These different connectivity modeling techniques may measure fundamentally different aspects of inter-regional coupling.

It is also unclear which connectivity measures are sensitive to specific pathologies and therefore are most appropriate to particular clinical questions. ROI->ROI analysis is useful for identifying low spatial resolution network properties and is computationally efficient due to the low number of correlations computed. ROI->voxel approaches reveal more spatial detail, at the cost of greatly increased calculation time. Voxel->voxel methods, such as ICA, are the most computationally demanding, but do not require a priori anatomical assumptions, and thus may be better suited for exploratory studies of network structure (14). In addition, techniques for ICA network identification have not been standardized and are quite sensitive to specification of the maximum number of identified components. Increasing the maximum number can cause large networks to split into smaller subsets. A major limitation of network analysis methods based on graph theory metrics is that group sizes larger than 40–50 are required to obtain stable estimates of network properties using short acquisition protocols, making them difficult to use in characterizing individual patients (24). Nevertheless, novel indices, like the hub disruption index, may be useful in characterizing an individual's relationship to a group (25). For all of these techniques, compensating for anatomical distortion from space occupying lesions presents a substantial analytical challenge.

### Barrier 3: Reliability and Reproducibility

Recently, there has been growing concern about the reliability and reproducibility of biomedical research (26). Our survey demonstrates that the neuroradiology community shares this concern with respect to rs-fMRI.

Identifying reliable and reproducible canonical brain networks has received great attention in the rs-fMRI literature, with studies showing reproducible networks in both adults and children (27, 28). Yet, the neuroradiology community remains uncertain about how these findings translate to individual patients. More individual participant test-retest studies may be needed to address this area of uncertainty.

Large test-retest data sets, focusing on rs-fMRI from over 36 laboratories around the world, have been made publicly available by the Consortium for Reliability and Reproducibility (CoRR) through the International Data-sharing Neuroimaging Initiative (29). The individual scans composing the large aggregate dataset have been collected using different acquisition parameters and experimental designs, allowing investigators to assess rsfMRI reliability and reproducibility. In addition, the impact of commonly encountered artifacts, such as motion, on interindividual variation can be explored (29). Publicly available datasets from the NIH supported Human Connectome Project (http://humanconnectome.org) are also being used to evaluate the reliability of rs-fMRI and functional connectivity summary measures (30).

In addition, there have not yet been any large scale validation studies to determine if the cognitive domains commonly mapped using intraoperative cortical stimulation can be identified using rs-fMRI. Most rs-fMRI validation studies compare to task-fMRI results, which are expected to have better specificity for specific functions, making simple comparisons difficult. Comparisons between cortical stimulation and other functional imaging modalities have previously shown good between modality correspondence (31), suggesting that this strategy may be useful.

### Barrier 4: rs-fMRI Analysis Issues

While a majority of survey respondents indicated that rs-fMRI data are relatively easy to collect, the majority also believed that rs-fMRI data are relatively difficult to process.

Resting state data analysis can be time intensive and, therefore, not always feasible during a typical demanding day on clinical service. Automatic transfer of images to a clinical image archiving system, followed by automated analysis, could facilitate clinical workflows. One popular analysis program, the CONN Toolbox (13), while well suited for automated analysis of group rs-fMRI data, has limited options for single subject statistical analysis. Nevertheless, a CONN Toolbox script optimized for clinical use and running on a typical laboratory computer requires 10–15 min to process data from a single subject, in addition to the time required to transfer images from PACS. Other toolboxes designed for clinical practitioners, such as CLINICA (32), are not yet widely used, but do hold promise for single subject analysis.

Hemodynamic signal artifacts resulting from physiological noise, including head motion, cardiac pulsation, and respiratory effects can severely compromise efforts to detect regional modulations in neural activity.

Participant head motion is particularly problematic, as it can bias estimated activity correlations between regions. Visual examination of a participant's scan immediately after completion, using a movie loop, allows a clinician to repeat scans when excessive head motion is detected. Nevertheless, even small inter-scan head movements (<0.5 mm) can bias correlation estimates, influencing between-group effect estimates (33). For this reason, motion correction using rigid body realignment is an obligate part of the rs-fMRI preprocessing pipeline, followed by inclusion of motion estimates in subsequent single-subject statistical modeling (34).

Even images from cooperative patients will have physiologic confounds that need to be addressed. Cardiac pulsation and respiration can cause spurious connectivity patterns (35). Bandpass filtering to remove fluctuations outside the frequency range of interest mitigates cardiac and respiratory effects and does not require external physiological recordings. Filtering frequencies lower than ∼0.01 Hz and > ∼0.2 Hz, reduces the effects of non-neuronal physiologic processes (36).

Global signal regression (20) is another method sometimes used for physiologic noise reduction (37). GSR uses a denoising covariate that contains information from both physiological noise and neural signal. Its re-centers the mean of the inter-regional correlation distribution, so that some positive

estimates; Outliers, head motion and global intensity outliers. Display threshold r = 0.4. Original data from NYCSC TRT: subject 16, session 1.

correlations appear to be negative. Its use may therefore confound attempts to distinguish sets of regions whose activity are either positively or negatively associated (38). For this reason, noise reduction techniques like anatomical CompCorr, that exclude the cortical signal from the denoising procedure, may be preferred in most circumstances (13) (**Figure 2**).

Systemic carbon dioxide (CO2) fluctuations alter BOLDcontrast and contribute to respiratory induced signal variation (39). To reduce CO<sup>2</sup> fluctuation effects, end-tidal CO<sup>2</sup> can be measured with a face-mask or nasal cannula and the measurements incorporated into the denoising pipeline (39).

Temporal signal-to-noise ratio (tSNR), the ratio of the mean signal over its temporal standard deviation (SD), reflects the ability to detect BOLD-contrast signal changes (40), and thus can be used in quality assurance. More recently, the Physiological Contributions in Spontaneous Oscillations index has been proposed as a more sensitive measure of functional connectivity strength (41).

These denoising techniques are not only effective at the group level (**Figures 2A,B**), but also can improve sensitivity and specificity for detecting networks at the individual level (**Figure 2C**).

In summary, the inter-regional associations estimated with rs-fMRI may be relatively weak compared to the customary task-fMRI effects, often being masked by physiological noise. The reproducibility of the two modalities may also differ. Varying acquisition and processing parameters can profoundly affect detection sensitivity (42) and there is ongoing debate regarding the role of GSR in pre-processing (43–45). Further, different data analysis families such as ROI-based correlation analysis, independent component analysis (ICA) detection of canonical networks, and graph theory metrics used to quantify local and global network properties, are likely to be sensitive to very different aspects of inter-regional functional connectivity (3).

To allow readers to reproduce the denoising pipeline variations shown in **Figure 2**, links are provided to scripts that preprocess and model the NYU CSC TRT dataset (www. neurometrika.org/tutorials/fc-denoising).

### Barrier 5: User Training

Traditionally, diagnostic radiology has been primarily an anatomical medical specialty. Functional MRI acquisition and interpretation is more physiological and statistical in nature and may therefore may require somewhat different training.

While many academic programs briefly expose trainees to the principles of functional MRI, it is presently not part of the standard curriculum in diagnostic radiology residency or neuroradiology fellowship programs in the U.S. More training in software systems for rs-fMRI analysis will facilitate clinical practice implementation. Relevant curricular offerings in systems neuroscience and statistical modeling could help trainees gain a deeper understanding of the origins of instrumental and physiological noise in rs-fMRI data and thereby optimize their data acquisition, analysis, and interpretation efforts.

## Barrier 6: Standardization, Regulatory, and Financial Issues

The lack of standardization of rs-fMRI acquisition and analysis methods may reflect a lack of consensus regarding the best approach to maximize inter-individual signal variability while concomitantly minimizing intra-subject measure variability (46). As task-fMRI analysis methods are relatively mature compared to their rs-fMRI counterparts, more vigorous engagement of professional societies with the rs-fMRI research community will promote achieving agreement concerning rs-fMRI analysis standards.

Of great importance from a practical viewpoint, there is currently no FDA-cleared software for rs-fMRI analysis on MRI consoles. Obtaining FDA marketing authorization for rs-fMRI clinical use will require validating its intended use as a "tool type" device and more clearly determining what the statistical information derived from rs-fMRI means for patient diagnosis and treatment. Overcoming these hurdles will require a concerted effort from the interested academic and commercial parties. MRI system vendors could have a major role in these activities, working with academic investigators to develop software tools and techniques in accordance with standard medical device development practices, thereby speeding the transition from research to practice.

Acquiring the expertise needed for rs-fMRI acquisition, analysis, and interpretation requires a substantial time commitment. Busy clinicians may be more motivated to obtain such training, and their associated hospitals be more willing to support them, if rs-fMRI had an associated Current Procedural Terminology (CPT) code. Before this can happen, however, rs-fMRI protocols must be standardized by neuroradiologists. Task-fMRI received a CPT code in the U.S. after relative standardization of the processing and analysis techniques. Societies such as the RSNA, ASNR, and ASFNR may be more likely to pursue the process of obtaining an rs-fMRI CPT code after clinical validation and standardization has been achieved.

Even after standardization and regulatory hurdles are overcome, it will be necessary to identify the clinical applications for which rs-fMRI can provide useful information to referring physicians from neurosurgery, neurology and psychiatry. For example, preoperative mapping of motor and language brain function, the most common clinical application of fMRI and rs-fMRI, has been widely integrated into pre-surgical planning protocols in academic centers (32, 47). While resting-state presurgical maps can reliably identify sensorimotor function (12, 48, 49), larger scale validation studies are still needed, and solving problems related to substantial subject level variability remains for language mapping (50, 51) (**Supplement Figure 1**). Individual subject level reliability still needs to be addressed with large studies before clinical services will routinely request rs-fMRI for clinical practice.

# LIMITATIONS

Our study has limitations. First, our response rate was 24% of the ASFNR membership and respondents may have tended to be more enthusiastic about using rs-fMRI in their research and clinical practice than non-respondents. Second, surveys were only sent to the ASFNR membership and thus non-member neuroradiologists who use rs-fMRI were not sampled. Third, for practical reasons, our survey was confined to members of an American professional organization. It will be of interest to survey a broader and more international sample of the neuroimaging community to assess the generality of our findings and interpretations.

# CONCLUSIONS

Despite some perceived impediments to expanding clinical rsfMRI use, neuroradiologists were generally enthusiastic about rs-fMRI in research and clinical applications, believing that their current workplace MRI systems are suitable for rs-fMRI acquisition. Many of the concerns associated with using rs-fMRI in clinical contexts are related to: (1) developing better methods for minimizing physiological noise effects, (2) improving methods for detecting the spatial characteristics of clinicallyrelevant brain processing systems in individual patients, and (3) overcoming remaining standardization, training, and regulatory hurdles.

### AUTHOR CONTRIBUTIONS

EO and TZ contributed to study design, data collection and analysis, and manuscript preparation.

### REFERENCES


### ACKNOWLEDGMENTS

The authors thank Dr. Chris Filippi, Mr. Ken Cammarata, and Ms. Kylie Mason for assisting us in surveying the ASFNR membership and Dr. Daniel Krainak for providing information regarding FDA regulatory issues.

### SUPPLEMENTARY MATERIAL

The Supplementary Material for this article can be found online at: https://www.frontiersin.org/articles/10.3389/fneur. 2019.00420/full#supplementary-material

of random effect fMRI group analyses. Neuroimage. (2016) 126:49–59. doi: 10.1016/j.neuroimage.2015.10.071


functional connectivity using MRI. Hum Brain Mapp. (2016) 37:1986-97. doi: 10.1002/hbm.23150


**Conflict of Interest Statement:** The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

Copyright © 2019 O'Connor and Zeffiro. This is an open-access article distributed under the terms of the Creative Commons Attribution License (CC BY). The use, distribution or reproduction in other forums is permitted, provided the original author(s) and the copyright owner(s) are credited and that the original publication in this journal is cited, in accordance with accepted academic practice. No use, distribution or reproduction is permitted which does not comply with these terms.

# Neurocysticercosis and Hippocampal Atrophy: MRI Findings and the Evolution of Viable or Calcified Cysts in Patients With Neurocysticercosis

Job Monteiro C. Jama-António, Clarissa L. Yasuda and Fernando Cendes\*

Department of Neurology, University of Campinas, UNICAMP, Campinas, Brazil

Neurocysticercosis (NC) is the most common parasitic infection of the central nervous system (CNS). Several studies have reported an association between NC and mesial temporal lobe epilepsy (MTLE). We intended to evaluate the frequency of hippocampal atrophy (HA), clinical evolution and imaging findings in patients with calcified neurocysticercotic lesions (CNLs).

### Edited by:

Hongyu An, Washington University in St. Louis, United States

### Reviewed by:

Jordi A. Matias-Guiu, Hospital Clínico San Carlos, Spain Bo Gao, Affiliated Hospital of Guizhou Medical University, China

> \*Correspondence: Fernando Cendes fcendes@unicamp.br

### Specialty section:

This article was submitted to Applied Neuroimaging, a section of the journal Frontiers in Neurology

Received: 18 January 2019 Accepted: 12 April 2019 Published: 30 April 2019

### Citation:

Jama-António JMC, Yasuda CL and Cendes F (2019) Neurocysticercosis and Hippocampal Atrophy: MRI Findings and the Evolution of Viable or Calcified Cysts in Patients With Neurocysticercosis. Front. Neurol. 10:449. doi: 10.3389/fneur.2019.00449 Methods: One hundred and eighty-one subjects (70 cases and 111 controls) were evaluated for the presence or absence of HA. We assessed the imaging findings, and the evolution of patients with NC treated or not with anthelmintics for NC.

Results: Hippocampal volumes were different between cases and controls (p < 0.001). Seventy percent of the cases presented HA. 52.2% of the patients without a history of anthelmintic treatment for NC had reports of epileptic seizures. There was an association between non-treatment and the later occurrence of epileptic seizures (p = 0.006). There was an association between perilesional edema on MRI and the presence of uncontrolled epileptic seizures (p = 0.004).

Conclusions: Hippocampal atrophy is frequent in patients with NCC. There was an association between no anthelmintic treatment in the acute phase of NC, perilesional edema, more pronounced hippocampal atrophy, and the occurrence of refractory seizures.

Keywords: neurocysticercosis, hippocampal atrophy, perilesional edema, magnetic resonance imaging, seizures, epilepsy, brain calcifications

## INTRODUCTION

Neurocysticercosis (NC) is the most common parasitic infection of the central nervous system (CNS), caused by the larval form of Taenia solium (1). A frequent cause of symptomatic seizures and epilepsy worldwide (2). It is a severe public health problem in several regions of Asia, Africa, and Latin America (3–5).

The earliest documented descriptions of parasitic infection date back from Egyptian Medicine (6). Aristotle was the first to report the presence of cysticerci in animals, between 389 and 375 b.C (7). In ancient Greece, the disease was known as a pig disease (8). From the nineteenth century, it was clear that the disease was transmitted by man and not by animals as it was thought (7). NC was considered a public health problem after the second half of the twentieth century (8, 9).

**52**

Taenia solium is an enteroparasite belonging to the Platyhelminthes phylum, the Cestoda class, the Taeniidae family, the genus Taenia, and the solium species (7).

In its adult form, Taenia solium measures typically 2–4 meters in length and consists of scolex (head), neck (neck) and strobile (body) (8). Adults live on average 3 years and can live up to 25 years, housed in the digestive tract of humans (1).

The scolex invaginates and attaches to the mucosa of the small intestine. After cell division, they become adult tapeworm, which later eliminates gravid proglottid containing thousands of eggs, and thus, the cycle restarts (5). The man can act as an intermediate host, in this case, the human contamination with Taenia solium eggs is processed by (6, 7): External autoinfection; hetero-infection by ingestion of water or food, contaminated with T. solium eggs, disposed of in the environment by carriers; internal autoinfection may occur by intestinal antiperistaltic movements, making possible the presence of gravid proglottid or eggs in the stomach (1).

The oncosphere, when it reaches its final location, undergoes a vesiculation process and loses its aculeus, the invaginated scolex of the future adult, cysticercus Cellulosae (5), forms internally in the vesicle wall. Once established, the larval cysts, through mechanisms of immune evasion (complement inhibition, cytokine release, and masking of host immunoglobulins) actively avoid host immune response (8).

The Taeniasis-cysticercosis complex is a neglected tropical disease (NTD), usually associated with low socioeconomic development (6, 10). It is estimated that 50.000.000 individuals are infected each year (5, 11, 12).

Praziquantel and albendazole have been considered efficient in NC etiologic therapy (13). Therapy with albendazole or praziquantel is indicated in symptomatic individuals with viable cysts on CT or MRI and with positive evidence of immunological evidence for cysticercosis in CSF (2). The purpose of anthelmintic therapy is to try to reduce the duration of the neuroimmunological phenomena involved in NC (5). In most patients, it accelerates the degeneration of cysts and improves symptoms (14).

The clinical manifestations of NC are pleomorphic according to the viability of the parasite, occurring during or after the inflammatory process caused by the presence of dead or degenerated or calcified forms in the cerebral parenchyma (1, 8, 10, 15, 16). Epileptic seizures occur in up to 70–90% of the symptomatic cases of NC and generally represent the primary or unique manifestation of the parenchymatous form of the disease (17, 18). Patients with seizures invariably have prominent inflammatory infiltrate around the cysts, including the presence of pro-inflammatory cytokines and an altered blood-brain barrier (19).

The association between acute symptomatic epileptic seizures and NC is already well established, but the association between drug-resistant epilepsy and NC is still controversial (20). The majority of patients with acute symptomatic seizures in the active phase of the disease experience symptom remission in the next 3 to 6 months, together with the disappearance of the active lesions (20). However, degenerate as well as calcified cysts can lead to chronic epileptic seizures due to hippocampal sclerosis, probably triggered by an inflammatory process, recurrent epileptic seizures and local damage (2).

Neuroimaging and histology studies provide evidence that some nodules are not completely solid, but contain remnants of parasitic membranes that undergo periodic morphological changes related to remodeling mechanisms, thus exposing the host's immune system to the trapped antigenic material, causing recurrent epileptic seizures (2, 21, 22).

The presence of punctiform cerebral calcifications in the correct clinical scenario is mainly indicative of chronic cerebral NC (13). Often these calcifications are the only evidence of the disease (21). However, it is difficult to determine the causality of the relationship between epilepsy and NC, since calcifications are observed in asymptomatic individuals living in endemic areas (23). The use of CT and MRI produce objective evidence regarding the diagnosis of NC (24). These neuroimaging techniques have improved the accuracy of the diagnosis.

The first reports on NC findings on CT were published in 1977; since then, a number of studies have described in detail the different forms of the disease (25). The radiological descriptions allowed the development of clinical classifications of NC based on the topography and evolutionary stage of the lesions and were of great importance for the determination of the rational therapeutic approach in the different forms of the disease (25).

The imaging changes suggestive of NC are dependent on the development stage of the larva. Thus, in the CT the main ones are the following (25):


The mean interval between cysticercus death and radiologically perceptible calcification is approximately 25 months (26).

Cysticercus in intraventricular topography is not always detected by CT since its density is similar to CSF. Therefore, they can only be inferred by the distortion of the ventricular cavity (25).

In MRI the cysts appear with signal properties similar to CSF in both the T1 and T2 sequences. The scolex is usually visualized within the cyst as a high-density nodule, a "hole-withdot" pathognomonic image, characterizing the viable phase (24).

Degenerating cysts (colloidal phase) present poorly defined contours due to edema (25). Some show ring enhancement after contrast administration. The cyst wall becomes thick and hypointense, with marked perilesional edema, best visualized in T2-weighted images (25). In the granular phase, cysts are visualized as ovoid signal areas in the T1 and T2 sequences, with surrounding edema or gliosis with hyperintense borders around the lesion (27). In the calcification phase, the cysts are usually not visualized (25). The susceptibility weighted imaging (SWI) sequence helps to visualize some calcifications. In the T1 and T2 sequences, calcifications can be visualized as small oval, hypointense images (10).

It has recently been demonstrated that calcified cysts can present perilesional edema and post-contrast enhancement, associated with recurrence of symptoms (28, 29).

These characteristics may serve as treatment-defining markers in these patients (30).

Hippocampal sclerosis is the most common structural brain injury associated with refractory mesial temporal lobe epilepsy (MTLE) (8, 31–37).

The histopathological mark of hippocampal sclerosis (HS) is the segmental loss of pyramidal (neuronal) cells, which may affect any segment of the "Ammon's horn," mainly CA1 and CA4, associated with a severe pattern of astrogliosis in the hippocampal formation, including the dentate gyrus (34, 37).

In MRI, HS is characterized by reduced volume and loss of the internal structure of the hippocampus, better visualized in T1-weighted images, observed as hypointense signal, as well as an increased signal in T2-weighted images and FLAIR (37). On the other hand, quantitative volumetric studies allow an objective evaluation of the unilateral or bilateral atrophy of hippocampi, which makes them useful for research applications (24).

The co-existence of NC and TLE associated with HS is common in regions where NC is endemic (31, 33). It is believed that, as in febrile seizures, NC functions as an initial precipitating lesion that would later lead to hippocampal sclerosis (9, 17, 26, 38, 39).

In the last decades, several studies have suggested an association between NC and hippocampal atrophy (HA) (5, 40). New MRI techniques allowed more detailed evaluation of cystic lesions, inflammatory response, and other associated abnormalities (14).

Our objective was to evaluate the frequency of hippocampal atrophy (HA) in patients with NC calcified lesions (NCC), describe the symptomatic evolution of patients treated and not treated for NC, and identify parenchymal alterations associated with the occurrence of epileptic seizures.

### METHODS

### Ethics Statement

All participants signed the informed consent form before performing the magnetic resonance (MRI) examination. This study was approved by the ethics and research committee (CEP-UNICAMP); CAAE Number: 55942116.5.0000.5404.

### Clinical Data

We included 181 subjects (70 cases and 111 controls). Individuals aged 18 years and older, followed by our outpatient's epilepsy clinic or headache clinic at the State University of Campinas (HC-UNICAMP) clinic hospital. We defined our primary variable of interest as the presence of active or calcified cysts in Computed Tomography (CT). We extracted information on the presence of active or calcified cysts from reports of radiological examinations that were available in the medical records. When they were not available, we assigned a qualified neurologist to evaluate CT scans, taking into account Carpio's criteria (41). Patients with a history of follow-up due to neurotuberculosis, neurotoxoplasmosis, tuberous sclerosis, and surgery for temporal lobe epilepsy were excluded from the study. We also excluded patients whose diagnosis was not confirmed after CT evaluation. Seventy patients participated in the study; 48 had no history of treatment for NC, 22 had a history of active cysticercosis and received treatment for NC between the years 1993–2013. The localization of cysts (calcified) observed on CT, were defined as temporal and extratemporal. Patients with multiple calcifications were classified as temporal lobe if they had a temporal lobe lesions, regardless of the location of the other lesions. The extratemporal category was assigned if the location of the lesion was only outside the temporal lobe. Regardless of whether or not they were treated for NC and the number of antiepileptic drugs used, those who had at least one seizure during the evaluation year were considered as individuals with uncontrolled seizures, and those who were 1 year or more without seizures were considered as with seizure control.

All participants performed MRI for volumetric analysis of hippocampus. Those who did not have recent MRI exams (<2 years before the study) were invited to perform further MRIs.

### Protocol of MR Image and Visual Analysis

Patient and control MRI scans were performed on a 3-T Philips Intera Achieva scanner (Philips, Best, The Netherlands), with acquisitions in the coronal, sagittal and axial planes, with coronal sections obtained perpendicularly along the axis of the hippocampal formation, to better study this structure.

### MRI Acquisition Protocol


### Volumetry of the Hippocampus

Patients and controls were matched for age and sex (with similar distribution about age, p = 0.211 and gender, p = 0.693). A group of 111 healthy subjects was used as controls (55.9% female, age 18–80 years, mean 45.05).

We selected the 3D T1-weighted images for volumetry. These were compressed in the neuroimaging informatics technology initiative (NIFTI) format through a web interface. Subsequently, the hippocampal volumes were obtained automatically using the volBrain online program (http://volbrain.upv.es). The automatic analyses were performed without knowledge of clinical data. All individual hippocampal values were corrected for total intracranial volumes. All values obtained were transformed into Z-score. The Z-score values of the corrected volumes or asymmetry index (defined by the ratio of the smallest to the largest hippocampus), which were equal to or lower than −2 were considered indicative of HA (**Table 1**).

### Visual Analysis of Images

In patients with a history of NC treatment, a visual analysis of the MRI examinations acquired on a 3T (as described previously) or in a 2.0T (Elscint Prestige, Haifa, Israel) scanner was performed by two investigators (JMCJA and FC). In addition, 54 MRIs were analyzed with the objective of evaluating the evolution of the cysts through the images. Th MRI acquisitions of these 54 patients were carried out between the years 2004 to 2018. The findings were correlated with the occurrence of a seizure described in the medical record during the period of MRI (equal to or <1 month). Further details are in **Tables 2**, **3**.

### Statistical Analysis

Data analysis was performed using SPSS software version 23 for mac. First, we did an exploratory analyses, measuring the frequency of categorical data and descriptive statistics for quantitative data.

To compare the groups (controls and cases), we performed a normality test (Kolmogorov-Smirnov). Then, the Mann-Whitney or Kruskall-Wallis test was performed to analyze numerical variables. Multivariate analysis was performed on numerical variables (controls, treated, and not treated for NC). The chi-square or Fisher's test were used to analyze the categorical variables. The significance was determined as p < 0.05 for all analyses.

### RESULTS

From an original sample of 211 participants, we included 181 (111 controls and 70 cases). Ninety-nine were female, mean age= 45.8, ±12.4. Hippocampal volumes of the controls were significantly different from the cases by the Man-Whitney test (p < 0.001, **Figure 1**). In a subgroup analysis (controls, patients treated, and patients untreated for NC), we observed that there was only a difference of controls compared to patients untreated for NC (p = 0.001; **Figures 2**, **3**). Groups had a similar gender distribution (p = 0.693).

TABLE 1 | Distribution of the Z-score values and asymmetry index of the hippocampus volumes of patients who had HA.


L, left; R, right; B, bilateral; HA, hippocampal atrophy.

TABLE 2 | Distribution of study variables and the level of significance.


The Kruskall-Wallis test showed a significant difference between the groups (p = 0.001). There was an association between non-treatment for NC and recurrence of seizures (p = 0.003, chi-square test). Hip, Hippocampal; bil, Bilateral; SD, Standard Deviation. Significant p-values are in bold.

### Case Analysis

Of the 70 cases, 22 (31.4%) were treated for NC, 48 (68.6%) were not (**Figure 4**). There was no difference in the volume of the hippocampi of treated and untreated patients for NC (p = 0.225). There was no age difference (p = 0.220) or sex distribution (p = 0.401) between groups.

### Location of Calcifications

In 34/70 (48.6%) the NC calcifications were localized in the temporal lobe: 14/34 (20%) in the left temporal lobe, 9/34 (12.9%) in the right temporal lobe and 11/34 (15.7%) in both temporal lobes. In 36/70 (51.4%) the NC calcifications were localized in extratemporal regions.

### Number of Calcifications

Twenty-six of 70 (37.1%) patients had one to two parenchymal calcifications, 24/70 (34.29%) had three to five calcifications, 14/70 (20%) had six to twenty calcifications, 6/70 (8.57%) had more than twenty calcifications.

### Clinical Manifestation

Only 1/70 (1.4%) of the patients did not present seizures in the acute phase or in the follow up.

### Hippocampus Atrophy

Forty-nine of the 70 (70%) patients presented HA. There was no difference between HA and the localization of calcifications (p = 0.2, Fisher exact test). Fifteen of the 22 (68.18%) patients treated and 34/48 (70.83%) of the untreated patients had HA. There was no association between the frequency of HA and treatment for NC (p = 0.83); however, patients who did not receive anthelmintic treatment in the acute phase had significantly smaller hippocampal volumes (p = 0.0001). There was no association between HA and sex (p = 0.96). Only 17/70 had a family history of epilepsy (p = 0.06). Further details are in **Table 2**.

### Epileptic Seizures Report

forty-four of the 69 (68.8%) patients had uncontrolled epileptic seizures; 36 of these 44 (81.8%) did not receive anthelmintic treatment for NC in the acute phase of the disease. There was an association between the uncontrolled epileptic seizures and non-treatment for NC (p = 0.003).

Thirty-four of the 44 (77.3%) patients with uncontrolled seizures presented HA and remaining 22.7% had well controlled seizures (p = 0.065).

### MRI Visual Analysis

Here we analyzed the patients with more than one MRI exam, and whose presence of viable cysts was confirmed by imaging tests.

Fifty-four MRI exams of 22 patients performed between 2004 and 2018 were analyzed. The average duration of followup was 15 years (range of 4–23 years). Five of 22 (22.72%) patients had active cysts in at least one of the exams. Two of 22 (9.09%) had ventricular dilatation, and 3/22 (13.63%) had diffuse cerebral atrophy.

Nineteen of 22 (86.4%) patients presented perilesional gliosis in at least one of the calcified lesions. However, there was no association between the presence of gliosis and the occurrence of seizure (p = 0.963). Sixteen of 22 (72.7%) presented perilesional edema around at least one of the calcified lesions. There was an association between the presence of perilesional edema and the occurrence of seizure in the weeks before the MRI exam (p = 0.004). Fourteen of 22 (63.6%) had contrast enhancement around at least one of the calcified lesions. There was no association between contrast enhancement and the occurrence of seizures (p = 0.51). Eight of these 22 (36.4%) had hippocampal atrophy. Further details are in **Table 4**.

### Evolution of Patients With Active Cysts

We evaluated an average of 3 exams for each patient, performed between 3 to 11 years after the first examination of the acute phase of cysticercosis (viable or degenerating cysts). Five of these 22 presented active cysts in initial MRIs.

In one case, we observed the occurrence of hippocampal atrophy 2 years after the beginning of the cyst degeneration process that was not present before (**Figure 5**).

The evolution of the cysts was variable (**Figures 5**–**7**): The process of calcification occurred between 3 and 4 years after the diagnosis of active cysts. However, in one specific case, the degenerative cysts maintained enhancements for about 10 years later (2007–2017, details in **Figure 7**).

### DISCUSSION

We observed a high frequency of hippocampal atrophy in patients with NC (70%), suggesting a possible association between NC and HA. This possibility has been considered for years by several authors, who have studied such an association (5, 38, 40, 42, 43).

In a study that sought to determine the relationship between HA, NC and seizure semiology in epileptic patients, the authors observed that HA is more frequent in patients with MTLE and calcified NC, compared to patients with extratemporal epilepsies (40). In another population study, the authors, when assessing the association between NC and HA in older adults living in an endemic area found a high prevalence of HA (68%) in patients with calcified NC compared to controls (26). In

TABLE 4 | Distribution of the main findings of visual MRI analysis and the level of significance in relation to the seizure occurrence.


There was an association between perilesional edema and recurrence of seizures (p = 0.004; Fisher's test). Significant p-values are in bold.


TABLE 3 | Main findings of visual MRI analysis of patients treated for NC and report of seizures in the same period.

In this table, we illustrate the gender, the date of the first MRI, the month and the year of the exam with parenchymal alteration, and the record of epileptic seizures in the same period.

FIGURE 1 | Hippocampal volumes of patients and controls. This graph demonstrates that there is a difference in the size of the hippocampus of NC patients compared to healthy controls. The Mann-Whitney test showed a significative difference between the hippocampal volume of patients and controls (p = 0.001). Evidence of a possible relationship between NC and hippocampal atrophy. HIP.NOR.RIGHT: hippocampus normalized right; HIP.LEFT: hippocampus left. Patient–Hip. Right, Mean = 3.50 cm; SD = 0.57; Range = 2.68; Hip. Left, Mean = 3.36 cm; SD = 0.60; Range = 2.88; Controls–Hip. Right, Mean = 3.92 cm, SD = 0.34, Range = 1.92; Hip. Left, Mean = 3.84 cm, SD = 0.31; Range = 1.89.

another study, the authors evaluated 324 patients with MTLE-HS undergoing temporal lobectomy, and they found a high prevalence of calcific NC, 126/324 (38.9%) (4). Another casecontrol study found a high frequency of calcified NC in patients with MTLE-HS (31).

During the last decades, anecdotal reports and small series of cases have brought this association to the attention of the medical community, describing patients with drugresistant MTLE-HS whose neuroimaging studies showed granular or calcified cysticerci located in the hippocampus or neighboring tissues (2). In some cases, the pathological

exams revealed HS with neuronal loss in the CA1 and CA4 layer, and gliosis, as well as the presence of an intense inflammatory reaction in the brain tissue around the calcified parasites (2).

In the active form of cysticercosis, inflammation involves the parasites, and is the most common mechanism for the occurrence of seizures in the acute phase on NC (3). This inflammation is due to the aggregation of mononuclear lymphocytes, plasma cells and variable numbers of eosinophils at the lesion site (3). Experimental studies have

hippocampal atrophy; (C) T1-weighted coronal image, showing diffuse cerebral atrophy, including bilateral hippocampal atrophy; (D) FLAIR sequence, with

hyperintense signal in the hippocampus (atrophy), and left frontal and perinsular hyperintense lesions.

suggested that the injection of Taenia granuloma material into the mouse hippocampus is highly epileptogenic, supporting the involvement of the hippocampus by the inflammatory responses of the brain of the degenerating cysticerci (38).

Current evidence shows that the relationship between NC and MTLE-HS has always coexisted in endemic areas (38). However, the extent of this occurrence remains to be determined, so in many cases it is considered as "dual pathology" (2, 9, 38). Most of the information on this association comes from series of patients with MTLE-HS that suggest a cause-and-effect relationship (2, 4, 26). As in the febrile seizures during childhood, NC would act as an initial precipitating lesion, which would cause damage to

FIGURE 6 | Illustration of the calcified NC associated with perilesional edema. (A–C) T1-weighted images, with calcification in the putamen. (B–D) Images in T2 and FLAIR, with edema around the calcification.

the hippocampus, leading to loss of neurons and synaptic reorganization of the cellular elements (9, 14, 38, 40, 44). In this conjecture, it has been suggested that cysticerci can lead to HS because they cause repetitive inter-ictal discharges, recurrent clinical and subclinical seizures or possibly epileptic status, which results in MTLE-HS, and in turn aggravate seizures (9, 26, 38). These parasites do not necessarily have to be located within the limbic system (17), suggesting a deleterious remote effect of NC-induced reactive seizures in hippocampal neurons (38).

On the other hand, parasitic cerebral lesions may lead to inflammation-mediated hippocampal damage associated or not with genetic susceptibility (9, 42, 45). In this view, the periodic remodeling of cysticercus occurs with the exposure of parasitic antigens bound to the host's immune system, which does not require recurrent seizures as a causal factor (9, 26). Although this has not been demonstrated in humans, there is experimental evidence showing that repeated exposure to endotoxin and increased levels of pro-inflammatory cytokines correlate with hippocampal damage, supporting the hypothesis of inflammation-mediated atrophy or hippocampal damage (2, 26).

Another possibility is that the presence of HS in patients with NC may be only a coincidence (31, 42), which in our view is less likely, given the high prevalence reported in this and other studies (40).

In cysticercosis calcification, recurrent seizures may result from inflammation related to exposure of the host immune system to parasitic remains (2). In the vicinity of the lesion, the tissue reaction usually consists of astrocytic gliosis and a small border of demyelination. Neurons are affected variably and tend to undergo degenerative changes (3). It seems reasonable to assume that the inflammation at the stage of nodular calcification is similar to that of the colloidal stage. Acute and recurrent seizures, if repeated, may cause additional hippocampal damage. Also, degenerate and calcified cysticerci can directly induce hippocampal sclerosis by damage mediated by local or remote inflammation of hippocampal neurons causing refractory epilepsy (2).

The format of this study did not allow us to directly establish a cause and effect relationship between NC and HA, however, in a case of active NC, we were able to demonstrate that hippocampal atrophy was related to the degenerate cysticercus, due to an inflammatory reaction. There was no HS before degeneration of the cysticercus, however, 3 years later the MRI signs of HS were observed (**Figure 2**). In this case, the hippocampus has probably been directly affected by the inflammatory response and gliosis that develops around the cyst and/or adjacent areas (38).

In addition to the high frequency of MTLE-HS in our patients with calcified NC, there was an association between the absence of anti-helmintic treatment in the acute phase of NC and later uncontrolled epileptic seizures, as well as smaller hippocampal volumes, something that may infer that anthelmintic treatment works as a protective factor. MTLE-HS is often pharmacoresistant and many patients reach seizure-free status only after surgical treatment (9).

The mechanism of involution of cysticercosis, which, contrary to what was previously thought, the final step (degeneration and calcification), is not completely inert (21, 46). It is known that NC is a potential cause of refractory epilepsy and that the presence of perilesional gliosis contributes to epileptogenicity (30). About half of the patients with only calcified lesions and recent ongoing seizures, developed perilesional edema at the time of seizure recurrence (28). A plausible explanation for the occurrence of perilesional edema may be that they are not all alike and may differ in the amount, in the form of calcium deposition, in the degree of antigens recognized by the host, in the level of residual inflammation, or by the proximity of a blood vessel (46), which favors the occurrence of perilesional edema. On the other hand, genetic factors may also be related (20). Some attest that this is due to dysfunction of the blood-brain barrier, probably due to the presence of inflammation and/or perilesional gliosis conditioned to the host's response to the newly recognized or released parasite antigen and/or to the positive regulation of the immune response of the host (28). Histopathological examination of calcification associated with multiple episodes of perilesional edema revealed

### REFERENCES


significant inflammation, which supports the concept that edema is inflammatory in nature (28).

Some authors argue that perilesional edema is the result of an inflammatory process directed at the sequestered parasite antigen (47), and therefore advocates specific measures to limit the inflammation process, which can be used to treat or prevent complications (28).

Another hypothesis is that perilesional edema occurs as a consequence of seizure activity (13). However, there are differences between edema associated with a flurry of seizures and perilesional edema, the first being more diffuse, with no defined maximum area of activity, presumably caused by the loss of fluid by damaged cells, while the second presents a peak, almost always accompanied by contrast enhancement, probably of vasogenic origin (28). In general, edema around calcification after seizures is considered an evident form of injury that is probably epileptogenic (20, 48). A previous study concluded that the presence of edema is a predictor of recurrence of seizures (30).

We conclude that there is a high frequency of AH in patients with NC, which may suggest an association between both. In addition, there was an association between no anthelmintic treatment and the later occurrence of uncontrolled seizures and smaller hippocampi, as well as between perilesional edema and seizures near the time of the MRI exam.

### ETHICS STATEMENT

This study was approved by the ethics and research committee (CEP-UNICAMP); CAAE Number: 55942116.5.0000.5404.

# AUTHOR CONTRIBUTIONS

All authors listed have made a substantial, direct and intellectual contribution to the work, and approved it for publication.

### FUNDING

This work was supported by FAPESP (Fundação de Amparo à Pesquisa do Estado de São Paulo) grant # 2013/07559-3, and CAPES (Coordenação de Aperfeiçoamento de Pessoal de Nível Superior).


of patients with drug-resistant epilepsy. Epilep Behav. (2017) 76:168–77. doi: 10.1016/j.yebeh.2017.02.030


**Conflict of Interest Statement:** The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

Copyright © 2019 Jama-António, Yasuda and Cendes. This is an open-access article distributed under the terms of the Creative Commons Attribution License (CC BY). The use, distribution or reproduction in other forums is permitted, provided the original author(s) and the copyright owner(s) are credited and that the original publication in this journal is cited, in accordance with accepted academic practice. No use, distribution or reproduction is permitted which does not comply with these terms.

# CSF Protein Concentration Shows No Correlation With Brain Volume Measures

Alexander Wuschek 1,2, Sophia Grahl 1,2, Viola Pongratz 1,2, Thomas Korn1,3,4, Jan Kirschke<sup>5</sup> , Claus Zimmer <sup>5</sup> , Bernhard Hemmer 1,3 and Mark Mühlau1,2 \*

<sup>1</sup> Department of Neurology, Klinikum Rechts der Isar, Technische Universität München, Munich, Germany, <sup>2</sup> TUM-Neuroimaging Center, Klinikum Rechts der Isar, Technische Universität München, Munich, Germany, <sup>3</sup> Munich Cluster for Systems Neurology (SyNergy), Munich, Germany, <sup>4</sup> Department of Experimental Neuroimmunology, Technische Universität München, Munich, Germany, <sup>5</sup> Department of Neuroradiology, Klinikum Rechts der Isar, Technische Universität München, Munich, Germany

Background: CSF protein concentrations vary greatly among individuals. Accounting for brain volume may lower the variance and increase the diagnostic value of CSF protein concentrations.

### Edited by:

Hongyu An, Washington University in St. Louis, United States

### Reviewed by:

Yann Quidé, University of New South Wales, Australia Christian Herweh, Heidelberg University Hospital, Germany

> \*Correspondence: Mark Mühlau mark.muehlau@tum.de

### Specialty section:

This article was submitted to Applied Neuroimaging, a section of the journal Frontiers in Neurology

Received: 07 February 2019 Accepted: 16 April 2019 Published: 03 May 2019

### Citation:

Wuschek A, Grahl S, Pongratz V, Korn T, Kirschke J, Zimmer C, Hemmer B and Mühlau M (2019) CSF Protein Concentration Shows No Correlation With Brain Volume Measures. Front. Neurol. 10:463. doi: 10.3389/fneur.2019.00463 brain volume. Methods: Brain volumes (total intracranial, gray matter, white matter volumes) derived

Objective: To determine the relation between CSF protein concentrations and

from brain MRI and CSF protein concentrations (total protein, albumin, albumin CSF/serum ratio) of 29 control patients and 497 patients with clinically isolated syndrome or multiple sclerosis were studied.

Finding: We found significant positive correlations of CSF protein concentrations with intracranial, gray matter, and white matter volumes. None of the correlations remained significant after correction for age and sex.

Conclusion: Accounting for brain volume derived from brain MRI is unlikely to improve the diagnostic value of protein concentrations in CSF.

Keywords: albumin, brain volume, CSF, protein, MRI

# INTRODUCTION

Cerebrospinal fluid (CSF) analysis is supportive of the diagnosis of many neurological diseases. CSF protein concentrations constitute a mainstay of CSF analysis. Despite age- and sex-dependent cut-offs (1–4), considerable interindividual variance may lower the diagnostic value of CSF protein concentrations. We aimed to reduce variance of CSF protein concentrations and, hence, to increase their diagnostic value by considering brain volumes derived from magnetic resonance imaging (MRI). This idea may not seem practical at first glance but, given latest developments with regard to modern hospital information systems and tools for automated MRI analysis, linking of multiple paraclinical data seems to be in reach even in clinical routine. We reasoned that, since most CSF proteins are both released into CSF (mainly ultrafiltration of blood plasma in the choroid plexus) and retrieved from CSF (drainage into the venous system mostly through arachnoid granulations) in certain circumscribed brain structures, differences in whole brain volumes may not perfectly parallel the net capacity of CSF protein filtration and drainage as only this would lead to independence between brain volumes and CSF protein concentrations. Thus, we studied the relation of CSF protein concentrations (total protein, albumin, and albumin CSF/serum ratio) and brain volumes (total intracranial volume, TIV; gray matter, GM; white matter, WM).

### METHODS

The study was approved by the local ethics committee of the medical faculty of the Technical University of Munich. Written informed consent was obtained. We selected patients, who received brain MRI and lumbar puncture, from our in-house database. As a surrogate of a healthy subjects, we firstly reviewed medical records for patients with transient symptoms (e.g., headache) but without a severe or chronic neurological disorder. Based on these strict criteria, we could only include 29 adult patients (age, 31.4 ± 9.1 years; females, 24) as control patients (CP). Given this relatively small number, we secondly selected patients with clinically isolated syndrome (CIS) or multiple sclerosis (MS) aged between 18 and 60 years resulting in a group of 497 patients (age, 35.7 ± 9.4; females, 342; EDSS, 1.45 ± 1.40; CIS, 50.3%; relapsing-remitting MS, 44.1%, primary and secondary progressive MS, each 2.8%). We chose this population, because the suspicion of MS usually prompts performing both cranial MRI and lumbar puncture in clinical setting. Therefore, we could assemble a relatively homogenous and large group of patients with both MRI and CSF data. This approach seemed justified, since total protein levels are normal in 75 percent of MS patients with mild elevation in the remainder (5, 6), whilst levels above 1,000 mg/L are unusual (7). To increase statistical power, we gathered patients with MS and its precursor CIS in one group. Protein levels and albumin CSF/serum ratios were determined by nephelometry (Siemens ProSpec <sup>R</sup> ). As described earlier (8), all brain MRI were acquired with the same protocol on the same 3T scanner (Achieva, Philips, Netherlands). We used a 3D gradient echo T1-weighted sequence (orientation, 170 contiguous sagittal 1 mm slices, reaching down to C4/C5; field of view, 240 × 240 mm; voxel size, 1.0 × 1.0 × 1.0 mm; repetition time (TR), 9 ms; echo time (TE), 4 ms) and a 3D fluid attenuated inversion recovery sequence (orientation, 144 contiguous axial 1.5 mm slices, reaching down to the foramen magnum; field of view, 230 × 185 mm; voxel size, 1.0 × 1.0 × 1.5 mm; TR, 10,000 ms; TE, 140 ms; TI, 2,750 ms). Volumes of GM and WM were obtained from the first segmentation step of the CAT12 toolbox (Version 916, http://dbm.neuro.uni-jena.de), an extension of SPM12 software (Version 6685, http://www.fil. ion.ucl.ac.uk/spm). TIV was estimated by a 'reverse brain mask method' (9) after lesion filling. For statistical analysis, we used unpaired t-tests, simple, and partial correlations in IBM SPSS Statistics for Windows (Version 25.0).

### RESULTS

First, we characterized our data set by testing for well-known associations of the demographic parameters of age and sex with CSF protein concentrations on the one hand and with brain TABLE 1 | Correlations of brain volumes with CSF protein concentrations.


r, Pearson correlation coefficient; n, sample size; TIV, total intracranial volume; GM, gray matter; WM, white matter; CSF protein, total protein concentration in the cerebrospinal fluid in mg/L; CSF albumin, total albumin concentration in the cerebrospinal fluid in mg/L; Q-Alb, albumin CSF/serum ratio × 10E-3.

volumes on the other hand. CSF protein concentrations were significantly higher in men than in women [independent-samples t-test; protein in mg/L, 624 ± 207 vs. 492 ± 196, t(524) = 6.99; albumin in mg/L, 318 ± 122 vs. 233 ± 100, t(524) = 8.39; albumin CSF/serum ratio, 6.8 ± 2.4 vs. 5.3 ± 2.4, t(524) = 6.41; all p < 0.001]. As expected (10), age correlated with CSF protein concentrations (linear correlation; protein, r = 0.17; albumin, r = 0.15; albumin CSF/serum ratio, r = 0.17; all p < 0.001). Although CP were significantly younger than CIS/MS patients [independent-samples t-test; 31.4 ± 9.1 vs. 35.7 ± 9.4, t(524) = 2.38, p = 0.018], none of the CSF protein concentrations significantly differed between the two groups [CP vs. CIS/MS; independent-samples t-test; protein in mg/L, 489 ± 149 vs. 535 ± 211, t(524) = 1.16, p = 0.25; albumin in mg/L, 234 ± 83 vs. 260 ± 116, t(524) = 1.21, p = 0.23; albumin CSF/serum ratio, 5.4 ± 1.9 vs. 5.8 ± 2.5, t(524) = 0.96, p = 0.34]. With regard to brain volumes, we could replicate well-known associations: men had larger TIV than women [independent-samples t-test; 1578 ± 141 vs. 1415 ± 136 ml, t(524) = 12.42, p < 0.001); GM volume negatively correlated with age (linear correlation, r = −0.37, p < 0.001].

Given the strong associations of age and sex with both CSF protein concentrations and brain volume measures, we report only results of partial correlation analyses with age and sex as covariates in detail. In the CP group, none of the CSF protein concentrations correlated with brain volume. In the CIS/MS group, we found statistically significant positive correlations of brain volumes with protein CSF concentrations. Yet Pearson correlation coefficients were in the same range as in the CP group suggesting a lack of statistical power to demonstrate these associations in this small group of only 29 subjects. In the CIS/MS group however, none of significant associations between CSF protein concentrations and brain volumes survived correction for TIV, age, and sex (**Table 1**).

### DISCUSSION

We related brain volumes, derived from high-resolution MRI as available in clinical routine, to CSF protein concentrations. Our data are plausible as we could replicate well-known associations. Men showed higher values of CSF protein concentrations than women. Protein concentrations increased with age. Moreover, we could replicate well-known associations of brain volumes with age and sex. Age and sex are very important clinical parameters; they are available and considered in (almost) every patient in clinical routine and go along with differences in both CSF protein concentration and brain volumes. Therefore, we felt that an association of CSF protein concentration and brain volumes, potentially meaningful in clinical routine, should remain significant after correction for both age and sex.

Accordingly, after having failed to demonstrate a relationship of brain volumes and CSF protein concentrations beyond that explained by age and sex in as many as 526 subjects, we conclude that accounting for individual brain volumes is unlikely to considerably decrease the variability of CSF protein concentrations and, hence, to increase their diagnostic value.

### REFERENCES


### ETHICS STATEMENT

The study was approved by the local ethics committee of the medical faculty of the Technical University of Munich.

### AUTHOR CONTRIBUTIONS

AW and MM contributed to the conception and design of the study. AW, SG, VP, TK, JK, CZ, BH, and MM participated in the acquisition and analysis of data. AW and MM contributed to drafting the text or preparing the tables.

### FUNDING

AW was funded by the Kommission für Klinische Forschung (KKF), Klinikum Rechts der Isar.

parameters. Mult Scler. (2017) 24:1115-1125. doi: 10.1177/1352458517712078


**Conflict of Interest Statement:** The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

Copyright © 2019 Wuschek, Grahl, Pongratz, Korn, Kirschke, Zimmer, Hemmer and Mühlau. This is an open-access article distributed under the terms of the Creative Commons Attribution License (CC BY). The use, distribution or reproduction in other forums is permitted, provided the original author(s) and the copyright owner(s) are credited and that the original publication in this journal is cited, in accordance with accepted academic practice. No use, distribution or reproduction is permitted which does not comply with these terms.

# The Effects of Fatiguing Aerobic Exercise on the Cerebral Blood Flow and Oxygen Extraction in the Brain: A Piloting Neuroimaging Study

Dapeng Bao<sup>1</sup> , Junhong Zhou2,3, Ying Hao<sup>4</sup> , Xuedong Yang<sup>5</sup> , Wei Jiao<sup>1</sup> , Yang Hu<sup>1</sup> and Xiaoying Wang4,5 \*

*<sup>1</sup> Sport Science Research Center, Beijing Sport University, Beijing, China, <sup>2</sup> The Hinda and Arthur Marcus Institute for Aging Research, Hebrew SeniorLife, Roslindale, MA, United States, <sup>3</sup> Harvard Medical School, Boston, MA, United States, <sup>4</sup> Peking University, Academy for Advanced Interdisciplinary Studies, Beijing, China, <sup>5</sup> Department of Radiology, Peking University First Hospital, Beijing, China*

### Edited by:

*Deqiang Qiu, Emory University, United States*

### Reviewed by:

*Bo Gao, Affiliated Hospital of Guizhou Medical University, China Xiaoyun Liang, Australian Catholic University, Australia Henk J. M. M. Mutsaerts, VU University Medical Centre, Netherlands*

> \*Correspondence: *Xiaoying Wang cjr.wangxiaoying@vip.163.com*

### Specialty section:

*This article was submitted to Applied Neuroimaging, a section of the journal Frontiers in Neurology*

Received: *24 December 2018* Accepted: *04 June 2019* Published: *21 June 2019*

### Citation:

*Bao D, Zhou J, Hao Y, Yang X, Jiao W, Hu Y and Wang X (2019) The Effects of Fatiguing Aerobic Exercise on the Cerebral Blood Flow and Oxygen Extraction in the Brain: A Piloting Neuroimaging Study. Front. Neurol. 10:654. doi: 10.3389/fneur.2019.00654* The fatigue in aerobic exercise affects the task performance. In addition to the fatigue in the muscular system, the diminished performance may arise from the altered cerebral blood supply and oxygen extraction. However, the effects of the fatiguing aerobic exercise on the ability of brain to regulate the cerebral blood flow (CBF) and to extract the oxygen are not fully understood. In this pilot study, we aim to quantify such effects via advanced functional MRI techniques. Twenty healthy younger elite athletes were recruited. In the screening visit, one circle ergometer test was used to screen the maximal relative oxygen consumption (VO2max). Eleven eligible participants then completed the next MRI visit after 7 days. These participants completed a 2-min pulsed arterial spin labeling (ASL) using the PICORE/QUIPSS II and 5-min asymmetric spin echo (ASE) scan at baseline and immediately after the aerobic circle ergometer test. The CBF was then measured using the ASL images and the oxygen consumption of the brain was quantified using oxygen extraction fractions (OEF) derived from the ASE images. The test time, VO2max, and anaerobic threshold were also recorded. As compared to baseline, participants had significant reduction of global CBF (*p* = 0.003). Specifically, the CBF in bilateral striatum, left middle temporal gyrus (MTG) and right inferior frontal gyrus (IFG) decreased significantly (*p* < 0.005, *K* > 20). No significant changes of the OEFs were observed. Participants with greater OEF within the right striatum at baseline had longer test time, greater anaerobic threshold and relative VO2max (*r* <sup>2</sup> > 0.51, *p* < 0.007). Those with longer test time had less reduction of CBF within the right IFG (*r* <sup>2</sup> = 0.55, *p* = 0.006) and of OEF within the left striatum (*r* <sup>2</sup> = 0.52, *p* = 0.008). Additionally, greater anaerobic threshold was associated with less reduction of OEF within the left MTG (*r* <sup>2</sup> = 0.49, *p* = 0.009). This pilot study provided first-of-its-kind evidence suggesting that the fatiguing aerobic exercise alters the cerebral blood supply in the brain, but has no significant effects on the ability of brain to extract oxygenation. Future studies are warranted to further establish the CBF and OEF as novel markers for physical and physiological function to help the assessment in the sports science and clinics.

Keywords: aerobic exercise, fatigue, fMRI, cerebral blood flow, oxygen extraction fractions

## INTRODUCTION

Physiological fatigue is one of the main contributors to the diminished performance in aerobic exercise. With the increase of the exercise load, multiple physiological factors (1, 2), such as the intramuscular metabolism, excitation-contraction coupling, are altered, causing the inability of muscles to produce enough power forming the voluntary motion (3). In addition to these peripheral factors, the fatigue may affect the functionality of the brain, including the regulation of cerebral hemodynamics and oxygen consumption (4).

The successful completion of the motion in exercise is dependent upon the capacity of neurons in the brain to process the afferent information from peripheral systems and send the feedback to musculoskeletal system via neurotransmitters appropriately. These important neural activities rely on the sustainable supply of the oxygenated blood and the extraction of oxygen. Previous studies have provided preliminary evidence showing the effects of the exercise-induced fatigue on the cerebral hemodynamics and oxygen consumption. Poulin et al. (5), for example, observed the global cerebral blood flow (CBF) of the brain, as measured using transcranial Doppler ultrasound (TCD), increased after the exercise of mild load (i.e., at 20 and 40% of the maximal oxygen uptake). In a separate study, Thomas and Stephane (6) observed that the oxygenation of the prefrontal cortex, as measured using near-infrared spectroscopy (NIRs), increased in the first minutes of the exercise but decreased when the exercise load increased exhaustively. However, the changes of the CBF and the extraction of the oxygen within the specific brain regions, as well as the underlying mechanism of such regulation in response to aerobic exercise (5), are still unclear.

The advanced functional magnetic resonance imaging (fMRI) techniques enable non-invasively quantifying the functional characteristics of the brain, including the cerebral blood flow and oxygen consumption with high-resolution images. The development of the oxygen extraction fraction (OEF) sequences in fMRI (7), for example, allows measuring oxygen uptake of the small brain regions (as small as within several voxels). In this pilot study, we aim to explore the effects of fatiguing aerobic exercise on the regulation of cerebral blood flow and the oxygen extraction of the brain via these advanced fMRI techniques. Specifically, we hypothesize that after the aerobic circle ergometer exercise with incremental load to the maximal oxygen consumption (VO2max), participants would have a significant decrease in their CBFs and OEFs, particularly in the cerebral regions associated with the voluntary movements and task motivation.

### METHODS

### Participants

After screening the performance of 800-meter race in 410 healthy young athletes, 20 of them were recruited in this study. All of them were "elite" as they were able to complete the 800 meter race within 123 s. They had no injury within the past 3 months and were without any self-reported and/or diagnosed metabolic or neurological diseases. Those with a relative VO2max

<55 ml/min/kg (8–10), as measured in the screening visit, were excluded for the following MRI test.

## ETHICS STATEMENT

This study was approved by Institutional Review Board of Beijing Sport University, and conducted according to the principles of the Declaration of Helsinki. All the participants provided written informed consent as approved by the institutional review board.

### Study Protocol

This study consisted of two visits: the screening visit and MRI visit. The participants were first screened based upon their relative VO2max measured in the cycle ergometer test in the screening visit. Eligible participants then completed the MRI visit 7 days after and the CBF and OEF were measured during the MRI visit.

### Screening Visit

All the twenty participants completed one cycle ergometer test on an electrical cycle ergometer (Monark839) on this visit. They were instructed to not perform heavy exercise at least 24 h prior to the visit or any other exercise on the same day of the visit. They were also asked to not take consuming food or beverages containing caffeine before the test. The temperature in the testing room was maintained between 20 and 25◦C, and the relative humidity was between 40% and 50%. The load of the ergometer was set at 90 Watts at the beginning of the test and increased progressively with 15 Watts per minute. The gas analyzer (AEI Technologies, TX USA) was used to assess the oxygen consumption. Before each test started, the gas analyzer was calibrated. Several metrics of the oxygen consumption, including the relative VO2max and anaerobic threshold, were measured. The study personnel used the Borg Rating of Perceived Exertion scale to assess the degree of the fatigue of the participants during the test (11). This scale scored from 6 to 20 and greater score represented higher fatigue. When the participant reported a score ≥19 (i.e., the exercise load with their maximal effort), the test stopped. Nine of the participants had the relative VO2max <55 ml/min/kg and were thus excluded.

### MRI Visit

The 11 eligible participants then completed the MRI visit 7 days after the screening visit. All the participants completed the MRI scan, consisting of a 2-min pulsed arterial spin labeling (ASL) scan (12, 13) and a 5-min asymmetric spin echo (ASE) scan (14) before and immediately after the cycle ergometer test. The same protocol of ergometer test as in the screening visit was used. The duration of the cycle ergometer test (i.e., test time), relative VO2max and anaerobic threshold was recorded and used in the following analyses.

A 3T MRI scanner (GE Medical System) with 8-channel standard head array coil was used to acquire the MRI data. A 3D FSPGR scan was collected for whole-brain highresolution anatomy. The ASL was used to acquire the CBF using the following parameters: PICORE/QUIPSS II; Slice thickness/gap = 8.0/2.0 mm; flip angle = 90 degrees, field of view = 230 mm × 230 mm, TR = 3,000 ms; TE = 3.1 ms, TI<sup>1</sup> (inversion time) = 700 ms, TI<sup>2</sup> = 1,500 ms; volumes = 50. The ASE protocol, consisting of one spin echo and 19 ASE scans, was used to acquire the OEF (TE = 65 ms, TR = 3,000 ms, field of view = 240 mm × 240 mm, 64 × 64 acquisition matrix, slice thickness = 5 mm, 20 slices, Nex = 2, Tau = 49).

### Data Processing

The CBF data were pre-processed using Statistical Parametric Mapping software (SPM8, Wellcome Department of Imaging Neuroscience, University College, London, UK) and ASLtbx (15, 16). The motion artifact was removed first using the realignment function. Specifically, the rigid body transform was used to estimate the motion time courses for all ASL's control and label images. Sinc interpolation of the ASL was then used to avoid the BOLD contamination. The time-matched control and label images were created, followed by subtraction to suppress BOLD contamination (16, 17). The CBF image series were generated based on a single compartment continuous ASL perfusion model using ASLtbx (16). Functional images were reoriented with the origin (i.e., the coordinate of x = 0, y = 0, and z = 0) set at the anterior commissure. Then the ASL images were co-registered to the corresponding anatomical images and then normalized to the MNI (Montreal Neurological Institute) space for group analysis (16). The registration performance of images was visually checked. These data were then smoothed using a Gaussian kernel of full-width half-maximum 8 mm. The CBF maps were constructed using an in-house program by applying a gray matter mask for the calculation, and the threshold (i.e., probability of gray matter) was set>0.8. The global and the regional CBFs of cluster with a size of at least 20 voxels were then obtained following the method proposed by Wang et al. (16).

The ASE data were acquired using different times ranging from 10 to 24 ms with an increment of 0.5 ms. These data were processed using a customized Matlab (MathWorks Inc. Natick, MA, USA) program (14, 18). To improve the signal-to-noise ratio, all the ASE images were first filtered using Gaussian lowpass filter (the kernel size was 3 × 3, and the standard deviation was 1.5).

The global and regional OEFs were then obtained using the method proposed by An and Lin (7). Specifically, the measurement of OEF and R2′ was derived from a theoretical model proposed by Yablonskiy and Haacke (19), in which a set of randomly orientated cylinders was used to characterize the signal behavior in static dephasing regime. The signal can be written as:

$$S(\mathbf{r}) = \rho \left(1 - \lambda\right) \cdot \mathbf{f}\left(\lambda, \mathbf{\hat{s}}\alpha, \mathbf{r}\right) \cdot \left(-\frac{TE}{T2}\right) \cdot \mathbf{g}\left(\mathbf{r}, T1, TR\right) \dots$$

where ρ was the effective spin density, λ was the volume fraction of muscle occupied by deoxyhemoglobin; where δω was the characteristic frequency shift and was defined as:

$$S\omega = \frac{4}{3}\pi \cdot \mathcal{V} \cdot \Delta\chi\_0 \cdot H\_{ct} \cdot B\_0 \cdot OEF\_\* $$

where γ was the gyromagnetic ratio; 1χ0was the susceptibility difference between the fully oxygenated and fully deoxygenated blood; Hct was the fractional hematocrit, B<sup>0</sup> was the main magnetic field strength, and 1χ<sup>0</sup> of 0.27 ppm per unit Hct in centimeter-gram-second units.

## Statistical Analysis

All the statistical analyses were performed by using the Matlab. To examine the effects of fatiguing aerobic exercise on the CBF, we first identified the regions with significant changes in CBF after the circle ergometer test using the paired-t test to compare the CBFs within each voxel before (i.e., baseline) and after the test. The significance level here was set as p < 0.01, and the threshold of cluster size (i.e., K) was set as 20. Then we further corrected the multi-comparison results using the false discovery rate (FDR). To examine the effects of fatiguing aerobic exercise on the OEF, we compared the global OEFs and regional OEFs before and after the test. Particularly, we focused on those FDRidentified regions with significant change in CBFs [i.e., regions of interest (ROIs)]. The FDR was also used to compare the OEF within the ROIs before and after the test. Secondarily, the association between the oxygen consumption (i.e., the VO2max, anaerobic threshold), the test time and the regional CBFs and OEFs were examined using linear regression. The Bonferroni correction was used in the multiple comparison.

# RESULTS

All the 11 participants [age: 20.3 ± 0.8 (mean ± S.D.) years, BMI: 21.6 ± 1.6] completed the circle ergometer test and two MRI scans. The time of the cycle ergometer test they maintained was 687 ± 72.1 s. The relative VO2max was 60.1 ± 3.3 ml/min/kg and the anaerobic threshold was 2882.3 ± 245.4 ml/min as measured by the gas analyzer.

Compared to the baseline, significant reduction in the global CBF was observed after exercise (p = 0.003, **Figure 1**). The CBFs in four brain regions, including left and right striatum, left middle temporal gyrus (MTG) and right inferior frontal gyrus (IFG), significantly decreased after completing the cycle ergometer test (K > 20, p < 0.005, **Figure 2**). However, no significant changes were observed in global and regional OEFs as compared to the baseline (p > 0.21, **Table 1**).

The OEF within the right striatum at baseline was associated with multiple functional performance, including the test time (r <sup>2</sup> = 0.63, p = 0.003, **Figure 3A**), the relative VO2max (r <sup>2</sup> = 0.51, p = 0.007, **Figure 3B**), and anaerobic threshold (r <sup>2</sup> = 0.66, p = 0.004, **Figure 3C**). Participants with greater OEF within the right striatum at baseline were able to maintain the test for a longer time, and/or had greater anaerobic threshold and relative VO2max. Neither the OEFs in other regions nor the CBFs at baseline were associated with those functional outcomes.

The percent change of CBF within the right IFG (r <sup>2</sup> = 0.55, p = 0.006, **Figure 4A**) and the change of OEF within the left striatum (r <sup>2</sup> = 0.52, p = 0.008, **Figure 4B**) was associated with test time. Participants who had less reduction of the CBF within the right IFG and/or of the OEF within the left striatum were able to maintain the test longer. Additionally, the anaerobic threshold was associated with the change of OEF within the left MTG (r <sup>2</sup> = 0.49, p = 0.009, **Figure 4C**), such that those with less reduction of OEF within the left MTG had greater anaerobic threshold.

*p* = 0.005) (A), left middle temporal gyrus (*p* < 0.0001) (B), and right inferior frontal gyrus (*p* = 0.005) (C) significantly decreased after the fatiguing aerobic exercise as

decreased (*p* = 0.003) after the aerobic exercise. Brighter color in the figure was greater CBF (unit: ml/100 g/min).

# DISCUSSION

compared to baseline.

The regulation of the blood flow and oxygen extraction in the cerebral regions is the fundamental for the neural activities, which are important for individual's capacity of enduring longterm aerobic exercises. By using advanced fMRI techniques measuring the CBF and OEF of brain regions, our pilot study has demonstrated the first-of-its-kind evidence that compared to the baseline, after the aerobic circle ergometer exercise with the load up to VO2max, the cerebral blood flow may decrease globally, and particularly within the left and right striatum, left MTG and right IFG that associate with voluntary motor control, sensory perception, and task motivation; but no significant changes in the global and regional OEFs are observed. Moreover, these neuroimaging metrics, which captures the metabolism of the brain, and their changes after the exercise are associated with the performance of the task (i.e., test time) and the energy consumption (i.e., anaerobic threshold, relative VO2max). These preliminary findings may thus provide unique insight into the mechanism underlying the regulation of cerebral hemodynamics



pertaining to the aerobic exercise, which are worthwhile to be confirmed in future study of larger sample size.

Studies have shown the benefits of aerobic exercise with mild to moderate physical load on brain health in young and old adults (20, 21). However, the effects of fatiguing aerobic exercise or tasks with high physical load on the functionalities of the brain remain unclear. The cerebral metabolism of oxygen (e.g., the metabolic rate of oxygen) relies on the CBF, OEF and the total oxygen content in the arterial blood (22). Here our results suggest for the first time that the diminished capacity of maintaining the

FIGURE 3 | The association between the baseline regional OEFs and functional performance. Participants with greater OEF within the right striatum at baseline sustained the cycle ergometer test longer [*r* <sup>2</sup> <sup>=</sup> 0.63, *<sup>p</sup>* <sup>=</sup> 0.003; (A)], and had greater the relative maximal oxygen consumption (VO2max) [*<sup>r</sup>* <sup>2</sup> = 0.51, *p* = 0.007; (B)], and anaerobic threshold [*r* <sup>2</sup> = 0.66, *p* = 0.004; (C)].

high-load aerobic exercise may be due at least in part to the decreased cerebral blood flow, and the altered ability of these brain regions to extract the oxygen maintains normally.

We observed the significant reduction of CBF within left and right striatum, left MTG and right IFG after the aerobic circle ergometer test. The striatum is the main structure of the basal ganglia, a central hub associated with multiple function, including the control of voluntary movement (23, 24) and task motivation (25, 26). Chaudhuri and Behan (27) have shown that the decreased activation of the basal ganglia alters the neural integrator and the cortical feedback. This dysfunction within the striato–thalamo–cortical loop is associated with the diminished physical function and increased fatigue in many neurodegenerative conditions, such as Parkinson's disease (28). Meanwhile, the MTG is associated with the multisensory integration (29) and the IFG has been linked to the motion inhibition and attention control (30). In our study, a continuous aerobic task with extremely high load up to 100% VO2max was used. The demand of the oxygen supply may thus increase over the maximal supply the vascular system is able to provide. As such, a potential "preserve" mechanism may be initiated: when the exercise is severely overloaded, the supply of oxygenated blood to the basal ganglia region, MTG and IFG decreases, leading to the diminished activation of striatum loop, less transmission of dopamine and declined sensory integration and attention. This helps prevent the body continuing the task of high risk, causing damages to our physiologic systems. Future studies are worthwhile to explore and confirm this potential mechanism by measuring the cerebral changes repeatedly along with the increase of task load.

We also observed that participants with greater resting OEF or less percent reduction in CBF was associated with greater time to maintain the aerobic test, and greater anaerobic threshold and relative VO2max. This may indicate that these markers derived from the cerebral hemodynamics are sensitive to the physical and physiologic function. Other studies have demonstrated the effects of exercise on the cerebral function and physical performance (20, 31). Leddy et al. (31), for example, reported that the aerobic exercise restored/enhanced the activation of the cerebral regions (e.g., anterior cingulate gyrus), as well as their physical function in those with post-concussion syndrome. Our study for the first time provide potential links between one's capacity to adapt to the aerobic exercise of high physical load to the cerebral function. The effects of biological aging and other pathological conditions on this relationship are worthwhile to be explored in future's longitudinal studies.

To measure the cerebral oxygen consumption, PET is still the most widely used method. However, it relies on the radiocontrast agent injected into the body (32), and many studies have shown that the radiocontrast agent is toxic and may cause adverse events (33, 34). Meanwhile, the low spatial resolution of other techniques, such as the TCD and NIRs, also limits their applications. We here implemented novel fMRI techniques (i.e., ASL and ASE sequences) to non-invasively measure the CBF and OEF in the brain. These advanced neuroimaging techniques shed light on characterizing the brain in future studies.

The limitation of this pilot study is that the sample size is small (n = 11) and currently we focus only on the cohort of elite young athletes. The effects of fatigue on the hemodynamics of the brain in other vulnerary populations, such as those suffering from the chronic fatigue syndrome, are needed to be explored. The observation in this pilot study may still be impacted by the vascular changes within the cerebral regions. Moreover, multiple underlying physiological characteristics may also contribute to the observed changes in CBF here, including the exerciserelated changes of adenosine triphosphate, hematocrit, and blood pressure, which, however, was not measured in this pilot study. Additionally, we focused the OEF on only the regions with significant changes in CBF. Future studies of larger sample size are thus warranted to explore and confirm the results of this pilot study by analyzing the regional OEF across multiple brain regions, and to explore the potential physiological pathways through which the fatigue affects the brain's hemodynamics by measuring those metrics. This pilot study nevertheless demonstrated the effects of fatiguing aerobic exercise on the cerebral hemodynamics and the extraction of oxygen in the brain using advanced neuroimaging techniques, revealing a potential preserve mechanism and providing several sensitive neuroimaging markers of physical and physiological function,

### REFERENCES


which may ultimately help the functional assessment in the sports science and clinics.

### ETHICS STATEMENT

This study was approved by Institutional Review Board of Beijing Sport University, and conducted according to the principles of the Declaration of Helsinki. All the participants provided written informed consent as approved by the institutional review board.

### AUTHOR CONTRIBUTIONS

DB, JZ, WJ, YHu, and XW designed the study. DB, YHa, and XY collected the data. JZ and YHa analyzed the data and performed statistical analyses. DB and JZ drafted the manuscript. All authors contributed to and approved the final version.

### FUNDING

This study was supported by Key Research and Development Projects of the Ministry of Science and Technology (grant number: 2018YFC2000602 and 2018YFC2000603). JZ was supported by the Irma and Paul Milstein Program for Senior Health Fellowship Award from the Milstein Medical Asian American Partnership (MMAAP) and Fudan Scholar program.


and cardiovascular fitness in aging. Front Aging Neurosci. (2013) 5:75. doi: 10.3389/fnagi.2013.00075


**Conflict of Interest Statement:** The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

Copyright © 2019 Bao, Zhou, Hao, Yang, Jiao, Hu and Wang. This is an open-access article distributed under the terms of the Creative Commons Attribution License (CC BY). The use, distribution or reproduction in other forums is permitted, provided the original author(s) and the copyright owner(s) are credited and that the original publication in this journal is cited, in accordance with accepted academic practice. No use, distribution or reproduction is permitted which does not comply with these terms.

# Blood Perfusion and Cellular Microstructural Changes Associated With Iron Deposition in Multiple Sclerosis Lesions

Huaqiang Sheng1,2 \*, Bin Zhao<sup>3</sup> and Yulin Ge<sup>2</sup> \*

<sup>1</sup> Department of Medical Imaging, Qianfoshan Hospital Affiliated to Shandong University, Jinan, China, <sup>2</sup> Department of Radiology, New York University School of Medicine, New York, NY, United States, <sup>3</sup> Department of Medical Imaging Research Institute, Shandong University, Jinan, China

### Edited by:

Achim Gass, Universitätsmedizin Mannheim (UMM), Germany

### Reviewed by:

Robert Zivadinov, University at Buffalo, United States Andrea Bink, University Hospital Zurich, Switzerland

### \*Correspondence:

Huaqiang Sheng 15898616257@163.com Yulin Ge yulin.ge@nyumc.org

### Specialty section:

This article was submitted to Applied Neuroimaging, a section of the journal Frontiers in Neurology

Received: 01 April 2019 Accepted: 26 June 2019 Published: 11 July 2019

### Citation:

Sheng H, Zhao B and Ge Y (2019) Blood Perfusion and Cellular Microstructural Changes Associated With Iron Deposition in Multiple Sclerosis Lesions. Front. Neurol. 10:747. doi: 10.3389/fneur.2019.00747 Background and Purpose: Susceptibility-weighted imaging (SWI) has emerged as a useful clinical tool in many neurological diseases including multiple sclerosis (MS). This study aims to investigate the relationship between SWI signal changes due to iron deposition in MS lesions and tissue blood perfusion and microstructural abnormalities to better understand their underlying histopathologies.

Methods: Forty-six patients with relapsing remitting MS were recruited for this study. Conventional FLAIR, pre- and post-contrast T1-weighted imaging, SWI, diffusion tensor imaging (DTI), and dynamic susceptibility contrast (DSC) perfusion MRI were performed in these patients at 3T. The SWI was processed using both magnitude and phase information with one slice minimal intensity projection (mIP) and phase multiplication factor of 4. MS lesions were classified into 3 types based on their lesional signal appearance on SWI mIP relative to perilesional normal appearing white matter (peri-NAWM): Type-1: hypointense, Type-2: isointense, and Type-3: hyperintense lesions. The DTI and DSC MRI data were processed offline to generate DTI-derived mean diffusivity (MD) and fractional anisotropy (FA) maps, as well as DSC-derived cerebral blood flow (CBF) and cerebral blood volume (CBV) maps. Comparisons of diffusion and perfusion measurements between lesions and peri-NAWM, as well between different types of lesions, were performed.

Results: A total of 137 lesions were identified on FLAIR in these patients that include 40 Type-1, 46 Type-2, and 51 Type-3 lesions according to their SWI intensity relative to peri-NAWM. All lesion types showed significant higher MD and lower FA compared to their peri-NAWM (P < 0.0001). Compared to Type-1 lesions (likely represent iron deposition), Type-2 lesions had significantly higher MD and lower FA (P < 0.001) as well as lower perfusion measurements (P < 0.05), while Type 3 lesions had significantly higher perfusion (P < 0.001) and lower FA (P < 0.05). Compared to Type-2, Type-3 lesions had higher perfusion (P < 0.0001) and marginally higher MD and lower FA (P < 0.05).

**73**

Conclusion: The significant differences in diffusion and perfusion MRI metrics associated with MS lesions, that appear with different signal appearance on SWI, may help to identify the underlying destructive pathways of myelin and axons and their evolution related to inflammatory activities.

Keywords: DTI (diffusion tensor imaging), susceptibility-weighted imaging, multiple sclerosis (MS), PWI = perfusion-weighted imaging, MRI - magnetic resonance imaging

### INTRODUCTION

Multiple sclerosis (MS) is an inflammatory autoimmune neurodegenerative disease of the central nervous system (CNS), characterized by inflammation, demyelination, gliosis and neuroaxonal loss in lesions. It is generally believed that the basic pathogenesis of MS is collapse of immune tolerance to CNS myelin or myelin-like antigens followed by pro-inflammatory phagocytosis, oxidative injury, antigen presentation and T cell co-stimulation (1). Demyelinating and axonal injury are further consequences, which are typical features of MS. As we have already known, the progressive neurodegenerative processes in MS take a great toll on physical disability and cognitive disorder (2), and can seriously impact the quality of life in patients. Recent studies have shown that the changes of iron content that are commonly seen in MS lesions may be related to inflammatory activities (e.g., active myelin phagocytosis and intracellular iron depletion) and oxidative tissue injury in the demyelinating disease (3–6). Some other studies have found that iron is closely related to the biosynthetic enzymes of myelin formation (7, 8). Public opinions are divergent, but the effect of iron deposition on cellular and microstructural changes in the MS lesions remains an unresolved issue.

MRI has had an enormous impact on MS and plays a critical role as a paraclinical tool in routine clinical practice. The multi-sequence or multi-contrast MR imaging not only improves the diagnosis but also provides different specificity for various elements of pathology including iron deposition and microstructural destruction (9). Susceptibility weighted imaging (SWI) (10), as a three dimensional high resolution gradient echo sequence, is extensively applied for detecting abnormal iron deposition or microbleeds in MS (11). Compared to conventional T1- and T2-weighted MRI, SWI is more superior in displaying paramagnetic dark or hypointense signals, including the iron content in various forms of hemosiderin, ferritin and ironladen macrophage (12–15) with high sensitivity even with only 1 gFe/g tissue iron changes (16). Studies have shown that MS lesions can also appear as an isointense or hyperintense signal on SWI with unclear pathophysiological implications (16, 17). It is therefore essential to identify the pathophysiological meaning of different SWI signal appearances of MS lesions using noninvasive imaging to fulfill this unmet need.

Recently, quantitative imaging measures have been increasingly used in MS research to better elucidate the hidden pathological mechanisms associated with tissue microstructural and inflammatory changes (9). Among these techniques, diffusion tensor imaging (DTI) (18–20) and dynamic susceptibility contrast MRI (DSC-MRI) (21, 22) are gaining more wide-spread utility in clinical practice and have shown great potential for detecting the cellular microstructural integrity and hemodynamic impairment at different stages of lesion evolution in MS, respectively. The aim of this study is to characterize the quantitative DTI-derived diffusion and DSC-derived perfusion parameters changes underlying different SWI signal intensities of MS lesions. We hypothesized that signal intensities detected on SWI in MS lesions may be a noninvasive biomarkers that can help clinicians to determine specific pathological processes associated with demyelination, axonal loss, and inflammatory processes in patients with relapsing-remitting MS.

## MATERIALS AND METHODS

### Subjects

The research protocol of this retrospective study followed the tenets of the Declaration of Helsinki and was approved by the New York University Langone Health (NYULH) Institutional Review Board. Forty-six clinically definite relapsing remitting MS patients (28 women, 18 men, mean age 35.9 ± 11.3 years) enrolled from January 2012 to December 2016, were used in this study. All patients were informed and signed the institutional review board approved written consent form. The median disease duration in these patients was 4.4 years (range 1.6–11.4 years) and the median expanded disability status scale (EDSS) was 3.5 (range 1.5–5.5). These patients had no history of cerebrovascular disease, evidence of small vessel ischemic disease and no substantial intracranial pathology besides MS lesions in MR imaging.

### Image Acquisition and Processing

All patient data were acquired on a 3.0T Trio (Siemens Medical Solutions, Erlangen, Germany) MR scanner using a 20-channel array head coil. The MRI protocol included the following sequences: (1) Fluid-attenuated inversion recovery (FLAIR) imaging (TR/TE=9420/134 ms, voxel size = 1 × 1 × 3 mm<sup>3</sup> ); (2) pre and post T1-weighted (T1W) imaging (TR/TE=630/15 ms, voxel size = 1 × 1 × 3 mm3); (3) susceptibility weighted imaging (SWI) (TR/TE=28/20 ms; FA=15◦ , voxel size =0.86 × 0.86 × 3 mm<sup>3</sup> ); (4) DTI with 30 directions (TR/TE = 7300/89 ms, voxel size = 3.0 × 3.0 × 3.0 mm<sup>3</sup> , b = 1000 s/mm<sup>2</sup> ); (5) dynamic susceptibility contrast (DSC) perfusion imaging (TR/TE = 956/32 ms, voxel size = 1.7 × 1.7 × 3.0 mm<sup>3</sup> ) applied to 13 axial slices centered at lateral ventricle body with 10 seconds injection delay. For DSC, a 3–5 cc/sec bolus of Gadolinium contrast agent (Gd-DTPA; Magnevist, Bayer Schering Pharma) was administered at a dose of 10– 20 cc (0.075 mmol/kg) to acquire 60 time points. The postcontrast T1-weighted imaging (the same sequence with precontrast) was performed 10 min after injection. The image slice thickness from all sequences above is the same for lesion identification and registration on different imaging contrast. All sequences had 45 slices (13.5 cm) coverage of brain except DSC. The total scan time for all sequences was about 45 min.

SWI data is processed using an in-house image-processing software (SPIN) (23). The raw magnitude and phase from each SWI scan used to generate minimal intensity projection (mIP) using phase multiplication factor of 4 to enhance the susceptibility effects. Instead of using multiple slices for mIP, one slice mIP was used in this study to keep the slice thickness the same with the rest sequences and to minimize the partial volume effects from multi-slice mIP. DTI data analysis was performed offline using DTI studio, by which tensor images were generated to construct mean diffusivity (MD) and fractional anisotropy (FA) (24). MD and FA are the scalar measures of the total diffusion (e.g., average of eigenvalues) within a voxel and the degree of anisotropy in a given voxel, respectively. DSC data was processed using the perfusion analysis software package in Olea Sphere (Olea Medical, Cambridge, MA). Data first underwent preprocessing consisting of motion correction followed by spatial and temporal filtering. The standard single value decomposition (SVD) technique was then applied to the preprocessed data to generate maps of mean transit time (MTT), CBF, and leakagecorrected CBV (25). Because the absolute values of CBF (ml/100 ml/min) and CBV (ml/100 ml) can only be determined up to a multiplicative constant, the comparisons between lesion types were used as relative measures (i.e., rCBF, rCBV) in this study. Lastly, the diffusion and perfusion maps were manually registered to their corresponding conventional T1 and FLAIR imaging as well as SWI images using tkregister2 (Free Surfer, Massachusetts General Hospital, Harvard Medical School) for manually ROI placement and analysis.

### Data and Statistical Analysis

As shown in **Figure 1**, according to signal intensity appearances on SWI mIP, MS lesions were classified into three distinct lesion types. Type-1: hypointense (i.e., higher susceptibility), Type-2: isointense, and Type-3: hyperintense lesions. To avoid the visual predisposition bias, a cut-off value of 30% difference of mean intensity, measured between lesions and perilesional region, was applied. Only lesions with a diameter of 5 mm or larger were included in the data analyses. These lesions were first blindly reviewed and classified by each of the two experienced radiologists, and finally determined by consensus between the two for lesions with inconsistent opinion. Quantitative data analyses of diffusion and perfusion measurements were performed with Image J (National Institutes of Health, Bethesda, MD) software. Lesions were identified on conventional FLAIR, T1-weighted, and SWI images, on which the anatomical regions of interest (ROIs) were manually selected and then transferred onto co-registered FA, MD, CBF, and CBV maps. For each lesion, the ROI was placed on both lesion and perilesional NAWM (peri-NAMW) region for comparison. In order to increase the accurate lesion selection and avoid partial volume, the image with the lesion target was zoomed-in 3 times bigger on ImageJ for better ROI placement. On this magnified view, the ROI placement of peri-NAWM was also improved. Mixed model analysis of covariance (ANCOVA) was used to compare the lesions of each type to the perilesional normal appearing white matter (peri-NAWM) and to compare lesions of different types to each other with respect to FA, MD, rCBF, and rCBV. A separate univariate analysis was conducted for each perfusion measure. When the value of P < 0.05, the difference is considered to be statistically significant.

# RESULTS

A total of 137 lesions were identified on conventional T2 weighted and post-contrast T1-weighted imaging in 46 patients with relapsing remitting MS that had both DTI and DSC data. Among them, there were 40 (or 29.2%) Type-1, 46 (or 33.6%) Type-2, and 51 (or 37.2%) Type-3 lesions (**Figure 1**). In addition, there were 11 enhancing lesions found in 6 patients; and 9 of these enhancing lesions were Type-3 lesions that showed hyperintensity on SWI, and another 3 enhancing lesions were Type-2 lesions that show isointensity on SWI. In contrast, none of the Type-1 lesions (hypointense on SWI) showed Gadolinium enhancement.

As shown in **Figure 2**, compared to peri-NAWM measurements, FA was significantly lower and MD was significantly higher in all types of lesions (P < 0.0001), indicating clear microstructural disruption in MS lesions. The mean FA values of Type-1, Type-2, and Type-3 lesions were 0.31 ± 0.05, 0.24 ± 0.07, 0.27 ± 0.08, respectively; and the mean FA values for their corresponding peri-NAWM were 0.49 ± 0.11, 0.52 ± 0.09, 0.45 ± 0.12 respectively. The mean MD values of Type-1, Type-2, and Type-3 lesions were 1.16 ± 0.27, 1.42 ± 0.34, 1.27 ± 0.36, respectively; and the mean MD values for their corresponding peri-NAWM were 0.71 ± 0.16, 0.68 ± 0.17, 0.82 ± 0.09, respectively. For perfusion measures, both CBF and CBV in Type-3 lesions (297.6 ± 126.5 ml/100 g/min, 385.9 ± 142.9 ml/100 g) showed significantly higher than peri-NAWM (216.9 ± 80.6 ml/100 g/min, 234.7 ± 75.6 ml/100 g) with P = 0.0002 and P < 0.0001, respectively. Type 2 lesions showed significantly lower CBF than peri-NAWM (158.6 ± 77.1 vs. 193.7 ± 82.3 ml/100 g/min, P = 0.04) and significantly lower CBV (206.4 ± 95.1 vs. 257.4 ± 89.1 ml/100 g, P = 0.009). However, Type-1 lesions didn't show a significant difference in perfusion measurements with peri-NAWM.

The DTI-derived mean FA and MD values as well as DSCderived rCBF and rCBV values of three types of lesions and their comparisons (P-values) are summarized in **Table 1**. Compared to Type-1 lesions, Type-2 lesions showed significantly higher MD and lower FA. Compared to Type-1 lesions, Type-3 lesions only showed significant difference in MD (P = 0.036) but not in FA. Compared to Type-3 lesions, Type-2 lesions showed

marginally higher MD and lower FA (P = 0.04). The mean rCBF and rCBV were the lowest in Type-2 lesions and were the highest in Type-3 lesions with Type 1 lesions being in the middle. The increased blood perfusion in Type-3 lesions may be associated with vascular inflammatory activities since most enhancing lesions (9 out of 11) were Type-3 lesions.

Examples of diffusion and perfusion imaging parameter characteristics of Type-1 lesions were shown in **Figure 3**.


The values were reported in mean ± standard deviation. The unit for MD is in mm<sup>2</sup> /s. The reported rCBF and rCBV are relative (i.e., ratio) measurements and FA is an index for the amount of diffusion asymmetry between 0 and 1 within a voxel, therefore, they don't have absolute units. The intensity of different types of lesions is relative to the surrounding white matter on SWI.

FIGURE 3 | Representative images of Type-1 lesions in two MS patients (top row from a 36-year-old male patient and bottom row from a 37-year-old female patient) include FLAIR (A,A′ ), Gd-enhanced T1-weighted (B,B′ ), and SWI (C,C′ ), as well as parameter maps of MD (D,D′ ), FA (E,E′ ), and CBF (F,F′ ). The hypointense lesions on SWI (arrows) are associated with a less significant change in diffusion and perfusion measurements, as compared perilesional NAWM.

As shown in one patient (in **Figure 3** top row), SWI was most sensitive in detecting iron-laden component of lesions. The hypointense Type-1 lesions demonstrated a significant increase in MD and decrease in FA but no change in perfusion measurements compared to perilesional NAWM. Similarly, in another patient (**Figure 3** bottom row), visible changes of MD and FA can be seen in another Type-1 lesion compared to peri-NAWM with uncertain perfusion changes. Representative Type-2 lesions were shown in **Figure 4,** in which SWI lesions that appeared as slightly hypointense (top row) or isointense (bottom row) showed a remarkable increase in MD and decrease in FA as well as reduced CBF. Such lesions in **Figure 4** (top row) also showed hypointensity on both FLAIR and post-contrast T1-weighted images. Two Type-3 lesions with Gadolinium enhancement were shown in **Figure 5,** in which there is a mild increase in MD and marked decrease in FA as well as increase in CBF. One MS lesion with both Type-2 and Type-3 components was shown in **Figure 6**, in which the lesion showed a mixed pattern of significant diffusion and perfusion changes associated with non-enhancing center region and the enhancing rim, respectively.

FLAIR (A,A′ ), Gd-enhanced T1-weighted (B,B′ ), SWI (C,C′ ), as well as parametric maps of MD (D,D′ ), FA (E,E′ ), and CBF (F,F′ ). The hyperintense lesions (arrow) on SWI showed gadolinium enhancement that is corresponding to increased perfusion and slightly increased MD as well as decreased FA.

# DISCUSSION

Conventional MRI offers the most sensitive way to detect MS lesions and their changes over time for ruling in or ruling out a diagnosis of MS and for disease followup monitoring. The addition of SWI, which is a quick scan of routine conventional MRI protocols, may provide in vivo pathophysiological insights into cellular microstructural injury and tissue hemodynamic changes. Our results of MS lesions on SWI, combining quantitative multi-contrast and multi-parameter MRI, suggest that the intensity-based lesion types on SWI may represent a specific stage of lesion evolution or a certain pathological substrate associated with demyelination/axonal injury or inflammatory activity. These

FIGURE 6 | An ring-enhancing lesion in a 32-year-old female MS patient on FLAIR (A), Gd-enhanced T1-weighted (B), and SWI (C), as well as on MD (D), FA (E), and rCBF (F) parametric maps. The lesion has both Type-2 (isointense center on SWI) and Type-3 (hyperintense rim on SWI) components. The center of the lesion demonstrated larger MD changes compared to the enhancing rim that has increased perfusion of the entire lesion.

hidden pathological changes including blood-brain barrier dysfunction can possibly be detected with SWI without using Gadolinium contrast agent (26) as shown in Type-3 lesions. Our data also confirm previous imaging-histopathological correlative evidence of iron deposition, demyelination and axonal loss (6, 27, 28). In particular, three major observations emerge from these data. First, a hyperintense (Type-3) lesion on SWI may be related to the underlying enhanced vascular inflammatory activity with increased BBB disruption that results in increased CBF and CBV. Second, hypointense (Type-1) lesions on SWI are likely to have less tissue destruction by diffusion measures compared to Type-2 lesions; they also have less inflammatory activity than Type-3 lesions by perfusion measures. Third, isointense SWI Type-2 lesions may represent a more chronic demyelinated plaque with irreversible tissue destruction (e.g., black holes) showing the most severe DTI-derived diffusion changes.

SWI is a 3D gradient-echo high-resolution sequence that is fully flow-compensated with long-echo and combines magnitude and filtered-phase information to enhance susceptibility effects due to paramagnetic substances, such as hemosiderin and deoxyhemoglobin (10). Unlike quantitative susceptibility mapping (QSM), SWI is considered to be a qualitative MRI technique for enhanced lesion detection, its unique image contrast is particularly useful to gauge tissue iron content and venous structures. Therefore, it is well-recognized that hypointense (Type-1) lesions of MS are corresponding to increased iron content, which is likely due to chronic inflammatory activity with an elevated number of microglia and macrophages that contain high amounts of iron (15, 29). Type-1 lesions are also likely caused by increased hemosiderin (30) from old blood products leaked from inflammation-induced damaged vessels. All Type-1 lesions in this study were not enhancing on post-contrast T1-weighted imaging even though some lesions showed slightly increased blood perfusion, indicating certain inflammatory activities or lesion reactivity with increased macrophage cells (21). Compared to Type-2 lesions, these Type 1 lesions showed less diffusion abnormalities, which are believed to be corresponding to a lower degree of cellular architecture destruction during a tissue repair stage (8, 31).

Most SWI studies of MS have been focusing on hypointense (Type-1) lesions. In this study, besides the hypointense SWI lesions, we have also characterized quantitative diffusion and perfusion imaging features of isointense and hyperintense SWI lesions. Out of 137 MS lesions, 33.6% are Type-2 isointense lesions and 37.2% are Type-3 hyperintense lesions. Since mIP is a post-processed image using phase multiplication (with a factor of 4 in this study) and minimal intensity projection algorithm (10), the true meaning of signal intensity on multislice mIP images is uncertain. Therefore, in this study, the mIP image was generated using only one slice, in order to avoid the mixture effects of projected intensities. Except for high susceptibility substances (e.g., non-heme iron or venous blood) that contribute to dark signal on SWI, the non-dark signal on SWI is likely due to the combined effects of the amount free water content and edema (non-free intra- or extra-cellular water) due to pro-inflammation activation status. SWI does not provide a typical T1- or T2-weighted imaging contrast. After being applied with phase information, it does not seem to be a standard T2<sup>∗</sup> contrast either. Although the non-iron laden isointense or hyperintense SWI lesions have been consistently shown in the literatures and represent most MS lesions (16, 17, 32, 33), their histopathological characteristics are unclear. In this study, we found most isointense Type-2 lesions on SWI are corresponding to isointense or hypointense (or black hole) lesions on T1-weighted imaging. The well-demarcated black hole lesions on T1-weighted imaging likely represent the hypocellular area characterized by necrotic fluid elements and the loss of tissue structures.

Our results of combining quantitative diffusion and perfusion measurements support the notion that SWI can be used as a promising alternative in determining the underlying histopathological hallmarks of MS lesions. We found that Type-1 lesions have less diffusion changes than Type-2 lesions and less perfusion changes than Type-3 lesions, despite Type-1 lesions usually containing iron deposition. According to the previous study, the origins of iron deposition in MS lesions may be the concentrated iron of macrophages, debris of oligodendrocyte and myelin, or hemosiderin of hemorrhagic products from the leaky vessels (34, 35). The exact role of iron in MS is unclear with views from both sides that iron can either contribute to chronic inflammation, oxidative stress and neurodegeneration (35) or contribute to tissue repair (31). The slightly increased perfusion measurements (e.g., CBF and CBV) of Type-1 lesions found in this study may support the increased inflammation and cell activities.

To the best of our knowledge, this is the first time to characterize the hyperintense signal (Type-3) lesions on SWI with diffusion and perfusion measurements. Type-3 lesions showed significantly increased rCBF and rCBV but less diffusion changes compared to Type-2 lesions, indicating local vascular inflammation induced vasodilation and increased perfusion in these lesions (21, 36). This is also indicated by that fact that 9 out of 11 enhancing lesions in these patients are Type-3 lesions. Based on QSM analysis, Zhang et al. (26) showed gadoliniumenhancing MS lesions had relative low QSM values than nonenhancing lesions, which are consistent with the findings in this study that these enhancing lesions appear hyper- rather hypointense on SWI. Another study (32) has also demonstrated that enhancing lesions are likely to be hyperintensities in contrast to the central dark vein on post-gadolinium SWI images, despite that gadolinium is a paramagnetic agent and has strong T2<sup>∗</sup> shortening effect. These results suggest that signal intensities on SWI may help better detect BBB dysfunction and identify subtle inflammatory activities that are not detected on post-contrast T1-weighted imaging (35). The marginal or no difference of diffusion measurements between Type-3 and Type-1 lesions, as well as between Type-3 and Type-2 lesions, indicate that there is a large span of variabilities for microstructural changes in MS lesions depending on stages of lesion development and evolution. However, Type-2 lesions demonstrated the highest MD and lowest mean FA, suggesting most severe architecture destruction and tissue loss in these lesions; and Type-1 lesions showed the lowest mean MD, which may suggest a certain level of water diffusion restriction [i.e., cytotoxic edema from hypoxia injury (37, 38)] occurs in these lesions during high level of macrophage activities.

There were several limitations associated with this study. First, due to the challenge for quantifying the absolute CBF and CBV using DSC MRI (39, 40) due to the uncertainties of scaling coefficients for relaxivity and AIF partial volume, we used relative perfusion measures for comparison between lesion types. For comparisons between lesions and peri-NAWM, although we used the actual CBF and CBV values from DSC SVD algorithms, the values reported in these tissues are supposed to be interpreted for comparisons only. Future longitudinal studies are warranted for validating some of the findings regarding the underlying histopathology of lesion development and progression, in particular with a large sample size of patients with enhancing lesions. Lastly, the definition of different types of lesions was using 30% signal intensity difference may be arbitrary, however, we found the classification based on such a threshold provided appropriate differentiable imaging features from DTI and DSC data.

# CONCLUSION

This study indicates that the addition of SWI to clinical MRI protocol may provide in vivo pathological insights, suggesting that the intensity-based lesion types on SWI may represent a specific stage of lesion evolution or a certain pathological substrate associated with iron deposition, demyelination/axonal injury or inflammatory activity. Further studies investigating the longitudinal evolution of lesion appearances on SWI and their quantitative correlations will be envisioned.

# DATA AVAILABILITY

All datasets generated for this study are included in the manuscript and/or the supplementary files.

# ETHICS STATEMENT

This study was carried out in accordance with the recommendations of name of guidelines, name of committee; with written informed consent from all subjects. All subjects gave written informed consent in accordance with the Declaration of Helsinki. The protocol was approved by the local ethics committee of the Institutional Review Board Human Research Protection Program.

### AUTHOR CONTRIBUTIONS

YG and HS participated in the design of this study, and they both performed the statistical analysis. HS collected

### REFERENCES


important background information. BZ carried out literature search and interpretation of the data. All authors contributed to the construction of manuscript and its critical revision.

relevance of classified lesions to disease status. J Neurol Neurophysiol. (2014) 2014:12. doi: 10.4172/2155-9562.S12-012


inflammatory multiple sclerosis patients. PLoS ONE. (2015) 10:e0119356. doi: 10.1371/journal.pone.0119356


quantitative cerebral blood flow using corrections for partial volume and nonlinear contrast relaxivity: a xenon computed tomographic comparative study. J Magn Reson Imaging. (2009) 30:743–52. doi: 10.1002/jmri. 21908

**Conflict of Interest Statement:** The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

Copyright © 2019 Sheng, Zhao and Ge. This is an open-access article distributed under the terms of the Creative Commons Attribution License (CC BY). The use, distribution or reproduction in other forums is permitted, provided the original author(s) and the copyright owner(s) are credited and that the original publication in this journal is cited, in accordance with accepted academic practice. No use, distribution or reproduction is permitted which does not comply with these terms.

# Altered Neurometabolic Profile in Early Parkinson's Disease: A Study With Short Echo-Time Whole Brain MR Spectroscopic Imaging

Martin Klietz 1†, Paul Bronzlik 2†, Patrick Nösel <sup>2</sup> , Florian Wegner <sup>1</sup> , Dirk W. Dressler <sup>1</sup> , Mete Dadak <sup>2</sup> , Andrew A. Maudsley <sup>3</sup> , Sulaiman Sheriff <sup>3</sup> , Heinrich Lanfermann<sup>2</sup> and Xiao-Qi Ding<sup>2</sup> \*

### Edited by:

Fabiana Novellino, Italian National Research Council (CNR), Italy

### Reviewed by:

Bruno J. Weder, University of Bern, Switzerland Gaetano Barbagallo, Università degli Studi Magna Græcia di Catanzaro, Italy

### \*Correspondence:

Xiao-Qi Ding ding.xiaoqi@mh-hannover.de

†These authors have contributed equally to this work

### Specialty section:

This article was submitted to Applied Neuroimaging, a section of the journal Frontiers in Neurology

Received: 01 February 2019 Accepted: 03 July 2019 Published: 17 July 2019

### Citation:

Klietz M, Bronzlik P, Nösel P, Wegner F, Dressler DW, Dadak M, Maudsley AA, Sheriff S, Lanfermann H and Ding X-Q (2019) Altered Neurometabolic Profile in Early Parkinson's Disease: A Study With Short Echo-Time Whole Brain MR Spectroscopic Imaging. Front. Neurol. 10:777. doi: 10.3389/fneur.2019.00777 <sup>1</sup> Department of Neurology, Hannover Medical School, Hanover, Germany, <sup>2</sup> Department of Neuroradiology, Hannover Medical School, Hanover, Germany, <sup>3</sup> Department of Radiology, University of Miami School of Medicine, Miami, FL, United States

### Objective: To estimate alterations in neurometabolic profile of patients with early stage Parkinson's disease (PD) by using a short echo-time whole brain magnetic resonance spectroscopic imaging (wbMRSI) as possible biomarker for early diagnosis and monitoring of PD.

Methods: 20 PD patients in early stage (H&Y ≤ 2) without evidence of severe other diseases and 20 age and sex matched healthy controls underwent wbMRSI. In each subject brain regional concentrations of metabolites N-acetyl-aspartate (NAA), choline (Cho), total creatine (tCr), glutamine (Gln), glutamate (Glu), and myo-inositol (mIns) were obtained in atlas-defined lobar structures including subcortical basal ganglia structures (the left and right frontal lobes, temporal lobes, parietal lobes, occipital lobes, and the cerebellum) and compared between patients and matched healthy controls. Clinical characteristics of the PD patients were correlated with spectroscopic findings.

Results: In comparison to controls the PD patients revealed altered lobar metabolite levels in all brain lobes contralateral to dominantly affected body side, i.e., decreases of temporal NAA, Cho, and tCr, parietal NAA and tCr, and frontal as well as occipital NAA. The frontal NAA correlated negatively with the MDS-UPDRS II (R = 22120.585, p = 0.008), MDS-UPDRS IV (R = −0.458, p = 0.048) and total MDS-UPDRS scores (R = −0.679, p = 0.001).

Conclusion: In early PD stages metabolic alterations are evident in all contralateral brain lobes demonstrating that the neurodegenerative process affects not only local areas by dopaminergic denervation, but also the functional network within different brain regions. The wbMRSI-detectable brain metabolic alterations reveal the potential to serve as biomarkers for early PD.

Keywords: Parkinson's disease, whole brain, MRI, spectroscopy, biomarker, early diagnosis

# INTRODUCTION

Parkinson's disease (PD) is characterized by symptoms of rigidity, bradykinesia, tremor, and postural instability. Diagnosis is based on clinical findings, with an accuracy of only 53% for disease duration <5 years, increasing to 88% for durations longer than 5 years (1). Magnetic resonance imaging (MRI) reveals in PD patients only unspecific brain changes and is used mainly to exclude differential diagnoses (2–6). Magnetic resonance spectroscopy (MRS) can be used to measure brain metabolites like N-acetyl-aspartate (NAA), choline (Cho), myo-inositol (mIns), total creatine (tCr), glutamine (Gln), and glutamate (Glu), which provide information about neuronal integrity (NAA), membrane turnover (Cho), gliosis (mIns), energy metabolism (Cr), and glutamatergic neuronal activity (Glu, Gln) in patients. Numerous MRS studies on PD have been reported previously (7–11). Due to methodical limitations of commonly used MRS techniques that suffered from limited spatial coverage, most studies reported PD-related metabolic changes in one or a few small brain structures. These results thus may not necessarily reflect the metabolic status within whole brain. Considering that human brain functions as organized networks with interactions between different multiple brain regions (10, 12), information about PD-related metabolic alterations within the whole brain with high spatial resolution may help to better characterize PD and understand the underlying pathologic mechanisms. A recently established whole-brain MR spectroscopic imaging (wbMRSI) technique provides the possibility to measure brain metabolites simultaneously over different larger brain scales in subjects in vivo (13), as well as in multiple specific small brain areas (14, 15). Therefore, we are going to study altered brain metabolism in PD patients systematically by use of the wbMRSI. As a first part of the project we aimed to obtain an overview about altered neurometabolic profile in early PD by exploring metabolic changes in eight brain lobes and cerebellum that composed the whole brain, with the results being reported in the following.

# PATIENTS AND METHODS

# Patients and Clinical Examinations

Human subject studies were carried out with approval from the local Ethics Committee of Hannover Medical School (No. 6167- 2016) and all subjects gave written informed consent. PD patients were recruited from those treated at the neurological wards and movement disorders outpatient clinic of Hannover Medical School. Inclusion criteria were the neurological diagnosis of PD according to the Movement Disorder Society (MDS) diagnosis criteria with a Hoehn and Yahr stage (H&Y) of 1 or 2 in the best medical on state and age of 75 or below. Definition of early stage PD by the H&Y stage is in accordance with (7–13), additionally, none of our patients complained of a significant amount of motor complications qualifying for advanced PD [see (16) for review]. Patients with atypical Parkinsonism and other known brain pathologies e.g., stroke, small vessel disease or tumor, were excluded. Additionally, patients with severe head tremor, dystonia or dyskinesia had to be excluded from this study.

A movement disorders specialist enrolled the PD patients. Twenty PD patients (48–72 years old, mean age 60.2 ± 7.2 years, 8 males) were included. Information about course of PD in the individual patient was collected, including disease duration, dominantly affected body side, main symptoms, medication, and comorbidities. PD symptoms were assessed by the Movement Disorders Society Unified Parkinson's Disease Rating Scale (MDS-UPDRS) (17). Patients were rated in best medication "on" state. PD specific medication was noted and levodopa equivalence dosage calculated (LED). Cognitive deficits were quantified by the established test for dementia and mild cognitive impairment DemTect (18). As controls 20 healthy participants matched in age and sex on a one-to-one basis were also studied. All patients and healthy controls were right-handed according to self-report.

## MR Examinations

All subjects underwent MR examinations at 3T (Verio, Siemens, Erlangen, Germany). The routine MRI protocol included a T2 weighted turbo spin echo (TSE) sequence, a T2 weighted gradient echo (GRE) sequence, a fluid attenuation inversion recovery (FLAIR) sequence, a T1 weighted 3D magnetization-prepared rapid gradient-echo (MPRAGE) sequence, and a volumetric spin-echo planar spectroscopic imaging (EPSI) acquisition (TR/TE = 1550/17.6 ms, field-of view of 280 × 280 × 180 mm<sup>3</sup> , matrix size of 50 × 50 with 18 slices with a nominal voxel volume of 0.31 ml (=5.6 × 5.6 × 10 mm<sup>3</sup> ), echo train length of 1,000 points, and bandwidth of 2,500 Hz) for wbMRSI, as described previously (13, 19). The scan time with EPSI was about 17 min. The EPSI-acquisition included also a second dataset obtained without water suppression, which was used for several processing functions, including measurement and correction of the resonance frequency offset at each voxel location, correction of lineshape distortions and to provide internal signal reference for the normalization of metabolite concentrations (20), while the MPRAGE images were used as anatomical reference. The EPSI, MPRAGE, FLAIR, TSE, and GRE scans were obtained with the same angulation.

# Data Processing

MPRAGE, FLAIR, TSE, and GRE images were inspected to recognize possible morphological abnormalities, which were done by two neuroradiologists. The EPSI data were processed using the MIDAS software package to obtain volumetric metabolite maps. Processing included zero-filling to 64 × 64 × 32 points and spatial smoothing, resulting in an interpolated basic voxel volume of 0.107 ml (4.375 × 4.375 × 5.625 mm<sup>3</sup> ) and an effective voxel volume of 1.5 ml (13, 19). The processing also included calculation of the fractional tissue volume contributing to each MRSI voxel, which used a tissue segmentation (18, 19) of the T1-weighted MPRAGE data to map gray matter (GM), white matter (WM), and cerebrospinal fluid (CSF). All resultant maps were then spatially transformed and interpolated to a standard spatial reference (20) at 2 mm isotropic resolution, which was associated with an atlas that mapped the individual brain lobes and the cerebellum. Mean regional metabolite concentrations were then determined in atlas-defined brain lobes and cerebellum, which composed the whole brain (**Figure 1**): The frontal lobe left (LFL) and right (RFL) including anterior parts of the striatum and pallidum, the temporal lobe left (LTL) and right (RTL) including posterior parts of the striatum and pallidum, the parietal lobe left (LPL) and right (RPL) including thalamus and subthalamic nucleus, the occipital lobe left (LOL) and right (ROL), and the cerebellum (Cbl). To obtain brain regional metabolite concentrations, especially to estimate the metabolites with small MRS signal amplitudes (Glu and Gln) separately, a modified data analysis approach suggested by Goryawala et al. was applied, i.e., the spectra were averaged by summing voxels within a region of interest (ROI) to obtain high-SNR spectra from atlas-registered anatomic regions, which was done following inverse spatial transformation of the atlas into subject space (14). Prior to averaging, the voxels were excluded if they had a spectral linewidth larger than 12 Hz

FIGURE 1 | Exemplary MR spectra of each brain lobe and cerebellum obtained from a PD patient (female, 62 years). Estimated anatomic assignment of brain areas to brain lobes of our study. Frontal lobe: BA 4, 6, 8, 9, 10, 11, 12, 24, 25, 32, 33, 44, 45, 46, 47, head of caudate nucleus, accumbens, anterior part of putamen and pallidum, anterior cingulum. Parietal lobe: BA 1, 2, 3, 5, 7, 23, 31, 39, 40, thalamus, subthalamic nucleus, posterior cingulum. Temporal lobe: BA 13, 14, 15, 16, 20, 21, 22, 26, 27, 28, 29, 30, 34, 35, 36, 37, 38, 41, 42, 43, hippocampus, posterior parts of putamen and pallidum, caudatus tail, amygdala. Occipital lobe: BA 17, 18, 19. RFL, right frontal lobe; LFL, left frontal lobe; RPL, right parietal lobe; LPL, left parietal lobe; RTL, right temporal lobe; LTL left temporal lobe; ROL right occipital lobe; LOL left occipital lobe; Cbl, cerebellum; BA, Broadman Area.

or a CSF fraction larger than 30%. The application of these selection criteria resulted in excluding more basic voxels in frontal lobes (57% of the voxels within the structure), temporal lobes (55%), and cerebellum (54%) than in parietal (47%) and occipital lobes (33%), due to more filed distortion by neighbored structures containing bone and air or locations containing more CSF spaces. Finally, there were altogether 6,442 of 13,534 basic voxel spectra accounted for the spectral averaging in nine brain regions. The averaged spectra were subsequently analyzed with the FITT program included in MIDAS, in which a Lorentz-Gauss lineshape was used for spectral fitting. Mean regional concentrations of NAA, Cho, Cr, Glu, Gln, and mIns were determined as a ratio to a signal equivalent to that from 100% tissue water and presented as institutional units (i.u.) (21). Cramer-Rao lower bound (CRLB) of the spectral analysis was used as quality criteria for estimated metabolite values, i.e., only metabolites estimated with a CRLB <30% for Gln [a larger CRLB was selected for Gln to minimize possible bias related to its lower concentration (14)] and <20% for all other metabolites as often recommended for MRS analysis (http:// s-provencher.com/lcmodel.shtml) were considered for further analyses. The fractional tissue volumes of CSF (FVCSF) and total brain tissue (FVBT) in each brain region were derived by using multi-voxel analysis based on nine atlas-defined anatomical regions. Correction for CSF volume contribution was applied as Met' = Met/(1-FVCSF).

### Statistical Analysis

The results from the patient studies were compared with those of 20 age- and sex-matched healthy controls. The normality of the data was checked with Shapiro-Wilk test, where more than 80% of the data were normally distributed (p > 0.05). Paired t-test was used for comparison of measured metabolite concentrations, spectral linewidths, and the fractional volumes of CSF in each of nine brain regions between patients and healthy controls. Wilcoxon signed rank test was additionally used for the few non-normally distributed data, which revealed the results with

### TABLE 1 | Patient characteristic.


H&Y, Hoehn and Yahr stage; MDS-UPDS, Movement Disorders Society Unified Parkinson's Disease Rating Scale; DemTect test for cognitive assessment; LED, Levodopa-equivalence dosage.

the same significance levels as those derived by using paired t-test, i.e., this part of the data showed no significant changes in patients by both parametric and nonparametric estimation. For simplicity, only the results of paired t- tests are given. In addition, the lobar metabolite concentrations of the patients were pooled in two hemispheres according to dominantly affected body side (right side in 12 patients and left side in 7 patients, one patient was excluded from this analysis due to no obvious dominantly affected body side), i.e., the contralateral brain lobes as well as ipsilateral brain lobes in respect to the affected/more affected body side, and compared to those of the healthy controls by using paired t-test. In addition, a nonparametric Spearman's correlation test was used to estimate possible correlations between clinical MDS-UPDRS and pooled lobar metabolites concentrations in patients. Corrections for multiple comparisons or multiple correlation tests were performed by using the falsediscovery rate (FDR) method, with the desired false-discovery rate to 0.05. Those results with p-values not significant after a FDR correction were considered as showing a tendency (if p < 0.05) or a weak tendency (p < 0.75) of corresponding alterations in patients. Statistical analyses were performed with SPSS version 23 (SPSS IBM, New York, U.S.A.).

# RESULTS

## PD Patient Characteristics

The 20 early stage PD subjects were clinically diagnosed. None of the patients was suspected to suffer from atypical Parkinsonism or was cognitively impaired. All patients reported an obvious positive response on dopaminergic treatment and were examined in the best medical on state. As summarized in **Table 1**, the Hoehn and Yahr stage of the PD subjects was 2 or less with mean disease duration of 6 years. Twelve PD subjects (60%) showed right side dominant symptoms, seven (35%) left side, and one patient (5%) presented mainly non-motor symptoms with no dominantly affected body side. Of our 20 PD subjects 4 presented with an akinetic-rigid type (20%), 8 showed a tremor dominant type (40%) or an equivalence type (40%), respectively. All the early PD subjects revealed good cognitive functions measured by the DemTect with a mean score of

15.5 ± 2.9. The patients displayed a mean value of 7.7 points in MDS-UPDRS part I as non-motor aspects of daily living (SD 4.5, min 2, max 20), and 7.0 in MDS-UPDRS part II as motor aspects of daily living (SD 4.3, min 2, max 17). Motor deficits in the best medical on of our PD subjects in the MDS-UPDRS part III scored in mean 15.4 points (SD 7.5, min 5, and max 31). Using the MDS-UPDRS part IV as scale for motor complications we measured a mean score of 0.35 points (SD 1, min 0, max 4).

### Whole-Brain MR Spectroscopic Imaging

Example averaged MR spectra of each brain lobe and cerebellum obtained from a PD patient (female, 62 years) are shown in **Figure 1**. The lobar and cerebellar concentrations of NAA, Cho, tCr, Glu, Gln, and mIns measured in the PD patients and healthy controls are drawn as Box-Whiskerplots in **Figure 2**, which shows that several lobes exhibit clear differences between the metabolite values of the patients and the controls.

Results of paired t-tests for comparisons of metabolite concentrations, spectral linewidths, and the fractional volumes of CSF and total brain tissue between patients and controls are shown in **Tables 2**, **3**. Paired t-tests revealed that in patients NAA was decreased significantly in the right temporal lobe (−8.6%, p = 0.010), in the right parietal lobe (−6.7% and p = 0.007), and the right occipital lobe (−8%, p = 0.005), with a tendency to decrease in the left parietal lobe (p = 0.034) and with a weak tendency to decrease in both frontal lobes (p = 0.073 and 0.061 respectively); Cho did not show significant changes but revealed a weak tendency to decrease in right temporal lobe (p = 0.066); tCr was decreased significantly in the right temporal lobe (−7.6%, p = 0.004) and with a tendency to decrease in the right parietal lobe (−4.8%, p = 0.033); glutamate was decreased significantly in the right temporal lobe (−9.9%, p = 0.006) and the right occipital lobe (−10.2%, p = 0.001); glutamine showed a tendency to increase in the left temporal lobe (20.0%, p = 0.045); mIns did not show significant differences between patients and controls (**Table 2**). Moreover, in comparison to healthy controls

TABLE 2 | Comparison of lobar and cerebellar metabolite concentrations between patients and controls with paired t-tests.



<sup>a</sup>Definition of brain regions: left and right frontal lobe (LFL/RFL), left and right temporal lobe (LTL/RTL), left and right parietal lobe (LPL/RPL), left and right occipital lobe (LOL/ROL), and cerebellum (Cbl).

<sup>b</sup>Number of sampled subjects. Note that due to data quality controls (Crammer-Rao lower bond <30% for Gln, and <20% for all other metabolites) several subject pairs were not sampled for Gln analysis. Metabolites were determined as a ratio to a signal equivalent to that from 100% tissue water and presented as institutional unit (i.u.). SD = standard deviation. p-values < 0.05 were presented in bold.

\*\*Significant after correction for multiple comparisons by using false-discovery rate (FDR).

\*p < 0.05 but not significant after FDR correction.

TABLE 3 | Comparison of the spectral linewidths and the fractional volumes of cerebrospinal fluid and total brain tissue between patients and controls with paired t-tests.


<sup>a</sup>Definition of brain regions: left and right frontal lobe (LFL/RFL), left and right temporal lobe (LTL/RTL), left and right parietal lobe (LPL/RPL), left and right occipital lobe (LOL/ROL), and cerebellum (Cbl).

<sup>b</sup>Number of sampled subjects.

<sup>c</sup>FVCSF represents the fractional volumes of cerebrospinal fluid.

<sup>d</sup>FVTB represents the fractional volumes of total brain tissue.

\*\*Significant after correction for multiple comparisons by using false-discovery rate (FDR).

\*p < 0.05 but not significant after FDR correction.

the patients showed significantly broader spectral linewidths in both frontal lobes and the right occipital lobe, and with a trend to increase in the left occipital lobe. Slightly decreased FVCSF in frontal lobes and in left parietal lobe, and a slight increase of FVTB in left frontal lobe were also found in PD patients (**Table 3**). On the other hand, the cerebellum did not reveal significant differences between patients and controls concerning metabolite concentrations, spectral linewidth and FVCSF.

The results of paired t-tests between pooled lobar metabolite concentrations of the patients and the controls are shown in **Table 4**, and those of the correlation tests of MDS-UPDRS to pooled lobar NAA concentrations in patients in **Table 5**. In comparison to healthy controls the patients revealed different grades of alterations in pooled lobar metabolite concentrations in respect to contralateral or ipsilateral hemispheres corresponding to the prominently affected body side.

In the contralateral hemisphere, NAA decreased significantly or with a trend in all 4 brain lobes (−6.82% and p = 0.039 in FL, −8.00% and p = 0.023 in TL, −7.96% and p = 0.001 in PL, and −6.07% and p = 0.049 in OL); Cho showed a trend to decrease in one lobe (−8.47% and p = 0.030 in TL); tCr significantly decreased in two lobes (−8.52% and p = 0.006 in TL, and −6.40% and p = 0.009 in PL); glutamate decreased only in TL (−11.06% and p = 0.007); mIns and Gln did not show significant alterations (**Table 4**). Spearman's correlation test revealed significant correlations between MDS-UPDRS scores and lobar NAA concentrations in contralateral frontal lobe in patients, i.e., significant negative correlations of the NAA in contralateral frontal lobe to UPDRS2 as motor activities of daily living (R = −0.585, p = 0.008), and to total UPDRS (R = −0.679, p = 0.001) (**Table 5**), and a trend of negative correction to UPDRS4 as treatment complications (R = −0.458, p = 0.048).

In the ipsilateral hemisphere, only NAA and Glu revealed a weak tendency to decrease in occipital lobe (p = 0.057 for NAA and 0.059 for Glu) (**Table 4**), and no significant correlations of NAA to MDS -UPDRS was found (**Table 5**).

### DISCUSSION

In this study we assessed changes of brain lobar and cerebellar metabolites in early stage PD. Our major findings are the significant alterations of NAA contents in PD subjects in the whole brain in comparison to age-matched healthy controls: NAA was decreased or showed a tendency to decreased values in a majority of brain lobes (6 of 8 lobes). Of course, variations in tissue water content, which was used to calibrate the metabolite contents, could impact the measured concentration of metabolites in spectroscopy, however, this would affect not only NAA but all metabolites and since they were not all altered significantly we found no evidence for changes in tissue water content. After pooling brain lobar metabolites in hemispheres contralateral and ipsilateral to the dominantly affected body side significantly decreased NAA contents was seen in all four contralateral brain lobes, where frontal lobar NAA revealed negative correlations to clinical scores of UPDRS2, UPDRS4, and total UPDRS. Moreover, we found tCr decreased in two contralateral brain lobes, and Cho and glutamate decreased each in one contralateral brain lobe. In parallel, a significant broadening of spectral linewidth was found in both frontal and occipital lobes.

The findings of decreased NAA, Cho, tCr, and Glu, and no changes of mIns were qualitatively consistent with those previously reported, despite several methodological differences on targeted brain regions or tissue type. For example, studies in de novo PD subjects showed a reduction in NAA in the motor cortex (22) and putamen (23). Decreased NAA/tCr and Glu/tCr in PD subjects with psychosis were reported (24). Brain NAA, Cho, and tCr were also measured over brain lobes with long echo TABLE 4 | Paired t-test of lobar metabolite levels<sup>a</sup> between patients and controls measured in brain hemisphere contralateral or ipsilateral to affected/more affected body side as indicated.



<sup>a</sup>Measured in ratio to brain internal water.

<sup>b</sup>Number of patient-control pairs. One patient was excluded from the analysis because of bilateral predominate symptoms.

Due to data quality criteria (Crammer-Rao lower bond less than 30% for Gln, and less than 20% for all other metabolites) data of several patients were not sampled for Gln by the analysis. Metabolites were determined as a ratio to a signal equivalent to that from 100% tissue water and presented as institutional unit (i.u.).

\*\*Significant after correction for multiple comparisons by using false-discovery rate (FDR).

\*p < 0.05 but not significant after FDR correction.

time wbMRSI by Levin et al., who found decreased NAA/tCr and Cho/tCR in gray matter of temporal lobe, decreased NAA in right occipital lobe, and decreased NAA/tCr (25), anyhow, a detailed comparison to present study is difficult due to different patient selections and not separating tissue type between gray and white matter in the present. Since NAA is localized within neurons and involved in synaptic processes a decrease of brain NAA could be due to either a reduction in brain tissue volume or due to reduced neuronal function and metabolism (13). As no decrease of brain tissue volume in PD subjects was found the observed decreases of NAA most likely reflect reduced neuronal function and metabolism, which is consistent with the observations of decreased tCr and Glu, suggesting alterations in brain energy metabolism (tCr) and glutamatergic neuronal activity (Glu). The observed increase of glutamine in LTL could be a reactive response to reduced glutamate. Our findings of increased spectral linewidth in bilateral frontal and occipital lobes are most likely due to magnetic susceptibility-induced local magnetic field distortions and suggest pathological accumulation of brain iron in these brain regions, which has also been observed by other MRI measurements (26–28).

Present observation that brain lobar metabolite alterations were altered differently across the hemispheres contralateral and ipsilateral to the dominantly affected body side provides more insight into PD related brain metabolite changes. In the contralateral hemisphere, the main metabolic alterations were observed, which included significantly decreased NAA in all 4 brain lobes. In the ipsilateral hemisphere, smaller metabolic changes were seen, e.g., with PD subjects having lower mean NAA values in all lobes but not reaching significance. These observations indicate that the brain metabolite alterations are dominant in the hemisphere contralateral to more affected body side, reflecting PD-associated asymmetrical reduction of neuronal function and metabolism in early stage PD, which is consistent with previously reported lateralization findings of decreased NAA/tCr (29) and reduced dopamine uptake (30) in contralateral basal ganglia in early PD. Within the dominant hemisphere the distributions of the metabolic alterations varied among the brain lobes. The greatest changes occurred in the temporal lobe, with involvement of 3 metabolites (−8.00% for NAA, −8,47% for Cho, and −8.52% for tCr), the next occurred in the parietal lobe with involvement TABLE 5 | Correlations of MDS-UPDRS to brain NAA concentrations in respect to the most affected body side of the patients estimated by Spearman's correlation test<sup>a</sup> .


<sup>a</sup>Twelve patients with more affected right body side and 7 with more affected left body side.

\*\*Significant after correction for multiple comparisons by using false-discovery rate (FDR).

\*p < 0.05 but not significant after FDR correction. Bold values for results with p < 0.05.

of 2 metabolites (−7.96% for NAA and −6.40% for tCr), while frontal and occipital lobes revealed decreases of one metabolite (−6.82% for FL and −6.07% for OL for NAA), showing the inhomogeneity of brain structures involved in PD process that may relate to their contained substructures. Interestingly, corresponding to the fact that the basal ganglia, which has been reported to be involved in PD pathological processes in previous MRS studies (22, 23, 31–36), are located in the temporal and frontal lobar atlas regions used in this study, we found that most metabolic changes occurred in temporal lobe, while the frontal NAA changes correlated significantly to MDS-UPDRS score describing clinical symptoms. Previous studies have reported a significant correlation of NAA/tCr to clinical symptoms (11, 31, 33, 34). However, this study found that tCr was also altered in PD, which is consistent with reduced concentrations of phosphates in the striatum and midbrain that has been interpreted as early mitochondrial dysfunction in PD patients (37). Therefore, the use of metabolite ratios to tCr may underestimate PD related metabolic changes.

This study performed the clinical evaluation and wbMRSI of the patients in their best medical condition; therefore, the low MDS-UPDRS part III score may underestimate the degree of disability as the treatment reduces this score by at least 20– 30%. This may also have contributed to the lack of a significant correlation of the MDS-UPDRS III with spectroscopic findings. However, the MDS-UPDRS part II score for impairment of daily living reflects more selectively the impact of the disease (17, 24). The observed correlations of NAA in the frontal lobe, which includes important areas of the telencephalic dopaminergically innervated structures, with the MDS-UPDRS II, therefore, indicate the potential of wbMRSI for assessment of disability in Parkinson's disease (38, 39).

Treatment of PD is difficult and a lot of therapeutic agents are available (40). The impact of treatment in MRS studies has been rarely investigated. While some studies reported patients' on or off status during examination and imaging (11, 22, 25, 41– 47), for a large number of studies the medication status is unclear (9, 31, 32, 37, 48–56). A recent study found significant metabolic changes in PD subjects between medical on and off state (41). These authors found a significant reduction in NAA, tCr, and mIns in the clinical off which is reversed for NAA und tCr under acute L-DOPA challenge (200 mg intake) (41). Another early PD spectroscopic study found no metabolic changes in the putamen before and after apomorphine therapy in 5 PD patients (57). Lucetti et al. reported an increased Cho/Cr ratio after 6 months of treatment with the dopamine agonist pergolide in early de-novo PD patients (43). Taking this limited amount of data together, PD therapy seems to impact spectroscopic measurements, however, dopamine agonists might not similarly influence spectroscopic changes of metabolite profile. Hence, the impact of different PD therapeutics is not yet clear and might vary in different brain regions. Dopaminergic innervated brain regions seem to show more likely PD medication dependent changes in metabolite profile. The short echo time wbMRSI offers a potential method to study these neurometabolic effects in more detail for a variety of brain regions in the future. Importantly, the impact of off state, acute levodopa challenge, and chronic treatment should be addressed in future studies.

It remains unclear which brain region and metabolites could be used as a valid spectroscopic marker for PD or atypical Parkinsonism. For this purpose wbMRSI could be very useful, because it provides a way to measure brain metabolites not only in large brain scales but also in multiple specific brain areas simultaneously, thus many hypotheses could be tested in one set of data (13, 19). Clinically established imaging diagnostics are often unspecific and in PD the MRI is often normal (2–6). In atypical Parkinsonism structural changes can be seen in MRI scans, e.g., changes like midbrain atrophy, putaminal rim, hot cross bun sign and others, are indicative of different atypical Parkinson syndromes, however, in the absence of these signs the diagnosis is difficult (4, 6, 58). DAT-SCAN as marker for degeneration of dopaminergic transporters on nigrostriatal projection neurons of the ventral midbrain is very sensitive for early detection of PD (59). The discrimination of Parkinson syndromes by the DAT-SCAN is not possible, because all syndromes have the common pathological hallmark of degeneration of midbrain dopaminergic neurons. By FDG-PET imaging PD and atypical Parkinsonism could be discriminated by specific metabolic patterns (60). Unfortunately, FDG-PET is offlabel for the usage in the differential diagnosis of Parkinsonism. Furthermore, the methods of DAT-SCAN and FDG-PET require radiotracers. As an alternative WbMRSI offers a potential method for discrimination of PD and Parkinsonism without exposure to radioactive substances.

Present study focused on obtaining an overview of early PD-related metabolic alterations within the whole brain. Therefore, the cortical lobes and the cerebellum were selected as regions of interest in order to cover the whole brain. However, an accompanied limitation is that any regional metabolic inhomogeneity within the lobar or cerebellar structures was not accounted, especially the different contributions of the white matter, the cortical gray matter and basal ganglia were not separately evaluated. As an ongoing project PD related metabolic changes in multiple specific brain areas will be investigated in our further study. These results may then provide information related to specific brain functional networks and contribute to our understanding of the pathophysiological processes underlying PD.

Limitations of this study include the lack of correction of the results for age and for partial volume. However, the age effect was minimized by matching the patients and controls on a one-by-one basis appropriately in respect to age (and to gender). The partial volume effect was minimized by including only basic voxels containing <30% CSF for obtaining the integrated spectrum of each lobar brain region, and by correction for CSF volume contribution to measured metabolite values. Further validation of this study is also needed in a larger sample of patients together with a more comprehensive clinical phenotyping of PD subtypes (61–63). The identification of PD related metabolic changes in the white matter, cortical gray matter and basal ganglia may help to understand the metabolic processes during disease progression and spreading of neurodegenerative pathology in the brain (33, 64).

### REFERENCES

1. Adler CH, Beach TG, Hentz JG, Shill HA, Caviness JN, Driver-Dunckley E, et al. Low clinical diagnostic accuracy of early vs advanced Parkinson disease: clinicopathologic study. Neurology. (2014) 83:406–12. doi: 10.1212/WNL.0000000000000641

### CONCLUSION

This study has shown that NAA changed nearly ubiquitous in all brain lobes with different grades and with a clear lateralization contralateral to the major symptoms in early PD subjects. This finding suggests that NAA may be a promising spectroscopic marker for early diagnosis of PD (8), which is also favored by the observations of significant negative correlations between frontal NAA levels and the clinical UPDRS scores. Future studies with larger cohorts of patients with different stages of PD are needed to verify these results.

In conclusion, this study has demonstrated that even in earlystage PD brain metabolic alterations are evident and involved in all brain lobar areas of the cerebral hemisphere contralateral to the dominant side of disability. This result indicates that PD affects not only brain local regions by dopaminergic denervation, but also the brain network within the hemisphere. The novel wbMRSI-detectable brain metabolic alterations in PD may serve as promising biomarkers for early PD diagnosis, differential diagnosis of Parkinsonism (32) and with emerging diseasemodifying drugs also for treatment monitoring (65, 66).

## ETHICS STATEMENT

Human subject studies were carried out with approval from the local Ethics Committee of Hannover Medical School (No. 6167-2016) and all subjects gave written informed consent.

### AUTHOR CONTRIBUTIONS

XD designed the study. MK, DD, and FW were responsible for recruitment. MK and PB performed the clinical characterization of the patients. PB and PN performed the MRIs. XD performed the statistical analysis. MK, FW, DD, MD, AM, SS, HL, and XD interpreted the data. MK, FW, and XD wrote the manuscript. PB, DD, MD, AM, SS, and HL coedited the manuscript. All authors had access to the data generated in the study including the statistical analysis and agree to submit the paper for publication.

### FUNDING

This research was supported by the DFG Grand to XD and the NIH Grand for AM. The sponsors had no influence on the carrying out of the study or the writing of the manuscript.

### ACKNOWLEDGMENTS

We would like to thank our research volunteers.


magnetic resonance imaging sequences: a specific marker of multiple system atrophy? Arch Neurol. (1999) 56:225–8.


MR spectroscopic imaging (MRSI). Magn Reson Med. (2009) 61:548–59. doi: 10.1002/mrm.21875


cohort of advanced Parkinson's disease patients. Front Neurol. (2018) 9:120. doi: 10.3389/fneur.2018.00120


by proton magnetic resonance spectroscopy. Eur J Neurol. (2000) 7:661–5. doi: 10.1046/j.1468-1331.2000.00111.x


**Conflict of Interest Statement:** The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

Copyright © 2019 Klietz, Bronzlik, Nösel, Wegner, Dressler, Dadak, Maudsley, Sheriff, Lanfermann and Ding. This is an open-access article distributed under the terms of the Creative Commons Attribution License (CC BY). The use, distribution or reproduction in other forums is permitted, provided the original author(s) and the copyright owner(s) are credited and that the original publication in this journal is cited, in accordance with accepted academic practice. No use, distribution or reproduction is permitted which does not comply with these terms.

# Deep Learning With EEG Spectrograms in Rapid Eye Movement Behavior Disorder

Giulio Ruffini 1,2 \*, David Ibañez <sup>2</sup> , Marta Castellano<sup>2</sup> , Laura Dubreuil-Vall <sup>1</sup> , Aureli Soria-Frisch<sup>2</sup> , Ron Postuma<sup>3</sup> , Jean-François Gagnon<sup>4</sup> and Jacques Montplaisir <sup>5</sup>

<sup>1</sup> Neuroelectrics Corporation, Cambridge, MA, United States, <sup>2</sup> Applied Neuroscience, Starlab Barcelona, Barcelona, Spain, <sup>3</sup> Department of Neurology, Montreal General Hospital, Montreal, QC, Canada, <sup>4</sup> Centre for Advanced Research in Sleep Medicine, Hôpital du Sacré-Cœur de Montréal, Montreal, QC, Canada, <sup>5</sup> Department of Psychiatry, Université de Montréal, Montreal, QC, Canada

### Edited by:

Jue Zhang, Peking University, China

### Reviewed by:

Marco Onofrj, Università degli Studi 'G. d'Annunzio' Chieti - Pescara, Italy Maria Salsone, Italian National Research Council (CNR), Italy

> \*Correspondence: Giulio Ruffini

giulio.ruffini@neuroelectrics.com

### Specialty section:

This article was submitted to Applied Neuroimaging, a section of the journal Frontiers in Neurology

Received: 31 March 2019 Accepted: 12 July 2019 Published: 30 July 2019

### Citation:

Ruffini G, Ibañez D, Castellano M, Dubreuil-Vall L, Soria-Frisch A, Postuma R, Gagnon J-F and Montplaisir J (2019) Deep Learning With EEG Spectrograms in Rapid Eye Movement Behavior Disorder. Front. Neurol. 10:806. doi: 10.3389/fneur.2019.00806 REM Behavior Disorder (RBD) is now recognized as the prodromal stage of αsynucleinopathies such as Parkinson's disease (PD). In this paper, we describe deep learning models for diagnosis/prognosis derived from a few minutes of eyes-closed resting electroencephalography data (EEG) collected at baseline from idiopathic RBD patients (n = 121) and healthy controls (HC, n = 91). A few years after the EEG acquisition (4±2 years), a subset of the RBD patients were eventually diagnosed with either PD (n = 14) or Dementia with Lewy bodies (DLB, n = 13), while the rest remained idiopathic RBD. We describe first a simple deep convolutional neural network (DCNN) with a five-layer architecture combining filtering and pooling, which we train using stacked multi-channel EEG spectrograms from idiopathic patients and healthy controls. We treat the data as in audio or image classification problems where deep networks have proven successful by exploiting invariances and compositional features in the data. For comparison, we study a simple deep recurrent neural network (RNN) model using three stacked Long Short Term Memory network (LSTM) cells or gated-recurrent unit (GRU) cells—with very similar results. The performance of these networks typically reaches 80% (±1%) classification accuracy in the balanced HC vs. PD-conversion classification problem. In particular, using data from the best single EEG channel, we obtain an area under the curve (AUC) of 87% (±1%)—while avoiding spectral feature selection. The trained classifier can also be used to generate synthetic spectrograms using the DeepDream algorithm to study what time-frequency features are relevant for classification. We find these to be bursts in the theta band together with a decrease of bursting in the alpha band in future RBD converters (i.e., converting to PD or DLB in the follow up) relative to HCs. From this first study, we conclude that deep networks may provide a useful tool for the analysis of EEG dynamics even from relatively small datasets, offering physiological insights and enabling the identification of clinically relevant biomarkers.

Keywords: PD, RBD, Deep learning, EEG, time-frequency analysis

# 1. INTRODUCTION

RBD is a parasomnia characterized by intense dreams with during REM sleep without muscle atonia (1), i.e., with vocalizations and body movements. Idiopathic RBD occurs in the absence of any neurological disease or other identified cause, is male-predominant and its clinical course is generally chronic progressive (2). Several longitudinal studies conducted in sleep centers have shown that most patients diagnosed with the idiopathic form of RBD will eventually be diagnosed with a neurological disorder such as Parkinson disease (PD) or dementia with Lewy bodies (DLB) (1–4). In essence, idiopathic RBD has been suggested as a prodromal stage of αsynucleinopathies [PD, DLB, and less frequently multiple system atrophy (MSA) (1, 4)].

RBD has an estimated prevalence of 15–60% in PD and has been proposed to define a subtype of PD with relatively poor prognosis, reflecting a brainstem-dominant route of pathology progression (see (5) and references therein) with a higher risk for dementia or hallucinations. PD with RBD is characterized by more profound and extensive pathology—not limited to the brainstem—, with higher synuclein deposition in both cortical and sub-cortical regions.

Electroencephalographic (EEG) and magnetoencepha lographic (MEG) signals contain rich information associated with functional processes in the brain. To a large extent, progress in their analysis has been driven by the study of spectral features in electrode space, which has indeed proven useful to study the human brain in both health and disease. For example, the "slowing down" of EEG is known to characterize neurodegenerative diseases (6–8). It is worth mentioning that the selection of disease characterizing features from spectral analysis is mostly done after an extensive search in the frequency-channel domain.

However, neuronal activity exhibits non-linear dynamics and non-stationarity across temporal scales that cannot be studied properly using classical approaches. Tools capable of capturing the rich spatiotemporal hierarchical structures hidden in these signals are needed. In Ruffini et al. (8), for example, algorithmic complexity metrics of EEG spectrograms were used to derive information from the dynamics of EEG signals in RBD patients, with good results, indicating that such metrics may be useful per se for classification or scoring. However, ideally we would like to use methods where the relevant features are found directly by the algorithms.

Deep learning algorithms are designed for the task of exploiting compositional structure in data (9). In past work, for example, deep feed-forward autoencoders have been used for the analysis of EEG data to address the issue of feature selection, with promising results (10). Interestingly, deep learning techniques, in particular, and artificial neural networks in general are themselves bio-inspired by the brain—the same biological system generating the electric signals we aim to decode. This suggests they may be well suited for the task.

Deep recurrent neural networks (RNNs), are known to be potentially Turing complete [see, e.g., (11) for a review], but general RNN architectures are notoriously difficult to train (12). In this regard, it is worth mentioning that "reservoir" based RNN training approaches are evolving (13). In earlier work, a particular class of RNNs called Echo State Networks (ESNs) that combine the power of RNNs for classification of temporal patterns and ease of training (14) was used with good results with the problem at hand. The main idea behind ESNs and other "reservoir computation" approaches is to use semi-randomly connected, large, fixed recurrent neural networks where each node/neuron in the reservoir is activated in a non-linear fashion. The interior nodes with random weights constitute what is called the "dynamic reservoir" of the network. The dynamics of the reservoir provides a feature representation map of the input signals into a much larger dimensional space (in a sense much like a kernel method). Using such an ESN, an accuracy of 85% in a binary, class-balanced classification problem (healthy controls vs. PD patients) was obtained using a relatively small dataset in Ruffini et al. (14). The main limitations of this approach, in our view, are the computational cost of developing the reservoir dynamics of large random networks and the associated need for feature selection (e.g., which subset of frequency bands and channels to use as inputs to simplify the computational burden).

In this paper we use a similar but simpler strategy as the one presented in Vilamala et al. (15), using Deep Convolutional Neural Networks with EEG signals, i.e., multi-channel time series. In comparison to Vilamala et al. (15), we reduce the number of hidden layers from 16 to 4, use a simpler approach for the generation of spectrograms, and do not rely on transfer learning from a network trained on a visual recognition task. Indeed, we believe such a pre-training would initialize the filtering weights to detect object-like features not present in spectrograms. The proposed method outperforms several shallow methods used for comparison as presented in the results section.

Lastly, we employ deep-learning visualization techniques for the interpretation of results. Once a network has been trained, one would like to understand what are the key features it is picking up from the data for classification. We show below how this can be done in the context of EEG spectrogram classification, and how it can be helpful in identifying physiologically meaningful features that would be hard to select by hand. This is also very important for the clinical translation of such techniques, since black-box approaches have been extensively criticized.

### 2. MATERIALS AND METHODS

### 2.1. Deep Learning in the Spectrogram Representation

Our goal here will be to train a network to classify subjects from the EEG spectrograms recorded at baseline in binary problems, with classification labels such as HC (healthy control), PD (idiopathic RBD who will later convert to PD), etc.

Here we explore first a deep learning approach inspired by recent successes in image classification using deep convolutional neural networks designed to exploit invariances and capture compositional features in the data [see e.g., (9, 11, 12)]. These systems have been largely developed to deal with image data, i.e., 2D arrays, possibly from different channels, or audio data [as in van den Oord et al. (16)], and, more recently, with EEG data as well (15, 17). Thus, inputs to such networks are data cubes (multichannel stacked images). In the same vein, we aimed to work here with the spectrograms of EEG channel data, i.e., 2D time-frequency maps. Such representations represent spectral dynamics as essentially images with the equivalent of image depth provided by multiple available EEG channels (or, e.g., current source density maps or cortically mapped quantities from different spatial locations). Using such representation, we avoid the need to select frequency bands or channels in the process of feature selection. This approach essentially treats EEG channel data as an audio file, and our approach mimics similar uses of deep networks in that domain.

RNNs can also be used to classify images, e.g., using image pixel rows as time series. This is particularly appropriate in the case of the data in this study, given the good performance we obtained using ESNs on temporal spectral data Ruffini et al. (14). We study here also the use of stacked architectures of long-short term memory network (LSTM) or gated-recurrent unit (GRU) cells, which have shown good representational power and can be trained using backpropagation (12, 18, 19).

Our general assumption is that some relevant aspects in EEG data from our datasets are contained in compositional features embedded in the time-frequency representation. This assumption is not unique to our particular classification domain, but should hold of EEG in general. In particular, we expect that deep networks may be able to efficiently learn to identify features in the time-frequency domain associated to bursting events across frequency bands that may help separate classes, as in "bump analysis" (20). Bursting events are hypothesized to be representative of transient synchrony of neural populations, which are known to be affected in neurodegenerative diseases such as Parkinson's or Alzheimer's disease (21).

Finally, we note that in this study we have made no attempt to fully-optimize the network architecture. In particular, no fine-tuning of hyper-parameters has been carried out using a validation set approach, something we leave for future work with larger datasets. Our aim has been to implement a proof of concept of the idea that deep learning approaches can provide value for classification and analysis of time-frequency representations of EEG data—while possibly providing new physiological insights.

### 2.2. Study Subjects

Idiopathic RBD patients (called henceforth RBD for data analysis class labeling) and healthy controls were recruited at the Center for Advanced Research in Sleep Medicine of the Hôpital du Sacrè-Cœur de Montréal as part of another study and kindly provided for this work. The protocol was approved by the Hôpital du Sacré-Cœur de Montréal Ethics Committee, and all participants gave their written informed consent to participate. For more details on the protocol and on the patient population statistics (age and gender distribution, follow up time, etc.), see Rodrigues-Brazéte et al. (7) and Ruffini et al. (8).

The dataset includes a total of 121 patients diagnosed with idiopathic RBD (of which 118 passed the first quality tests) and 85 healthy controls (of which only 74 provided sufficient quality data) without sleep complaints and in which RBD was excluded. EEG data was collected in every patient at baseline, e.g., when patients were still RBD. After 1–10 years of clinical follow-up 14 RBD patients converted to PD, 13 to DLB, while the rest remained idiopathic RBD (see **Figure 1**).

In addition to EEG recording at baseline (further described below) participants also underwent a complete neurological examination by a neurologist specialized in movement disorders and a cognitive assessment by a neuropsychologist. The only data used from the follow-up evaluation, which was conducted on average 10 years after baseline, was the updated diagnosis change, if any, from RBD into PD or DLB, or the confirmation of the RBD diagnosis. These data elements have been used here as ground truth in the DCNN training and in the performance evaluation on the test set as set up in the cross validation procedure.

RBD was diagnosed based on AASM Version II (https://aasm.org/aasm-updates-scoring-manual-version-

2-2-with-new-option-for-monitoring-respiratory-effort-

during-hsat/). This included a history of dream enactment behaviors and a subsequent assessment of overnight polysomnography (PSG) evaluation including video recording and EMG evaluation (22). EEG was acquired at the end of the PSG recording session in awake state.

PD was diagnosed following the Movement Disorder Society Clinical Diagnostic Criteria for Parkinson's disease (PD) (23). In early recordings, the criteria was the standard at that time based on Hughes et al. (24). DLB diagnosis was based on standard procedures described in McKeith et al. (25). Some subjects may have gone through neuroimaging (MRI, as no DAT Scan was available in Canada) for confirmation or differential diagnosis, but not in a systematic way in the overall PD/DLB population.

No healthy controls reported abnormal motor activity during sleep or showed cognitive impairment on neuropsychological testing. Only a subset of healthy controls was followed up. In general, patients were recruited within a year of RBD diagnosis. However, we note as a limitation that the cohort was recruited during a period of 15 years, which may have affected the recruiting conditions.

### 2.3. EEG Dataset

All RBD patients with a full EEG montage for resting-state EEG recording at baseline and with at least one follow-up examination (without EEG) after the baseline visit were included in the study. The first valid EEG for each patient enrolled in the study was considered baseline.

As in related work (7, 8, 14), the raw data in this study consisted of resting-state EEG collected from awake subjects using 14 scalp electrodes. The recording protocol consisted of conditions with periods of with eyes open of variable duration (∼2.5 min) followed by periods with eyes closed in which patients were not asked to perform any particular task. EEG signals were digitized with 16-bit resolution at a sampling rate of 256 S/s. The amplification device bandpass filtered the EEG data between 0.3 and 100 Hz with a notch filter at 60 Hz

to minimize line power noise. All recordings were referenced to linked ears.

# 2.4. Preprocessing and Generation of Spectrograms

To generate spectrograms (here called frames), EEG data from each channel was processed using Fourier analysis (FFT) after detrending blocks of 1 s with a Hann window (FFT resolution is 2 Hz) (see **Figure 1**). Twenty second 14 channel artifact-free epochs were collected for each subject, using a sliding window of 1 s. FFT amplitude bins in the band 4–44 Hz were used. The resulting data frames are thus multidimensional arrays of the form [channels (14)] x [FFTbins (21)] x [Epochs (20)]. To avoid biases, the number of frames per subject was fixed as a tradeoff between data per subject and number of subjects included, to 148, representing about 2.5 min of data. We selected a minimal suitable number of frames per subject so that each subject provided the same number of frames. For training, datasets were balanced for subjects by random replication of subjects in the class with fewer subjects. For testing, we used a leave-pair-out strategy [LPO, see (26)], with one subject from each class. Thus, both the training and test sets were balanced both in terms of subjects and frames per class. Finally, the data was centered and normalized to unit variance for each frequency and channel.

### 2.5. Network Architectures

We have implemented three architectures: DCNN and stacked RNN, as we now describe, plus a shallow architecture for comparison (see **Figure 2**).

### 2.5.1. DCNN Architecture

The network (which we call SpectNet), implemented in Tensorflow (27), is a relatively simple four hidden-layer convolutional net with pooling (see **Figure 2**). Dropout has been used as the only regularization. All EEG channels may be used in the input cube. The design philosophy has been to enable the network to find local features first and create larger views of data with some temporal (but not frequency) shift invariance via max-pooling.

The network has been trained using a cross-entropy loss function to classify frames (not subjects). It has been evaluated both on frames and, more importantly, on subjects by averaging subject frame scores and choosing the maximal probability class, i.e., using a 50% threshold. For development purposes, we have also tested the performance of this DCNN on a synthetic dataset consisting of Gaussian radial functions randomly placed on the spectrogram time axis but with variable stability in frequency, width and amplitude (i.e, by adding some jitter top these parameters). Frame classification accuracy was high and relatively robust to jitter (∼95–100%, depending on parameters), indicating that the network was capable of learning to detect burst-like features with time-translational invariance and frequency specificity.

### 2.5.2. RNN Architecture

The architectures for the RNNs consisted of stacked LSTM (12, 18) or GRU cells (19). The architecture we describe here consists of three stacked cells, where each cell uses as input the outputs of the previous one. Each cell used 32 hidden units, and dropout

was used to regularize it. The performance of LSTM and GRU variants was very similar.

### 3. RESULTS

### 3.1. Classification Performance Assessment

used for comparison. (C) Deep RNN using LSTM or GRU cells.

Our goal is to classify subjects (e.g., HC or PD converter labels) rather than frames. The performance of the networks has been evaluated in the balanced dataset using two metrics in a leave-pair out cross-validation framework—where the data from a subject in each class is left out for validation (LPO). First, using the accuracy metric (probability of good a classification), and second, by using the area under the curve (AUC) using the Wilcoxon-Mann-Whitney statistic (26). To map out the classification performance of the DCNN for different parameter sets, we have implemented a set of algorithms based on the Tensorflow package (27) as described in the following pseudocode:

TABLE 1 | Performance in different problems using a single EEG channel (P4, see Figure 4).


From left to right: architecture used and problem addressed (groups); Number of subjects in training and test sets per group (always balanced); train and test average performance on frames; test accuracy and LPO cross-validation area-under-the-curve metric (AUC) (26). Results to ±1%.

### REPEAT N times (experiments): 1- Choose (random, balanced) training and test subject sets (leave-pair-out)

2- Augment smaller set by random

```
replication of subjects
 3- Optimize the NN using stochastic
  gradient descent with frames as inputs
 4- Evaluate per-frame performance on
  training and test set
 5- Evaluate per-subject performance
   averaging frame outputs
END
Compute mean and standard deviation of
 performances over the N experiments
```
For each frame, the classifier outputs the probability of the frame belonging to each class [using softmax, see, e.g., (12)] and, as explained above, after averaging over frames per subject we obtain the probability of the subject belonging to each class. This provides an interesting score in itself. Classification is carried out by choosing the class with maximal probability.

The results from classification are shown in **Table 1** for the HC vs. PD problem and the HC+RBD vs. PD+DLB problem, which includes more data. Sample results for the RNN architecture (which are very similar to DCNN results) are provided in **Figure 3**. For comparison, using a shallow architecture neural network resulted in about 10% less ACC or AUC (in line with our results using support vector machine (SVM) classifiers (Soria-Frisch et al., in preparation), which required feature selection). On the other hand, in Ruffini et al. (14), a peak accuracy of 85% was reached in the balanced problem of HC vs PD, although this required appropriate feature selection (a selection of channels and bands), and in Ruffini et al. (8) similarly high AUC performance was reached using global (in channel and frequency space) complexity metrics.

**Figure 4** provides the performance in the HC vs. PD problem using different EEG channels (statistics computed using a smaller number of folds).

### 3.2. Interpretation

Once a DCNN has been trained, it can be used to explore which inputs optimally excite network nodes, including the output nodes that provide the classification (29). The algorithm for doing the latter consists essentially in maximizing a particular class score using gradient descent, starting from, e.g., a random noise image. An example of the resulting images using the trained DCNN above can be seen in **Figure 5**, where image corresponds to the input that maximizes each class output, e.g., HC vs. PD. This is a particularly interesting technique in our diagnosis/prognosis problem and provides new insights on the class-specific features in EEG of each class. In the case of a HC vs. PD trained network, we can see alterations in the alpha and theta spectral bands, appearing differentially in the form of bursts in each class. In the difference spectrograms we can observe the disappearance of alpha bursts in exchange with bursting at lower frequencies. This findings are consistent with others relating to alterations and slowing of EEG (6– 8, 28, Soria-Frisch et al., in preparation), and in particular of

FIGURE 4 | Sample images produced by maximizing network outputs for a given class. (Left) From a network was trained using P4 electrode channel data on the problem of HC vs. PD. The main features are the presence of 10 Hz bursts in the image maximizing HC classification (Top) compared to more persistent 6 Hz power in the pathological spectrogram (Middle). The difference of the two is displayed at the bottom. (Right) Network was trained using P4 electrode data on the problem of HC vs. PD+DLB (i.e., HC vs. RBDs that will develop an α-synucleinopathy or SNP). The main features are the presence of 10 Hz bursts in the HC class maximizing image (Top) compared to more persistent 6–8 Hz power bursting in the pathological spectrogram (Middle).

longitudinal alpha frequency and theta frequency band relative power increases in PD with dementia (30). However, they point out in more detail what the network has learned as feature to separate the classes: bursting in the observed bands. This adds a dimension (time) to the usually identified features (power, slowing).

# 4. DISCUSSION

Our results using deep networks are complementary to earlier work using machine learning to analyze this type of data using SVMs and ESNs. However, we deem the use of deep learning methods to be particularly interesting for various reasons. First, they largely mitigate the need for feature selection (in this case the choice of spectral bands and channels). Here we preprocessed the EEG data to obtain spectrograms as a way to simplify the learning task given the limitations in data availability (given enough data, it would seem natural to work with raw or minimally cleaned up multichannel EEG data). Secondly, the employed method represents an improvement over related prior efforts, increasing performance by about 5–10% in AUC (28, Soria-Frisch et al., in preparation).

The obtained results and especially the ones derived from the use of feature visualization are in agreement with the findings of slowing of EEG in PD with respect to HC and RBD patients as observed in previous studies, i.e., power increase in lower frequency bands and decrease in higher ones. More specifically, the shifting of bursting events in the alpha band to lower frequencies is especially interesting and may suggest potential mechanistic explanations regarding the effects of disease on the alpha band underlying circuitry. This underscores the fact that the DCNN can pick up relevant discriminative features without explicitly being tuned to do so, which is not the case for those previous studies with hand-picked features.

The performance of the network was higher with the task of discriminating HC and converters than RBD non-converters and converters, which is expected and probably reflecting different time courses of disease in subjects. This reflects a limitation in our study, namely, that RBD diagnosis and recruitment may have happened along different timepoints for each subject, creating a confound in the analysis.

We note that another limitation in the used dataset is the presence of healthy controls without follow up, which may be a confound for the network—worsening its performance, as some controls may actually be prodromal PD, for example [around 2.2% (31)]. We hope to remedy in the future this by enriching our database with improved diagnosis and follow up methodologies. In addition to dataset quality improvements, future steps include the exploration of this approach with larger datasets as well as a more systematic study of network architecture and regularization schemes. This includes the use of deeper architectures, improved data augmentation methods, alternative data segmentation and normalization schemes. With regard to data preprocessing, we should consider improved spectral estimation using more advanced techniques such as state-space estimation and multitapering—as in Kim et al. (32), and working with cortically or scalp-mapped EEG data prior creation of spectrograms.

Although here, as in Vilamala et al. (15), we worked with time-frequency pre-processed data, the field will undoubtedly steer toward working with raw data in the future when larger datasets become available—as suggested in Schirrmeister et al. (33). Working with time-frequency power representations is definitely a limitation, given current view indicating that neural processing involves both amplitude and phase of signals, e.g., as in communication through coherence or, more generally, oscillation-based communication (34).

In closing, we note that the techniques used in this pilot study can be extended to other EEG related problems, such as braincomputer interfaces, sleep scoring, detection of epileptiform activity or EEG data pre-processing, where the advantages of deep learning approaches may prove useful as well.

# DATA AVAILABILITY

The data analyzed in this study was obtained from Jacques Montplaisir's team at the Center for Advanced Research in Sleep Medicine affiliated to the University of Montréal, Hopital du Sacre-Coeur, Montréal during a previous study. The dataset used for deep learning in this paper (EEG spectrograms) is available upon request to giulio.ruffini@neuroelectrics.com.

## ETHICS STATEMENT

Idiopathic RBD patients (called henceforth RBD for data analysis labeling) and healthy controls were recruited at the Center for Advanced Research in Sleep Medicine of the Hôpital du Sacré-Cur de Montréal as part of another study and kindly provided for this work. All patients with a full EEG montage for resting-state EEG recording at baseline and with at least one follow-up examination after the baseline visit were included in the study. The first valid EEG for each patient enrolled in the study was considered baseline. Participants also underwent a complete neurological examination by a neurologist specialized in movement disorders and a cognitive assessment by a neuropsychologist. No controls reported abnormal motor activity during sleep or showed cognitive impairment on neuropsychological testing. The protocol was approved by the hospital's ethics committee, and all participants gave their written informed consent to participate. For more details on the protocol and on the patient population statistics (age and gender 1distribution, follow up time, etc.), see Rodrigues-Brazéte et al. (7) and Ruffini et al. (8).

# AUTHOR CONTRIBUTIONS

DI and MC preprocessed the EEG data to produce artifact free spectrograms. LD-V contributed to the implementation of the neural networks and revised the manuscript. AS-F contributed to the design of the study, interpretation of the results, and revised the manuscript. J-FG and JM designed the data collection study, collected the EEG data, and revised the manuscript. RP contributed to the study design, data collection, and revision of the manuscript. GR designed the analysis pipeline, implemented the neural networks, analyzed the results, and wrote the manuscript.

### FUNDING

This work has been partially funded by The Michael J. Fox Foundation for Parkinson's Research under Rapid Response Innovation Awards 2013. Part of this research was supported by grants from the Canadian Institutes of Health Research (CIHR) and W. Garfield Weston Foundation to J-FG, JM, and RP.

# ACKNOWLEDGMENTS

This manuscript has been released as a Pre-Print at bioRxiv (35). J-FG holds a Canada Research Chair in Cognitive Decline in Pathological Aging. JM holds a Canada Research Chair in Sleep Medicine.

### REFERENCES


**Conflict of Interest Statement:** Neuroelectrics and Starlab authors disclose commercial interests in the development of EEG derived biomarkers. GR is a co-founder of Starlab and Neuroelectrics.

The remaining authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

Copyright © 2019 Ruffini, Ibañez, Castellano, Dubreuil-Vall, Soria-Frisch, Postuma, Gagnon and Montplaisir. This is an open-access article distributed under the terms of the Creative Commons Attribution License (CC BY). The use, distribution or reproduction in other forums is permitted, provided the original author(s) and the copyright owner(s) are credited and that the original publication in this journal is cited, in accordance with accepted academic practice. No use, distribution or reproduction is permitted which does not comply with these terms.

# Lenticulostriate Arteries and Basal Ganglia Changes in Cerebral Autosomal Dominant Arteriopathy With Subcortical Infarcts and Leukoencephalopathy, a High-Field MRI Study

### Edited by:

*Hongyu An, Washington University in St. Louis, United States*

### Reviewed by:

*Bruno J. Weder, University of Bern, Switzerland Jay Chol Choi, Jeju National University, South Korea*

### \*Correspondence:

*Zihao Zhang zhzhang@ibp.ac.cn Yun Yuan yuanyun2002@126.com*

*†These authors have contributed equally to this work*

### Specialty section:

*This article was submitted to Applied Neuroimaging, a section of the journal Frontiers in Neurology*

Received: *26 February 2019* Accepted: *26 July 2019* Published: *09 August 2019*

### Citation:

*Ling C, Fang X, Kong Q, Sun Y, Wang B, Zhuo Y, An J, Zhang W, Wang Z, Zhang Z and Yuan Y (2019) Lenticulostriate Arteries and Basal Ganglia Changes in Cerebral Autosomal Dominant Arteriopathy With Subcortical Infarcts and Leukoencephalopathy, a High-Field MRI Study. Front. Neurol. 10:870. doi: 10.3389/fneur.2019.00870* Chen Ling1†, Xiaojing Fang1,2†, Qingle Kong3,4,5, Yunchuang Sun<sup>1</sup> , Bo Wang3,4,5 , Yan Zhuo3,4,5, Jing An<sup>6</sup> , Wei Zhang<sup>1</sup> , Zhaoxia Wang<sup>1</sup> , Zihao Zhang3,4,5 \* and Yun Yuan<sup>1</sup> \*

*<sup>1</sup> Department of Neurology, Peking University First Hospital, Beijing, China, <sup>2</sup> Department of Neurology, Peking University International Hospital, Beijing, China, <sup>3</sup> State Key Laboratory of Brain and Cognitive Science, Beijing MR Center for Brain Research, Institute of Biophysics, Chinese Academy of Sciences, Beijing, China, <sup>4</sup> CAS Center for Excellence in Brain Science and Intelligence Technology, Beijing, China, <sup>5</sup> University of Chinese Academy of Sciences, Beijing, China, <sup>6</sup> Siemens Shenzhen Magnetic Resonance Ltd., Shenzhen, China*

Background and Purpose: Cerebral autosomal dominant arteriopathy with subcortical infarcts and leukoencephalopathy (CADASIL) mainly affects the cerebral small arteries. We aimed to analyze changes in the lenticulostriate arteries (LSAs) and the basal ganglia in patients with CADASIL using high-field magnetic resonance imaging (7.0-T MRI).

Methods: We examined 46 patients with CADASIL and 46 sex- and age-matched healthy individuals using 7.0-T MRI. The number and length of the LSAs, and the proportion of discontinuous LSAs were compared between the two groups. The Mini-Mental State Examination score, the modified Rankin Scale, the Barthel Index, and the MRI lesion load of the basal ganglia were also examined in patients with CADASIL. We analyzed the association between LSA measurements and the basal ganglia lesion load, as well as the association between LSA measurements and clinical phenotypes in this patient group.

Results: We observed a decrease in the number of LSA branches (*t* = −2.591, *P* = 0.011), and an increase in the proportion of discontinuous LSAs (*z* = −1.991, *P* = 0.047) in patients with CADASIL when compared with healthy controls. However, there was no significant difference in the total length of LSAs between CADASIL patients and healthy individuals (*t* = −0.412, *P* = 0.682). There was a positive association between the number of LSA branches and the Mini-Mental State Examination scores of CADASIL patients after adjusting for age and educational level (β = 0.438; 95% CI: 0.093, 0.782; *P* = 0.014). However, there was no association between LSA measurements and the basal ganglia lesion load among CADASIL patients.

**103**

Conclusions: 7.0-T MRI provides a promising and non-invasive method for the study of small artery damage in CADASIL. The abnormalities of small arteries may be related to some clinical symptoms of CADASIL patients such as cognitive impairment. The lack of association between LSA measurements and the basal ganglia lesion load among the patients suggests that changes in the basal ganglia due to CADASIL are caused by mechanisms other than anatomic narrowing of the vessel lumen.

Keywords: cerebral autosomal dominant arteriopathy with subcortical infarcts and leukoencephalopathy (CADASIL), 7.0-Tesla magnetic resonance imaging (7.0-T MRI), time-of-flight-magnetic resonance angiography (TOF-MRA), lenticulostriate arteries, basal ganglia

### INTRODUCTION

Cerebral autosomal dominant arteriopathy with subcortical infarcts and leukoencephalopathy (CADASIL) is an inherited small vessel disease caused by mutations in the NOTCH3 gene (1, 2). The main clinical features of CADASIL include recurrent transient ischemic attack (TIA) and ischemic stroke, migraine with or without aura, progressive cognitive decline, and mood disturbances (3–5). Granular osmiophilic material (GOM) deposits in the basement membrane of vascular smooth muscle cells (VSMCs) represent the pathological hallmark of CADASIL (1, 6). Magnetic resonance imaging (MRI) also plays a crucial role in the diagnosis and clinical evaluation of CADASIL. Diffuse white matter hyperintensities (WMHs), multiple lacunar infarctions (LIs), and cerebral microbleeds (CMBs) are the typical MRI abnormalities in patients with CADASIL (7, 8).

Small arteries, especially the cerebral small arteries are mainly affected in CADASIL. The lenticulostriate arteries (LSAs) are the major cerebral small arteries supplying blood to the basal ganglia, a region of the brain that is particularly susceptible in CADASIL (9, 10). Ultrastructural analysis is the commonly used method for studying changes of small cerebral arteries in CADASIL. Investigations using such methods have revealed that small cerebral arteries in patients with CADASIL exhibited significantly thickened vessel walls, which contain deposits of various collagen and extracellular matrix proteins (11–13). However, it is difficult to conduct large-scale histopathological investigations to more fully elucidate the changes of small cerebral arteries, such as the LSAs in CADASIL, because of the limitations in obtaining post-mortem brain samples. Magnetic resonance angiography (MRA) provides an effective, non-invasive method for observing cerebral blood vessels in vivo. However, because of the limitations in signal-to-noise ratio, traditional 3.0-T MRA is incapable of visualizing the intracranial small arteries. Recently, several studies have confirmed the superiority of 7.0-T time-of-flight MRA (TOF-MRA) for examining the intracranial small arteries, especially the LSAs (14–16), providing a powerful tool for the study of CADASIL arteriopathy.

In the present study, we aimed to examine changes of the LSAs and the basal ganglia in patients with CADASIL using 7.0-T MRI, and to analyze the association between LSA measurements and the basal ganglia lesion load, as well as the association between LSA measurements and clinical phenotypes in this patient population.

### MATERIALS AND METHODS

### Patients

The present study was approved by the institutional review board and ethics committee at the Peking University First Hospital, and the study was conducted in accordance with the ethical standards laid down in the 1964 Declaration of Helsinki and its later amendments. Fifty patients with CADASIL and 53 sex- and age-matched healthy controls were recruited and examined after obtaining written informed consent. The diagnosis of CADASIL was based on the gene sequencing results. The positive gene result was defined as the presence of a heterozygous missense mutation, which is pathogenic according to previous studies, in the NOTCH3 gene. If GOM deposits on the basement membrane of VSMCs were found in skin biopsy of the patient, the gene results were considered positive, even though the mutation was not previously reported (5). Healthy controls had no known cerebrovascular disease or related risk factors (e.g., TIA, stroke, diabetes, hypertension, dyslipidemia, cardiac disease, psychiatric illness, major head trauma, or Alzheimer's disease), as confirmed via clinical interviews and examinations. Eleven of the controls admitted to being current or former smokers and 13 controls admitted to alcohol consumption.

The following clinical and demographic data were collected for each patient at the time of inclusion: age, sex, disease duration (determined based on the first occurrence of neurological symptoms), history of hypertension (defined as blood pressure at the time of presentation (≥140/90 mmHg) or previous diagnosis of hypertension), history of diabetes (defined by previous diagnosis), history of hyperlipidemia (defined by previous diagnosis), and history of smoking/alcohol consumption (defined as those who are currently consuming alcohol/smoking tobacco at least once a week, or those who have quit smoking or drinking less than a year ago). We also recorded the clinical symptoms of the patients, such as TIA/stroke, cognitive impairment, etc.

**Abbreviations:** ARWMC, age-related white matter change; CADASIL, cerebral autosomal dominant arteriopathy with subcortical infarcts and leukoencephalopathy; CI, confidence interval; CMBs, cerebral microbleeds; CNR, contrast-to-noise ratio; ICC, intraclass correlation coefficient; LIs, lacunar infarctions; LSAs, lenticulostriate arteries; MMSE, Mini-Mental State Examination; MRI, magnetic resonance imaging; TIA, transient ischemic attack; TOF-MRA, time-of-flight magnetic resonance angiography; WMH, white matter hyperintensity.

Patients with cognitive impairment were defined as those whose Mini-Mental State Examination (MMSE) scores were lower than the lower quartile of the age- and educational level-matched healthy controls (17). The degrees of dependence of all patients were determined using the modified Rankin Scale (mRS) and the Barthel Index (BI) (18).

### Brain MRI Analysis

All patients underwent MRI examination using a 7.0-T wholebody research MR system (Siemens Healthineers, Erlangen, Germany). The following imaging sequences were included in the scanning: T1-weighted (T1w) magnetization-prepared rapid gradient-echo for the localization and the identification of LIs, 3-dimensional (3D) high-resolution TOF-MRA for displaying the LSAs, T2-weighted (T2w) fluid-attenuated inversion recovery (FLAIR) for identifying the WMHs and LIs, and susceptibility weighted imaging (SWI) for detecting CMBs. The imaging parameters of the sequences are summarized in **Supplementary Table 1**.

3D reconstruction and analysis of MRA images were performed using a non-commercial software (Horos <sup>R</sup> ; https:// horosproject.org) (19, 20). Firstly, we examined the whole circle of Willis to exclude the presence of structural abnormalities in large vessels. We excluded the imaging data from three patients and seven controls because of poor image quality caused by head motion, and moreover, excluded the data from one patient because of a history of head trauma. Finally, 46 patients and 46 sex- and age-matched healthy controls were included.

We then counted the number of stems and branches of the LSAs derived from the first segment of the bilateral anterior cerebral arteries (ACAs) and from the first segment of the bilateral middle cerebral arteries (MCAs). Only the blood vessels pointing toward the anterior perforated substances were counted. Stems were defined as LSAs that originated directly from the ACAs or MCAs. The branches were defined as daughter vessels originating from the parent LSA stems, without any subordinate branches (single vessels) (21, 22). If the trunk had no branches, it was recorded as both stem and branch (**Supplementary Figure 1**). Secondary outcome measures included the maximum length of the LSAs and the proportion of discontinuous LSAs in each participant. To measure the maximum length of the LSAs, maximum intensity projections (MIPs) were reconstructed for coronal slabs (thickness: 28 mm) using Horos, and the lengths of the LSAs were determined as the straight axial distance from the highest point of the middle cerebral artery to the end of the longest perforating artery (23, 24). We used the total length of the left and right LSAs in the final data analysis. To evaluate discontinuous LSAs, we first identified arteries with signal interruption on coronal MIP images. We then returned to the axial image to identify and measure the contrast-to-noise ratio (CNR) of the vessel lumen. The CNR was calculated by mean (signallumen) mean (signaltissue) (25). The same region-of-interest was selected to obtain the CNR of 20 arteries. CNR in the normal signal region of arteries was 2.38 ± 0.58, while that in areas with interrupted signal was 1.37 ± 0.31. Discontinuous LSAs were defined as arteries with more than one region with CNR <1.7, and the proportion of discontinuous LSAs was calculated by number of discontinuous LSAs total number of LSAs .

Circular or elliptical lesions with a diameter of 3–15 mm, with a surrounding rim of high signal intensity on the FLAIR sequence, and with the same signal as the cerebrospinal fluid on both T1 and FLAIR sequence were defined as LIs; WMH was defined as a high signal intensity region with a diameter ≥5 mm on FLAIR sequence; Circular lesions with a diameter of 2–10 mm on SWI sequence were defined as CMBs (26). The number of LIs and CMBs on the right and left sides of the basal ganglia region (including the basal ganglia, and the internal and external capsule) were counted manually. The WMH load for the right and left sides of the basal ganglia region was measured using the basal ganglia subscale of the age-related white matter change (ARWMC) scores [**Supplemental Table 2**; (27)]. The basal ganglia lesions of a 51-year-old patient, as well as the LSAs of a 39-year-old patient and an age- and sex-matched control are shown in **Figure 1**.

The number and length of LSAs, as well as the proportion of discontinuous LSAs were examined by a clinician and a radiologist (LC and KQL), and the MRI lesions were assessed by two clinicians (LC and FXJ). The mean of the measurements was used for final statistical analyses. The intraclass correlation coefficient (ICC) for the number of LSA stems of healthy controls was 0.769 [95% confidence interval (CI): 0.618–0.865], that for the number of LSA branches of healthy controls was 0.873 (95% CI: 0.782–0.928), that for the number of LSA stems of CADASIL patients was 0.854 (95% CI: 0.750–0.916), and that for the number of LSA branches of CADASIL patients was 0.867 (95% CI: 0.773–0.924). The ICC for the length of LSAs of healthy controls was 0.992 (95% CI: 0.985–0.995), that for the length of LSAs of CADASIL patients was 0.990 (95% CI: 0.981–0.994), that for the proportion of discontinuous LSAs of healthy controls was 0.873 (95% CI: 0.782–0.928), and that for the proportion of discontinuous LSAs of CADASIL patients was 0.840 (95% CI: 0.729–0.908). Among patients with CADASIL, the ICC for ARWMC scores of the basal ganglia was 0.955 (95% CI: 0.920– 0.975), that for number of LIs in the basal ganglia was 0.869 (95% CI: 0.775–0.925), and that for number of CMBs in the basal ganglia was 0.990 (95% CI: 0.982–0.994).

### Statistical Analysis

Statistical analyses were performed using SPSS version 20.0 (SPSS Inc., Chicago, IL, USA). The normality of the data was analyzed using the Kolmogorov-Smirnov test. Normally distributed data were compared using independent two samples t-tests (t), whereas non-normally distributed data were compared using Mann–Whitney U-tests (z). When grouping by history of smoking/alcohol consumption, all data were analyzed using nonparametric tests due to the small sample size, and the P-value was Bonferroni corrected (Bonferroni corrected P = P ∗ 2). Chi-square tests were used to compare the ratios.

We performed univariate analyses using Spearman rank correlation and Mann–Whitney U-tests. Adjustments were achieved using multiple linear regression (for ARWMC scores and MMSE scores), or logistic regression (for the presence of LIs and the presence of CMBs). ARWMC scores for the basal

ganglia, the presence of LIs in the basal ganglia, the presence of CMBs in the basal ganglia, and MMSE scores were used as dependent variables. Age and LSA measurements, or age and educational level were included as independent variables in the final regression model. Because of the small number of patients with hypertension, diabetes, or hyperlipidemia, we did not include these variables in the final analyses. The ICC was analyzed using a two-way random effects model. Statistical significance was defined as P < 0.05.

# RESULTS

### Clinical Manifestations of Patients With CADASIL

As shown in **Table 1**, among the 46 patients with CADASIL, 38 were symptomatic (45.16 ± 8.90 years; range, 28–63 years), while eight were asymptomatic (34.25 ± 6.52 years; range, 23–43 years). Hypertension, diabetes, and hyperlipidemia were noted in six, two, and six symptomatic patients, respectively. Thirteen symptomatic patients reported a history of smoking, while 17 reported a history of alcohol consumption. Among the asymptomatic patients, with the exception of two patients with a history of smoking and one patient with a history of alcohol consumption, there were no other risk factors for cerebrovascular disease. The mean age at the onset of CADASIL, among symptomatic patients, was 39.53 ± 7.95 years (range, 24–56 years), and the median duration of disease was 5.5 years (range, 0–17 years). There were 35 patients had a history of TIA/stroke and nine patients suffered from cognition impairment.

TABLE 1 | Demographic and clinical features of patients with CADASIL.


*MRI, magnetic resonance imaging; CADASIL, cerebral autosomal dominant arteriopathy with subcortical infarcts and leukoencephalopathy; TIA, transient ischemia attack.*

### Changes in LSAs in Patients With CADASIL

There was no significant difference in age (t = 1.036, P = 0.303) between the CADASIL patients and healthy controls. The number of LSA branches was lower in patients than in controls (t = −2.591, P = 0.011), whereas no significant difference in the number of stems was found between the groups (z = −1.617, P = 0.106). An increased proportion of discontinuous LSAs was also observed in patients with CADASIL (z = −1.991, P = 0.047). However, there was no significant difference in the total length of LSAs between the patients and healthy individuals (t = −0.412, P = 0.682) (**Table 2**).

In addition, we observed fewer LSA branches in patients with a history of alcohol consumption (z = −2.247, Bonferroni corrected P = 0.049) than in those without such a history, but we did not find a significant change in the number of LSA branches in patients with a history of smoking (z = −2.221, Bonferroni corrected P = 0.053) (**Figure 2**). There was no significant difference in the proportion of discontinuous LSAs between smokers/drinkers and non-smokers/nondrinkers (smokers vs. non-smokers: z = −1.145, Bonferroni corrected P = 0.504; drinkers vs. non-drinkers: z = −0.034, Bonferroni corrected P = 1.946). Moreover, there was no significant difference in the total length of LSAs between smokers/drinkers and non-smokers/non-drinkers (smokers vs. non-smokers: z = −1.019, Bonferroni corrected P = 0.616; drinkers vs. non-drinkers: z = −1.305, Bonferroni corrected P = 0.384).

# Association Between LSA Measurements and Basal Ganglia Lesion Load of CADASIL Patients

In CADASIL patients, we did not find an association between LSAs measurements and the basal ganglia lesion load in univariate analyses. However, we found that patients with a history of alcohol consumption had more CMBs (z = −2.026; P = 0.043) and more LIs (z = −2.424; P = 0.015) in the basal ganglia (**Supplementary Table 3**).

After adjusting for age, there was no significant association between the number of LSAs and ARWMC scores (β = −0.049; 95% CI: −0.219, 0.121; P = 0.565) for the basal ganglia, the presence of LIs (OR = 0.878; 95% CI: 0.698, 1.103; P = 0.264) in the basal ganglia, or the presence of CMBs (OR = 1.011; 95% CI: 0.799, 1.279; P = 0.927) in the basal ganglia in CADASIL patients. Moreover, we observed no significant association between the proportion of discontinuous LSAs and ARWMC scores (β = 0.002; 95% CI: −0.032, 0.035; P = 0.912) for the basal ganglia, the presence of LIs (OR = 1.041; 95% CI: 0.988, 1.098; P = 0.134) in the basal ganglia, or the presence of CMBs (OR = 0.982; 95% CI: 0.932, 1.035; P = 0.500) in the basal ganglia in the patient group after adjusting for age. There was also no significant association between the length of LSAs and ARWMC scores (β = −0.030; 95% CI: −0.099, 0.039; P = 0.387) for the basal ganglia, the presence of LIs (OR = 0.980; 95% CI: 0.895, 1.073; P = 0.665) in the basal ganglia, or the presence of CMBs (OR = 0.951; 95% CI: 0.860, 1.051; P = 0.326) in the basal ganglia in the patient group after adjusting for age (**Table 3**). However, after adjusting for age and LSAs measurements, alcohol consumption increased the risk of CMBs and LIs in the basal ganglia in the patient group (CMBs: OR = 6.000; 95% CI: 1.472, 24.454; P = 0.012; Nagelkerke R <sup>2</sup> = 0.199) (LIs: OR = 6.299; 95% CI: 1.394, 28.456; P = 0.017; Nagelkerke R <sup>2</sup> = 0.278).

### Association Between LSA Measurements and Clinical Phenotypes of CADASIL Patients

Using univariate analysis, we found an association between the number of LSA branches and MMSE scores of CADASIL patients (ρ = 0.413; P = 0.014). However, no significant association was found between the number of LSA branches and the mRS/BI scores (mRS scores: ρ = −0.142, P = 0.348; BI scores: ρ = 0.039, P = 0.799). The proportion of discontinuous LSAs and the length of the LSAs were not significantly associated with all the three clinical scores of the patients (**Table 4**). After adjusting for age and educational level, there was still a positive association between the number of LSA branches and MMSE scores in CADASIL patients (β = 0.438; 95% CI: 0.093, 0.782; P = 0.014). No significant difference in LSA measurements was found between patients with a history of TIA/stroke and patients without a history of TIA/stroke (the number of LSA branches: z = −0.478, P = 0.633; the length of LSAs: z = −0.167, P = 0.867; the proportion of discontinuous LSAs: z = −1.363, P = 0.173).

# DISCUSSION

In the present study, we aimed to analyze the changes in LSAs among patients with CADASIL using a 7.0-T MRI. Our findings indicate that patients with CADASIL exhibit fewer LSA branches and a higher proportion of discontinuous LSAs than healthy individuals. Although there was no association between the measurements of LSAs and the basal ganglia lesion load in patients with CADASIL, we observed a positive association between the number of LSA branches and MMSE scores in CADASIL patients.

Clinical manifestations and MRI features in our patients were consistent with those previously reported in Chinese CADASIL cohorts (3, 5, 28). The median number and length of LSAs in our cohort of healthy individuals were similar to that observed in previous studies (14, 23, 24), confirming the reliability of our imaging strategies. We observed that alcohol consumption aggravated damage to the LSAs, consistent with the well-known harmful effects of drinking on the cerebrovascular system. Therefore, controlling alcohol intake is especially important for patients with CADASIL. Presently, our results did not show a significant decrease in the number of LSAs in patients with a history of smoking. However, as the Bonferroni corrected P-value is close to 0.05 (Bonferroni corrected P = 0.053), we presume that further increase in sample size may lead to a significant decrease in the number of LSAs in patients with a history of smoking.

In the present study, patients with CADASIL exhibited a decrease in the number of LSA branches and an increase in the proportion of discontinuous LSAs. Since TOF-MRA is based on the in-flow effect of cerebral blood flow, these results suggested that there was an interruption in blood flow. Some post-mortem studies have indicated that patients with CADASIL exhibit stenosis and occlusion of the cerebral small arteries (29, 30), whereas other studies have suggested abnormalities in hemodynamics and vasomotor activities in both patients with


*LSAs, lenticulostriate arteries; Age, number of LSA branches and length of LSAs are presented as mean* ± *SD. Number of LSA stems and proportion of discontinuous LSAs are presented as median with 95% CI.* \**Significant difference.*

*a Independent two samples t-test, and 95% CI of the difference is shown.*

*<sup>b</sup>Mann-Whitney U-test, and p25–p75 of the transformation rank is shown.*

*<sup>c</sup>Chi-square test.*

CADASIL and mouse models of CADASIL (31–33). Because of the small luminal diameter of LSA branches and the limited resolution of MRA devices, decreases in blood flow caused by stenosis or hemodynamic changes may go undistinguished and manifest as decreased number or discontinuity of arteries on MRA images. Further studies are required to determine which of these two factors plays a major role. A previous study involving 22 patients with CADASIL and 11 healthy controls reported that there were no changes in the number of LSAs in patients with CADASIL (23), inconsistent with our findings. Because our study included 46 patients as well as 46 age- and sex-matched controls, these discrepancies may be related to the differences in sample size. Thus, 7.0- T MRI is a promising and non-invasive method for the study of small artery damage in CADASIL, which may aid evaluation of the clinical condition of CADASIL patients in the future.

In our study, there was no association between LSA measurements and the basal ganglia lesion load, consistent with the findings of a previous 7.0-T MRI-based study on patients with CADASIL (23). LIs and WMHs are usually thought to be caused by hypoperfusion, which can also be attributed to hemodynamic abnormalities other than arterial stenosis. Additional studies have suggested that WMHs can also be attributed to a dysfunctional blood-brain barrier (34, 35). Indeed, hemodynamic abnormalities and dysfunctional blood-brain barrier have been observed in studies involving both patients with CADASIL and animal models of CADASIL (31, 33, 36, 37). Thus, we speculate that basal ganglia lesions in patients with CADASIL may be caused by hemodynamic abnormalities or a dysfunctional blood-brain barrier. This may explain why we were unable to identify an association between LSA changes and the basal ganglia lesion load. In addition, although the resolution of 7.0-T MRA has improved, it is still impossible to observe vessels with diameter <250µm by in vivo imaging (23). Therefore, the possibility that the stenosis of the lumen of smaller vessels leads to the lesions of the basal ganglia cannot be ruled out. It is also possible that the number of patients in our study was too small to yield a significant association. Further, we observed that alcohol consumption significantly increased the risk of CMBs and LIs in the basal ganglia in the patient group, highlighting the importance of controlling alcohol intake among patients with CADASIL.

TABLE 3 | Age adjusted association between LSA measurements and MRI lesion load of the basal ganglia in CADASIL patients.


*LSAs, lenticulostriate arteries; LIs, lacunar infarctions; ARWMC, age-related white matter change; CMBs, cerebral microbleeds; CI, confidence interval. <sup>a</sup>Multivariate linear regression analysis ("enter" model).*

*<sup>b</sup>Binary logistic regression ("enter" model).*

TABLE 4 | Association between LSA measurements and clinical phenotypes of CADASIL patients.


*LSAs, lenticulostriate arteries; MMSE, Mini-Mental State Examination; mRS, modified Rankin Scale; BI, Barthel Index.* \**Significant difference. Data were analyzed using Spearman rank correlation (*ρ*).*

In addition, we found a positive association between the number of LSA branches and MMSE scores in CADASIL patients, suggesting that abnormalities of small arteries may be related to some clinical symptoms of CADASIL patients. There may be two explanations for this association. Firstly, in recent years, the importance of the basal ganglia in cognition has been reported by many studies, and it is known to participate in several cognitive pathways such as executive function, procedural memory, and attention (38, 39). Studies of type 1 diabetes have suggested that reduced cerebral blood flow in the bilateral caudate nucleus-thalamus is associated with abnormal executive function (39). Therefore, the impaired blood supply and basal ganglia dysfunction caused by LSA abnormalities in CADASIL patients may directly lead to cognitive impairment. Secondly, the LSAs are a part of the cerebral perforating artery system, and therefore the LSA abnormalities we observed may indirectly reflect changes to the whole cerebral perforating artery system. Abnormal cerebral perfusion and brain tissue damage caused by changes to the cerebral perforating artery system could further lead to cognitive impairment in CADASIL patients. However, the above hypothesis lacks direct evidence and needs further research to confirm its validity.

The present study possesses several limitations of note. Because CADASIL is a rare disease, our analysis is inherently limited by weaknesses in the case-control study design, including imperfect matching, inevitable recall bias, and difficulty in determining causal relationships. In addition, the method for measuring the length of LSAs was simplified because of the lack of a software for tracking and reconstructing LSAs. Although this modified method may reflect the extensive stenosis of LSAs, such measurements are easily affected by the curvature of blood vessels and may not reflect the true length of these vessels.

# CONCLUSIONS

We have shown that patients with CADASIL exhibit fewer LSA branches and a higher proportion of discontinuous LSAs than healthy individuals when examined using 7.0- T MRI. This suggests that 7.0-T MRI is a promising and non-invasive method for the study of small artery damage in CADASIL. The abnormalities of small arteries may be related to some of the clinical symptoms of CADASIL patients such as cognitive impairment. However, since we observed no association between the LSA measurements and the basal ganglia lesion load, the changes in the basal ganglia due to CADASIL are most likely caused by mechanisms other than the anatomic narrowing of the vessel lumen, such as hemodynamic abnormalities or a dysfunctional bloodbrain barrier.

# DATA AVAILABILITY

The raw data supporting the conclusions of this manuscript will be made available by the authors, without undue reservation, to any qualified researcher.

# ETHICS STATEMENT

The present study was approved by the institutional review board and ethics committee at Peking University First Hospital and has been performed in accordance with the ethical standards laid down in the 1964 Declaration of Helsinki and its later amendments. Fifty patients with CADASIL and 53 age-matched healthy controls were recruited and examined after obtaining written informed consent.

# AUTHOR CONTRIBUTIONS

YY and ZZ contributed conception and design of the study. CL and XF collected and organized the database, performed the MRI analysis and the statistical analysis. CL and QK performed the MRI analysis. CL wrote the first draft of the manuscript. QK, YS, WZ, and ZW wrote sections of the manuscript. BW, YZ, and JA participated in the design and improvement of MRI scan sequences. All authors contributed to manuscript revision, read, and approved the submitted version.

### FUNDING

This work was supported by the Ministry of Science and Technology (2011ZX09307-001-07); the National Key Research and Development of China (2016YFC1300605); the National Natural Science Foundation of China (81471185); the Beijing Municipal Natural Science Foundation (7184226);

## REFERENCES


the Young Elite Scientists Sponsorship Program by CAST (2017QNRC001); and the Ministry of Science and Technology of China (2015CB351701).

# ACKNOWLEDGMENTS

We thank Meng Yu (Department of Neurology, Peking University First Hospital) and Junlong Shu (Department of Neurology, Peking University First Hospital) for their help with MRI data analysis. We also thank Qingqing Wang (Department of Neurology, Peking University First Hospital) and Weili Yang (Department of Neurology, Peking University First Hospital) for administrative assistance.

### SUPPLEMENTARY MATERIAL

The Supplementary Material for this article can be found online at: https://www.frontiersin.org/articles/10.3389/fneur. 2019.00870/full#supplementary-material

in CADASIL mice. Ann Neurol. (2016) 79:387–403. doi: 10.1002/ana. 24573


subcortical infarcts and leukoencephalopathy. Stroke. (2010) 41:2812–6. doi: 10.1161/STROKEAHA.110.586883


**Conflict of Interest Statement:** JA was employed by Siemens Shenzhen company.

The remaining authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

Copyright © 2019 Ling, Fang, Kong, Sun, Wang, Zhuo, An, Zhang, Wang, Zhang and Yuan. This is an open-access article distributed under the terms of the Creative Commons Attribution License (CC BY). The use, distribution or reproduction in other forums is permitted, provided the original author(s) and the copyright owner(s) are credited and that the original publication in this journal is cited, in accordance with accepted academic practice. No use, distribution or reproduction is permitted which does not comply with these terms.

# Simultaneous EEG-fMRI for Functional Neurological Assessment

Giulia Mele, Carlo Cavaliere\*, Vincenzo Alfano, Mario Orsini, Marco Salvatore and Marco Aiello

*IRCCS SDN, Naples, Italy*

The increasing incidence of neurodegenerative and psychiatric diseases requires increasingly sophisticated tools for their diagnosis and monitoring. Clinical assessment takes advantage of objective parameters extracted by electroencephalogram and magnetic resonance imaging (MRI) among others, to support clinical management of neurological diseases. The complementarity of these two tools can be now emphasized by the possibility of integrating the two technologies in a hybrid solution, allowing simultaneous acquisition of the two signals by the novel EEG-fMRI technology. This review will focus on simultaneous EEG-fMRI technology and related early studies, dealing about issues related to the acquisition and processing of simultaneous signals, and including critical discussion about clinical and technological perspectives.

### Edited by:

*Brad Manor, Institute for Aging Research, United States*

### Reviewed by:

*Junhong Zhou, Harvard Medical School, United States Bo Gao, Affiliated Hospital of Guizhou Medical University, China*

> \*Correspondence: *Carlo Cavaliere carlocavaliere1983@yahoo.it*

### Specialty section:

*This article was submitted to Applied Neuroimaging, a section of the journal Frontiers in Neurology*

Received: *28 February 2019* Accepted: *22 July 2019* Published: *13 August 2019*

### Citation:

*Mele G, Cavaliere C, Alfano V, Orsini M, Salvatore M and Aiello M (2019) Simultaneous EEG-fMRI for Functional Neurological Assessment. Front. Neurol. 10:848. doi: 10.3389/fneur.2019.00848* Keywords: EEG, fMRI, multimodal image analysis, functional connectivity, EEG spectra

### INTRODUCTION

The incidence of neurodegenerative and psychiatric diseases has increased in the last decades, requiring finer, and advanced tools, ranging from electrophysiology to neuroimaging, for a reliable diagnostic accuracy. Electrophysiology, and specifically the electroencephalogram (EEG), represents a consolidated, and widespread tool supporting the diagnosis of neurological diseases. Unlike imaging techniques, EEG offers an excellent temporal resolution, recording the electric brain activity in the order of milliseconds through electrodes placed on the scalp. Through EEG signal processing techniques, and dedicated experimental setup, quantitative parameters on the spectrum of frequencies, amplitudes, and coherence can be achieved.

Conversely, computed tomography (CT), and mainly MRI, provide a morphological view of brain (1), with an excellent spatial resolution, allowing a multiparametric assessment of the brain tissue properties, both in terms of structural and functional information. In this context, similarly to EEG but at different temporal scales (milliseconds vs. seconds), functional MRI (fMRI) allows for non-invasive investigation of brain functional activation both during resting state and task execution, enriching the panel of parameters achievable by MRI (e.g., structural connectivity revealed by diffusion tensor imaging, metabolites concentrations revealed by magnetic resonance spectroscopy, and perfusion revealed by arterial spin labeling). This complementarity of information is deeply exploited by multimodal acquisition systems that are developed to overcome single modality drawbacks and to improve the compliance of the patients. Both in preclinical and clinical settings (2–6), first multimodal imaging techniques attempted to combine functional information derived by nuclear medicine modalities (positron emission tomography—PET, and single photon emission computed tomography—SPECT) with structural data achieved by CT and MRI, in order to complement diagnostic and prognostic approach to different kind of patients (7). In neurology, simultaneous PET/MRI paved the way for a Mele et al. EEG-fMRI in Neurology

more comprehensive investigation of brain organization and physiology, allowing to investigate, within a single integrated exam, the cerebral connectivity in terms of structural, functional, and metabolic connectome (8, 9). Recently, to fully investigate healthy and pathological brain function, novel tools have been developed to simultaneously acquire EEG and fMRI signals, integrating the optimal temporal and spatial resolution of both techniques and overcoming the limitations of single modalities.

In this review, simultaneous EEG-fMRI technology, detailing current applications using both resting state and task approaches and discussing future perspectives will be focused.

### EEG

EEG is one of the most used techniques for studying brain electrical activity. The first acquisition of an electroencephalograph was made more than 50 years ago by Berger, who recorded brain electrical activity via a radio equipment. The discovery of EEG and, consequently, of cerebral electrical activity definitely changed the way of approaching to the study of brain structures and functions, and over the time became a fundamental tool in both clinical and research fields (10). Brain electrical activity is derived from the synchronizations of a pool of cortical neurons, in particular of pyramidal cells. These cells present a different electrical charge along the neuron, resulting negative on dendrites and positive in the rest of cell. This difference determines an electric dipole that can be acquired by EEG electrodes, and represented as a series of positive and negative waves. However, the electric field derived by a single pyramidal cell is not enough to obtain a detectable EEG signal. For this reason, the electrodes record a pool of cells arranged parallel to each other, and producing radial and tangential dipoles (11, 12). The EEG is acquired through the positioning of electrodes on the scalp according to the international 10–20 system (13), which takes into account four main reference point: nasion, inion, and the two preauricular points (A1, A2) (14). The electrodes are fixed to the scalp by means of a conductive paste and recorded a lot of brain oscillations including delta rhythm (0.5–4 Hz), theta rhythm (4–8 Hz), alpha rhythm (8– 13 Hz), beta rhythm (13–30 Hz), and gamma rhythm (above 30 Hz) (15) (**Figure 1**) Moreover, during the task execution, it is possible to record evoked potentials that allow to study different neuronal processes (16). The evoked potentials can be divided according to latency. In fact, the potentials that occur within the 100 ms post stimulus are usually due to the nature of the stimulus itself, while the subsequent components reflect the cognitive processes related to the perception of the stimulus (Shravani et al., 2009).Technological innovations have led to the development of high-density EEG systems with a high number of channels/electrodes for quantitative EEG and

brain connectivity studies (17). Currently, within a clinical setting (configuration with about 20 electrodes), the EEG is used to characterize numerous diseases including metabolic or drug alterations, sleep disorders, epileptic syndromes, neurodegenerative diseases, traumatic brain injury, and tumor lesions, and the characterization of comatose patients and brain death (18).

### fMRI

The fMRI is one of the main non-invasive techniques that allow to measure brain function. The mechanism that subtends the signal of fMRI is called blood oxygen level dependent (BOLD) effect, that describes the variation in the magnetic status of the red blood cells linked to the hemoglobin oxygenation. Indeed, the form of hemoglobin without oxygen is deoxyhemoglobin, which has paramagnetic property, while oxyhemoglobin has diamagnetic property. In resting conditions, the balance between these two elements concentrations in the vascular brain system, provides a signal indistinguishable from the surrounding parenchyma. When a stimulus was applied, the hemoglobin balance in specific brain areas changes, initially in favor to deoxyhemoglobin concentration and so determining a decrease

**Abbreviations:** CT, Computed Tomography; MRI, Magnetic Resonance Imaging; PET, Positron Emission Tomography; EEG, Electroencephalography; fMRI, Functional magnetic resonance imaging; BOLD, Blood oxygenation-leveldependent; rsfMRI, resting state functional Magnetic Resonance Imaging; JME, Juvenile Myoclonic Epilepsy; DMN, Default Mode Network; BNG, Basal Ganglia; SRN, Self-reference; PTSD, Post-Traumatic Stress Disorder; AD, Alzheimer Disease; SM, Multiple Sclerosis.

of signal, and following switching in favor to oxyhemoglobin concentration and a signal increase (19). The detection of these signal changes translates into a series of images, that can be analyzed to show the activations of specific brain areas, following the execution of specific tasks. It is important to understand that BOLD effect is an indirect measure of neuronal activation, depending from neurovascular coupling and so by different interplay, such as alteration in blood flow and volume and complex interactions between the activated neurocircuitry with astrocytic and vascular targets. Briefly, neuronal activation induced by the stimulus determines a neurotransmitter release in the synaptic cleft and its uptake by the astrocytic process, in the so-called tripartite synapse (20, 21). The secondary astrocytic activation triggers intracellular Ca2+fluctuations in astrocyte end-feet that elicit cellular molecular and hemodynamic changes recorded by fMRI through the release of vasoactive peptides (22). This complex cascade of events that subtend neurovascular coupling and BOLD effect is also responsible of the time delay between neuronal activation and BOLD signal fluctuation that distinguish fMRI from direct electrophysiological measures.

In this context, while task-related fMRI has been applied in many studies to investigate specific functions and/or brain areas (23), more recently, resting-state fMRI approach is coming out to analyze spontaneous physiological fluctuations without the need of patient's compliance, pathway's integrity and command following, sometime impossible in several kind of patients (24, 25).

Since from its development, fMRI technique has been applied to characterize brain functional connectivity in several physiological conditions (26, 27) and many diseases, including brain tumors (28), multiple sclerosis (29), Alzheimer's diseases (AD) (30, 31), epilepsy (32), but also psychiatric disorders (33, 34).

### SIMULTANEOUS EEG-fMRI

Simultaneous EEG-fMRI acquisition is used to evaluate the correlation between electrical brain activity and hemodynamic mutation. fMRI with high spatial resolution does not provide adequate temporal sampling due to the slow BOLD response (in order of seconds) unlike EEG that instead offers a high temporal resolution (in the order of milliseconds), but with a poor localization of signal sources (35). The integration of these two tools in a hybrid simultaneous acquisition allows to overcome the intrinsic limitations of both the techniques and to increase the plethora of analyses that can be performed, and in turns, of the information that can be achieved (36). Simultaneous acquisition also guarantees an identical registration, as regards the mental state of the subject, the execution of the task and the inference of the recording environment. This does not happen by recording the two methods separately, especially if the recording takes place in different environments and with cognitive unstable patients (37).

As for technological issues, the acquisition of simultaneous EEG/fMRI involves the use of specialized EEG hardware that is safe and compatible with the MR environment and comfortable to the participant. Improper use of the equipment may result in considerable risks. Regarding safety, a potential risk for the subjects comes from electrodes and heating of conducting leads during MR radio frequency transmission, resulting in discomfort

or even burns (38). To reduce the risk of subject discomfort or injuries, there are several precautions, for example fMRI sequences should be based on gradient echo-echo planar imaging (GE-EPI); for anatomical reference scans, low specific absorption rate (SAR) sequences should be used, in particular GE-T1 weighted sequences; for all sequences in EEG-fMRI protocol, it should be ascertained that their SAR does not exceed the SAR of the GE-EPI sequence. Otherwise, extensive safety testing with temperature sensors is necessary. Staff performing EEG-MRI studies must have received appropriate training, as injuries due to MR-compatible EEG equipment cannot be ruled out if the equipment is accidentally used out of specifications, especially in the case of body coil transmission (39). The adoption of these guidelines is particularly important in vigilance-reduced subjects (sleeping or sedated subjects) or, generally, in subjects who cannot give notice of any discomfort reliably (children).

Regarding the compliance of the subjects, it is important when using EEG/fMRI to make sure that they have a good understanding of all steps involved, that they are comfortable with all steps, and that there are no accidents that could cause discomfort leading to movement and resulting in failure of the experiment (40). The participants should understand that nothing will be painful even if some steps may be slightly uncomfortable, such as slight abrasion of the scalp during placement of EEG electrodes; this helps eliminate much of the anxiety that the participant might otherwise have, in order to complete the experiment properly and safely.

Moreover, the data obtained from the simultaneous acquisition of EEG-fMRI are strongly influenced by artifacts. On the one hand the presence of the helmet generates a variation in the homogeneity of the magnetic field that involved a variation in images quality, on the other hand the presence of the

magnetic field itself generates broad-band artifacts, which almost completely cover the electroencephalographic signal (**Figure 2**) (41). Moreover, the movement of the electrodes caused by pulserelated in the static magnetic field generates a ballistocardiogram artifact that is influenced by the spatio-temporal variability of cardiac cycles in place during recording (**Figure 3**) (42). For this reason, researchers developed different methods to remove artifact, such as independent component analysis (ICA), that is considered the best method for remove ballistocardiogram artifact (43), or Fourier transform that can be used to correct gradient artifacts (44).

Scientific articles published since 2014 on PubMed website, using as key word "simultaneous EEG-fMRI" have been collected in order to include studies with simultaneous acquisitions of EEG-fMRI both in resting state and during tasks execution (**Figure 4**).

### Simultaneous Resting-State EEG-fMRI

Brain is a dynamic system that generates activity even in a state of rest (**Table 1**). This can be revealed by EEG recording through the detection of neural waves with different frequency and amplitude and by fMRI through the estimation of different resting state networks linked to specific cerebral functions. The simultaneous acquisition of rsfMRI and EEG makes it possible to consider the brain as a series of systems or networks that interact with each other (47, 51). The interactions are dependent by the concurrent variation of BOLD fluctuations and brain electrical activities. There are many fields of application of simultaneous acquisition of rsfMRI and EEG (**Table 1**). First studies have focused on methodological issues in healthy subjects, analyzing the reconstruction of EEG signal sources, based on fMRI information, and mainly oriented to a connectivity analysis. However, it is considered necessary to implement the study sample in order to validate the theory (47). The authors demonstrated that simultaneous approach using a 64 channel MR-compatible EEG cap in seventeen adult volunteers is useful to validate whole-brain connectomes extracted by each modality and to elaborate predictive model of dynamic functional connectivity (47). Another study (36) correlated theta and delta frequencies of the temporal lobe with simultaneous fMRI acquisition in fourteen healthy sleep-deprived subjects in awake and drowsy states. The study identified, for the first time, a different brain regional source for the delta and theta rhythms, although their analysis also includes the fastest rhythms, such as alpha, beta and gamma. This kind of approach produces a greater differentiation of the slow rhythms, but decreases the localization of the sources generating different EEG bands. The electrical-BOLD correlation seemed to be stronger for frequencies lower than 1 Hz, and influenced by the spatial relationship between the resting state networks analyzed and the recording zones (48). This relationship has also been used to investigate the basis of some specific electrical oscillations such as the mu rhythm. In a study conducted on thirty-six healthy subjects, simultaneous acquisition of EEG-fMRI has allowed to identify a positive correlation between the power of mu rhythm and the BOLD signal in areas including the anterior cingulate cortex and the anterior insula, confirming the multiple origin of this specific rhythm (50). Concerning neurological diseases applications, a study on eighteen subjects affected by juvenile myoclonic epilepsy demonstrated the added value of the EEGfMRI acquisition to unveil the pathophysiology of the disease, highlighting the relationships between the frontal networks and the epileptic discharges (46). Another study (45) detected a reduced association between occipital alpha band power and the fluctuation of the BOLD signal in frontal and temporal cortices and in the thalami of fourteen AD patients. In psychiatry, other authors demonstrated a close relationship between the temporal dynamics of default mode network and post-traumatic stress disorder (PTSD) severity in thirty-six veterans, compared to twenty combat-exposed controls (49). It becomes clear that the simultaneous recording of EEG-fMRI can give substantial information on the relationships between the hemodynamic response and neuronal activity. In particular, the resting state acquisition can be fundamental for underling the variability of brain activity and above all to define the structures generally involved in the triggering EEG waves in resting state. In this case, increasing the sample size and using different methods of analysis could validate previous results and disentangle inconsistent or controversial findings.

### Simultaneous Task EEG-fMRI

The execution of tasks allows to establish, according to the cognitive domain studied, which cerebral areas are assigned to the specific task (**Table 2**). According to studies performed with a recognition memory task, EEG-fMRI experiments have demonstrated that theta-alpha low frequency oscillations (4– 13 Hz) are linked to the functional activation of a network involving the hippocampus, the striatum and the prefrontal cortex. These findings confirmed the theory that the hippocampus acts as a modulator of brain activity by acting through low frequency oscillations (52). Hippocampus seems to have an important role also during sleep. In fact, it was demonstrated that hippocampus activity increases during light sleep in relationship with alpha activity (58). It could confirm the idea that memory fixation could occur in light sleep phases,

### TABLE 1 | A summary of the resting state EEG-fMRI studies since 2014.


### TABLE 2 |A summary of the task EEG-fMRI studies since 2014.


in motor rehabilitation.

Mele et al.

although the acquired subjects had not performed any learning task (58). As for decision making assessment, a simultaneous approach has been employed to investigate common neural substrates for perceptual decisions and accumulation of evidences, highlighting a common role for the posterior medial frontal cortex in both the processes (54). In another study using a two-choice decision-making paradigm, the authors demonstrate that an increase in theta band power, associated with a choice with a negative feedback, corresponds to the activation of fronto-parietal areas; at contrary, an increase in the power of the beta band, associated with a positive feedback, reflects the activation of subcortical are as involved in the reward network (55). Other authors employed the gambling task paradigm in 20 healthy controls to analyze the concurrent activation of large areas related to the reward and punishment, such as posterior cingulate, medial pre-frontal cortex and ventral striatum (56).

A further application of EEG-fMRI is represented by neuro feedback, which allows the modulation of the brain activities, although up to now the information that come back to the patient belong to only EEG (57) or fMRI scan (53, 56, 59). As for EEG neurofeedback, several authors have compared brain activation during motion imaginations and movement execution in healthy subjects, suggesting a role for this approach in the rehabilitation of patients affected by post-stroke paralysis (57). As for fMRI neurofeedback, two studies have investigated the correlation between EEG rhythms and BOLD signal following behavioral modulation. The first one, in a sample of 34 healthy subjects, reported that the modulation of thalamic nuclei activation during the retrieval of happy autobiographical memories, is able to modulate both the alpha activity and the BOLD signal (59). The second one, performed by the same group, in a population of patients affected byPTSD, showed that emotional control training can improve the alpha rhythm and the functional connectivity between the amygdala and the prefrontal cortex, and this enhancement was correlated with a better clinical performance (53).

Up to now, only one article reported the implementation of a novel simultaneous real time fMRI and EEG neurofeedback (60). The authors demonstrated that the training of emotional self-regulation in healthy subjects, based on retrieval of happy autobiographical memories, can modulate both amygdala BOLD fMRI activation and beta band EEG power asymmetry (60). Summarizing, major evidences derived from task-related EEG-fMRI focus on emotional and cognitive processes. This certainly represents a great starting point for understanding and discovering everything concerning psychiatric and neurological syndromes that still remain a big question mark. Although the multimodal approach determines several issues that can complicate the research process, simultaneous EEG-fMRI

FIGURE 5 | Source analysis. (Left) Sources localization of the EEG frequencies for a time period of 3 s, accomplished through the LORETA analysis. (Right) 3-s period electroencephalographic pattern of a healthy subject. Image obtained on a 40 years-old healthy volunteer with hybrid EEG-fMRI system and included for illustrative purpose only.

acquisition remains one of the most appreciated approach, which certainly allows a complete view of brain activity, without affecting the state of patients and subjects participating in the study.

### EEG-fMRI ANALYSIS METHODS

Data analysis is a fundamental step for EEG-fMRI research studies and, in general, for simultaneous multimodal acquisitions. The various analysis used can be contained in two macro-areas: symmetric analysis and integrated analysis (61). Briefly, the symmetrical approach involves the simultaneous analysis of the data extracted from the two methods, while the integrated analysis exploits the data collected by one of the two methods, to understand and validate the data collected from the other one. In this way, it is possible to generate a unique model that facilitates the understanding of brain activity (62). In particular, integrated analyzes include two methods of applications: EEG-informed fMRI (63) and fMRI-informed EEG (64). The first one uses brain electrical activity to predict hemodynamic variations (15). The second one, uses the activation maps extracted by the fMRI to correct and analysis the EEG sources (**Figure 5**) (65).

Nevertheless, the optimal procedure for the analysis of the simultaneous EEG-fMRI data is still an open issue that needs further investigation in order to extract meaningful quantitative biomarkers, useful to characterize physiological and pathological brain activity, taking advantages by mutual information.

### FUTURE PERSPECTIVES

Even if MRI and EEG complement each other considering their different spatial and temporal resolution, the characterization of molecular processes that subtend resting state analysis or a specific task is not achievable through these tools. For this reason, a trimodal approach integrating an MR-compatible EEGsystem in the hybrid MR–PET scanner has been proposed and successfully implemented (66). The trimodal acquisition certainly allows a broader and integrative view of the brain

# REFERENCES


activity, although technical issues derived by the PET attenuation of the EEG cap are debated (67, 68).

In an exploratory pilot study, 10 healthy subjects are analyzed in order to implement the value of the single technique and explore the human brain through the different information provided with the same physiological and psychological condition of the subject. The results of these early studies pave the way for further research on different patient populations to exploit the mutual clinical potential of the methods (69).

This kind of approach appeared promising, ensuring the same physiological conditions for all measurements, with the possibility to acquire other synergistic information like perfusion and diffusion changes via MR-based methods.

### CONCLUSIONS

Simultaneous EEG-fMRI acquisition represents a reference tool to evaluate the correlation between brain electrical activity and BOLD signal. This technique appeared essential to investigate physiological brain networks in healthy subjects, introducing new evidences about the electrical neural activity and the neurovascular coupling underpinning the BOLD signal. Moreover, it offers the possibility to characterize the relationship between EEG spectrum and regional brain activation, providing new insights on neurological and psychiatric diseases and, hopefully, new treatment targets.

Despite the increasing use of EEG-fMRI, as other multimodal techniques, the question about the optimal integrated and standardized analysis is still open, representing the true challenge that follows the technological development.

### AUTHOR CONTRIBUTIONS

GM and CC substantial contributed to the conception and design of the work. GM, VA, and MO prepared the literature database and drafted the work. CC, MS, and MA revised critically the manuscript for important intellectual content. MS provided approval for publication of the content.


systematic review. Schizophr Res. (2017) 189:9–18. doi: 10.1016/j.schres.2017. 02.026


EEG-fMRI study in healthy male subjects. Neurophysiol. (2019) 130:1375–86. doi: 10.1016/j.clinph.2019.04.715


**Conflict of Interest Statement:** The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

Copyright © 2019 Mele, Cavaliere, Alfano, Orsini, Salvatore and Aiello. This is an open-access article distributed under the terms of the Creative Commons Attribution License (CC BY). The use, distribution or reproduction in other forums is permitted, provided the original author(s) and the copyright owner(s) are credited and that the original publication in this journal is cited, in accordance with accepted academic practice. No use, distribution or reproduction is permitted which does not comply with these terms.

# Applications of Deep Learning to Neuro-Imaging Techniques

Guangming Zhu, Bin Jiang, Liz Tong, Yuan Xie, Greg Zaharchuk and Max Wintermark\*

Neuroradiology Section, Department of Radiology, Stanford Healthcare, Stanford, CA, United States

Many clinical applications based on deep learning and pertaining to radiology have been proposed and studied in radiology for classification, risk assessment, segmentation tasks, diagnosis, prognosis, and even prediction of therapy responses. There are many other innovative applications of AI in various technical aspects of medical imaging, particularly applied to the acquisition of images, ranging from removing image artifacts, normalizing/harmonizing images, improving image quality, lowering radiation and contrast dose, and shortening the duration of imaging studies. This article will address this topic and will seek to present an overview of deep learning applied to neuroimaging techniques.

Keywords: artificial intelligence, deep learning, radiology, neuro-imaging, acquisition

### Edited by:

Hongyu An, Washington University in St. Louis, United States

### Reviewed by:

Yann Quidé, University of New South Wales, Australia Chia-Ling Phuah, Washington University School of Medicine in St. Louis, United States

> \*Correspondence: Max Wintermark max.wintermark@gmail.com

### Specialty section:

This article was submitted to Applied Neuroimaging, a section of the journal Frontiers in Neurology

Received: 13 March 2019 Accepted: 26 July 2019 Published: 14 August 2019

### Citation:

Zhu G, Jiang B, Tong L, Xie Y, Zaharchuk G and Wintermark M (2019) Applications of Deep Learning to Neuro-Imaging Techniques. Front. Neurol. 10:869. doi: 10.3389/fneur.2019.00869

### INTRODUCTION

Artificial intelligence (AI) is a branch of computer science that encompasses machine learning, representation learning, and deep learning (1). A growing number of clinical applications based on machine learning or deep learning and pertaining to radiology have been proposed in radiology for classification, risk assessment, segmentation tasks, diagnosis, prognosis, and even prediction of therapy responses (2–10). Machine learning and deep learning have also been extensively used for brain image analysis to devise imaging-based diagnostic and classification systems of strokes, certain psychiatric disorders, epilepsy, neurodegenerative disorders, and demyelinating diseases (11–17).

Recently, due to the optimization of algorithms, the improved computational hardware, and access to large amount of imaging data, deep learning has demonstrated indisputable superiority over the classic machine learning framework. Deep learning is a class of machine learning that uses artificial neural network architectures that bear resemblance to the structure of human cognitive functions (**Figure 1**). It is a type of representation learning in which the algorithm learns a composition of features that reflect a hierarchy of structures in the data (18). Convolutional neural networks (CNN) and recurrent neural networks (RNN) are different types of deep learning methods using artificial neural networks (ANN).

AI can be applied to a wide range of tasks faced by radiologists (**Figure 2**). Most initial deep learning applications in neuroradiology have focused on the "downstream" side: using computer vision techniques for detection and segmentation of anatomical structures and the detection of lesions, such as hemorrhage, stroke, lacunes, microbleeds, metastases, aneurysms, primary brain tumors, and white matter hyperintensities (6, 9, 15, 19). On the "upstream" side, we have just begun to realize that there are other innovative applications of AI in various technical aspects of medical imaging, particularly applied to the acquisition of images. A variety of methods for image generation and image enhancement using deep learning have recently been proposed, ranging from removing image artifacts, normalizing/harmonizing images, improving image quality, lowering radiation and contrast dose, and shortening the duration of imaging studies (8, 9, 15).

**123**

As RNNs are commonly utilized for speech and language tasks, the deep learning algorithms most applicable to radiology are CNNs, which can be efficiently applied to image segmentation and classification. Instead of using more than billions of weights to implement the full connections, CNNs can mimics mathematic operation of convolution, using convolutional and pooling layers (**Figure 1**) and significantly reduce the number of weights. CNNs can also allow for spatial invariance. For different convolutional layers, multiple kernels can be trained and then learn many location-invariant features. Since important features can be automatically learned, information extraction from images in advance of the learning process is not necessary. Therefore, CNNs are relatively easy to apply in clinical practice.

There are many challenges related to the acquisition and post-processing of neuroimages, including the risks of radiation exposure and contrast agent exposure, prolonged acquisition time, and image resolution. In addition, to expert parameter tuning of scanners always required to optimize reconstruction performance, especially in the presence of sensor non-idealities and noise (20). Deep learning has the opportunity to have a significant impact on such issues and challenges, with fewer ethical dilemmas and medical legal risks compared to applications for diagnosis and treatment decision making (21). Finally, these deep learning approaches will make imaging much more accessible, from many perspectives, including cost, patient safety, and patient satisfaction.

Published deep learning studies focused on improving medical imaging techniques are just beginning to enter the medical literature. A Pubmed search on computeraided diagnosis in radiology, machine learning, and deep learning for the year 2018 yielded more than 5,000 articles. The number of publications addressing deep learning as applied to medical imaging techniques is a small fraction of this number. Although many studies are not focused on neuroimaging, their techniques can often be adapted for neuroimaging. This article will address this topic and will seek to present an overview of deep learning applied to neuroimaging techniques.

Copyright American Journal of Neuroradiology.

FIGURE 2 | Imaging value chain. While most AI applications have focused on the downstream (or right) side of this pathway, such the use of AI to detect and classify lesions on imaging studies, it is likely that there will be earlier adoption for the tasks on the upstream (or left) side, where most of the costs of imaging are concentrated.

# USING DEEP LEARNING TO REDUCE THE RISK ASSOCIATED WITH IMAGE ACQUISITION

There are many risks associated with different image acquisitions, such as ionizing radiation exposure and side effect of contrast agents. Deep learning based optimizing acquisition parameters is crucial to achieve diagnostically acceptable image quality at the lowest possible radiation dose and/or contrast agent dose.

## MRI

Gadolinium-based contrast agents (GBCAs) have become indispensable in routine MR imaging. Though considered safe, CBCAs were linked with nephrogenic systemic fibrosis, which is a serious, debilitating, and sometimes life-threatening condition. There is ongoing discussion regarding the documented deposition of gadolinium contrast agents in body tissues including the brain, especially for those patients who need repeated contrast administration (22). Recent publications have reported the gadolinium deposition in the brain tissue, most notably in the dentate nuclei and globus pallidus (23, 24). This deposition can probably be minimized by limiting the dose of gadolinium used (25). Unfortunately, low-dose contrastenhanced MRI is typically of insufficient diagnostic image quality. Gong et al. (26) implemented a deep learning model based on an encoder-decoder CNN to obtain diagnostic quality contrast-enhanced MRI with low-dose gadolinium contrast. In this study 60 patients with brain abnormalities received 10% lowdose preload (0.01 mmol/kg) of gadobenate dimeglumine, before perfusion MR imaging with full contrast dosage (0.1 mmol/kg). Pre-contrast MRI and low-dose post-contrast MRI of training set were introduced as inputs, and full dose post-contrast MRI as Ground-truth. The contrast uptake in the low-dose CE-MRI is noisy, but does include contrast information. Through the training, the network learned the guided denoising of the noisy contrast uptake extracted from the difference signal between low-dose and zero-dose MRIs, and then combine them to synthesize a full-dose CE-MRI. The results demonstrated that the deep learning algorithm was able to extract diagnostic quality images with gadolinium doses 10-fold lower than those typically used (**Figure 3**).

# CT

Computed Tomography (CT) techniques are widely used in clinical practice and involve a radiation risk. For instance, the radiation dose associated with a head CT is the same as 200 chest X-rays, or the amount most people would be exposed to from natural sources over 7 years. CT acquisition parameters can be adjusted to reduce the radiation dose, including reducing kilovoltage peak (kVp), milliampere-seconds (mAs), gantry rotation time, and increasing acquisition pitch. However, all these approaches also reduce image quality. Since an insufficient number of photons in the projection domain can lead to excessive quantum noise, the balance between image quality and radiation dose is always a trade-off.

Various image denoising approaches for CT techniques have been developed. Iterative reconstruction has been used, but sparsely, in part due to significant computational costs, time

delays between acquisition and reconstruction, and a suboptimal "waxy" appearance of the augmented images (27, 28). Traditional image processing methods to remove image noise are also limited, because CT data is subject to both non-stationary and non-Gaussian noise processes. Novel denoising algorithms based on deep learning have been studied intensively and showed impressive potential (29). For example, Xie et al. (30) used a deep learning method based on a GoogLeNet architecture to remove streak artifacts due to missing projections in sparseview CT reconstruction. The artifacts from low dose CT imaging were studied by residual learning, and then subtracted from the sparse reconstructed image to recover a better image. These intensively reconstructed images are comparable to the fullview projection reconstructed images. Chen et al. (28, 31) applied a residual encoder-decoder CNN, which incorporated a deconvolution network with shortcut ("bypass") connections into a CNN model, to reduce the noise level of CT images. The model learned a feature mapping from low- to normal-dose images. After the training, it achieved a competitive performance in both qualitative and quantitative aspects, while compared with other denoising methods. Kang (27) applied a CNN model using directional wavelets for low-dose CT reconstruction. Compared to model-based iterative reconstruction methods, this algorithm can remove complex noise patterns from CT images with greater denoising power and faster reconstruction time. Nishio et al. (32) trained auto-encoder CNN for pairs of standard-dose (300 mA) CT images and ultra-low-dose (10 mA) CT images, and then used the trained algorithm for patch-based image denoising of ultra-low-dose CT images. The study demonstrated the advantages of this method over block-matching 3D (BM3D) filtering for streak artifacts and other types of noise. Many other deep learning-based approaches have been proposed in radiation-restricted applications, such as adversarially trained networks, sharpness detection network, 3D dictionary learning, and discriminative prior-prior image constrained compressed sensing (33–36).

Reconstruction algorithms to denoise the output low-quality images or remove artifacts have been studied intensively (27, 28, 30–32). Gupta et al. (37) implemented a relaxed version of projected gradient descent with a CNN for sparse-view CT reconstruction. There is a significant improvement over total variation-based regularization and dictionary learning for both noiseless and noisy measurements. This framework can also be used for super-resolution, accelerated MRI, or deconvolution, etc. Yi et al. used adversarially trained network and sharp detection network to achieve sharpness-aware low-dose CT denoising (34).

Since matched low- and routine-dose CT image pairs are difficult to obtain in multiphase CT, Kang et al. (38) proposed a deep learning framework based on unsupervised learning technique to solve this problem. They applied a cycle-consistent adversarial denoising network to learn the mapping between lowand high-dose cardiac phases. Their network did not introduce artificial features in the output images.

### Sparse-Data CT

The reconstruction of Sparse-data CT always compromises structural details and suffers from notorious blocky artifacts. Chen et al. (39) implemented a Learned experts' assessmentbased reconstruction network (LEARN) for sparse-data CT. The network was evaluated with Mayo Clinic's low-dose challenge image data set and was proved more effectively than other methods in terms of artifact reduction, feature preservation, and computational speed.

# PET

Radiation exposure is a common concern in PET imaging. To minimize this potential risk, efforts have been made to reduce the amount of radio-tracer usage in PET imaging. However, low-dose PET is inherently noisy and has poor image quality. Xiang et al. combined 4-fold reduced time duration 18F-fluorodeoxyglucose (FDG) PET images and co-registered T1-weighted MRI images to reconstruct standard dose PET (40). Since PET image quality is to a first degree linear with true coincidence events recorded by the camera, such a method could also be applied to reduced dose PET. Kaplan and Zhu (41) introduced a deep learning model consisting an estimator network and a generative adversarial network (GAN). After training with simulated 10x lower dose PET data, the networks reconstructed standard dose images, while preserving edge, structural, and textural details.

Using a simultaneous PET/MRI scanner, Xu et al. (42) proposed an encoder-decoder residual deep network with concatenate skip connections to reconstruct high quality brain FDG PET images in patients with glioblastoma multiforme using only 0.5% of normal dose of radioactive tracer. To take advantage of the higher contrast and resolution of the MR images, they also included T1-weighted and T2-FLAIR weighted images as inputs to the model. Furthermore, they employed a "2.5D" model in which adjacent slice information is used to improve the prediction of a central slice. These modifications significantly reduced noise, while robustly preserving resolution and detailed structures with comparable quality to normal-dose PET images.

These general principles were also applied by Chen et al. to simulated 1% dose 18F-florbetaben PET imaging (43). This amyloid tracer is used clinically in the setting of dementia of unknown origin. A "positive" amyloid study is compatible with the diagnosis of Alzheimer's disease, while a negative study essentially rules out the diagnosis (44, 45). Again, simultaneous PET/MRI was used to acquire co-registered contemporaneous T1-weighted and T2-FLAIR MR images, which were combined as input along with the 1% undersampled PET image. They showed the crucial benefit of including MR images in terms of retaining spatial resolution, which is critical for assessing amyloid scans. They found that clinical readers evaluating the synthesized full dose images did so with similar accuracy to their own intra-reader reproducibility. More recently, the same group has demonstrated that the trained model can be applied to true (i.e., not simulated) ultra-low dose diagnostic PET/MR images (**Figure 4**).

## ACCELERATE IMAGING ACQUISITION AND RECONSTRUCT UNDER-SAMPLED K-SPACE

Image acquisition can be time-consuming. Reducing raw data samples or subsample k-space data can speed the acquisition, but result in suboptimal images. Deep learning based reconstruction methods can output good images from under-sampled datasets.

Compared to most other imaging modalities, MRI acquisition is substantially slower. The longer acquisition time limits the utility of MRI in emergency settings and often results in more motion artifact. It also contributes to its high cost. Acquisition time can be reduced by simply reducing the number of raw data samples. However, conventional reconstruction methods for these sparse data often produce suboptimal images. Newer reconstruction methods deploying deep learning have the ability to produce images with good quality from these under-sampled data acquired with shorter acquisition times (46). This approach has been applied in Diffusion Kurtosis Imaging (DKI) and Neurite Orientation Dispersion and Density Imaging (NODDI). DKI and NODDI are advanced diffusion sequences that can characterize tissue microstructure but require long acquisition time to obtain the required data points. Using a combination of q-Space deep learning and of simultaneous multi-slice imaging, Golkov et al. (47) were able to reconstruct DKI from only 12 data points and NODDI from only 8 data points, achieving an unprecedented 36-fold scan time reduction for quantitative diffusion MRI. These results suggest that there is considerable amount of information buried within the limited number of data points that can be retrieved with deep learning methods.

Another way to reduce acquisition time is to subsample kspace data. However, naive undersampling of k-space will cause aliasing artifact once the under-sampling rate exceeds the Nyquist conditions. Hyun et al. (48) trained a deep learning network, using pairs of subsampled and fully sampled k-space data as inputs and outputs respectively, to reconstruct images from subsampled data. They reinforced the subsampled k-space data with a few low-frequency k-space data to improve image contrast. Their network was able to generate diagnostic quality images from sampling only 29% of k-space.

Lee et al. (49) investigated deep residual networks to remove global artifacts from under-sampled k-space data. Deep residual networks are a special type of network that allows stacking of multiple layers to create a very deep network without degrading the accuracy of training. Compared to non-AI based fastacquisition techniques such as compressed sensing MRI (which randomly sub-samples k-space) and parallel MRI (which uses multiple receiver coils), Lee's technique achieved better artifact reduction and use much shorter computation time.

Deep learning techniques for acceleration and reconstruction are not limited to static imaging, but are also applicable for dynamic imaging, such as cardiac MRI. Due to inherent redundancy within adjacent slices and repeated cycles in dynamic imaging, the combination of under-sampling and using Neural Networks for reconstruction seem to be the perfect solution. Schelmper's (50) trained CNN to learn the redundancies and the spatio-temporal correlations from 2D cardiac MR images. Their CNN outperformed traditional carefully handcrafted algorithms in terms of both reconstruction quality and speed. Similarly, Majumdar (51) address the problem of real-time dynamic MRI reconstruction by using a stacked denoising autoencoder. They produced superior images in shorter time, when compared to CS based technique and Kalman filtering techniques.

Hammernik et al. (52) introduced a variational network for accelerated Parallel Imaging-based MRI reconstruction. The

reconstruction time was 193 ms on a single graphics card, and the MR images preserved the natural appearance as well as pathologies that were not included in the training data set. Chen et al. (53) also developed a deep learning reconstruction approach based on a variational network to improve the reconstruction speed and quality of highly undersampled variable-density single-shot fast spin-echo imaging. This approach enables reconstruction speeds of ∼0.2 s per section, allowing a realtime image reconstruction for practical clinical deployment. This study showed improved image quality with higher perceived signal-to-noise ratio and improved sharpness, when compared with conventional parallel imaging and compressed sensing reconstruction. Yang et al. (54) proposed a deep architecture based on Alternating Direction Method of Multipliers algorithm (ADMM-Net) to optimize a compressed sensing-based MRI model. The results suggested high reconstruction accuracy with fast computational speed.

Several studies also used generative adversarial networks to model distributions (low-dimensional manifolds) and generating natural images (high-dimensional data) (35, 55). Mardani et al. (56) proposed a compressed sensing framework using generative adversarial networks (GAN) to model the low-dimensional manifold of high-quality MRI. This is combined with a compressed sensing framework, a method known as GANCS. It offers reconstruction times of under a few milliseconds and higher quality images with improved fine texture based on multiple reader studies.

# ARTIFACTS REDUCTION

Image denoising is an important pre-processing step in medical image analysis, especially in low-dose techniques. Much research has been conducted on the subject of computer algorithms for image denoising for several decades, with varying success. Many attempts based on machine learning (57) or deep learning (58, 59) have been successfully implemented for denoising of medical images.

Standard reconstruction approaches involve approximating the inverse function with multiple ad hoc stages in a signal processing chain. They depend on the details of each acquisition strategy, and requires parameter tuning to optimize image quality. Zhu et al. (20) implemented a unified framework system called AUTOMAP, using a fully-connected deep neural network to reconstruct a variety of MRI acquisition strategies. This method is agnostic to the exact sampling strategy used, being trained on pairs of sensor data and ground truth images. They showed good performance for a wide range of k-space sampling methods, including Cartesian, spiral, and radial image acquisitions. The trained model also showed superior immunity to noise and reconstruction artifacts compared with conventional handcrafted methods. Manjón and Coupe (59) used two-stage strategy with deep learning for noise reduction. The first stage is to remove the noise using a CNN without estimation of local noise level present in the images. Then the filtered image is used as a guide image within a rotationally invariant non-local means filter. This approach showed competitive results for all the studied MRI acquisitions.

### Low Signal-To-Noise Ratio

MR images often suffers from low signal-to-noise ratio, such as DWI and 3D MR images. Jiang et al. (60) applied multi-channel feed-forward denoising CNNs, and Ran et al. (61) applied residual Encoder-Decoder wasserstein GAN, respectively, to restore the noise-free 3D MR images from the noisy ones.

Another MRI acquisition suffering from an inherently low-signal-to-noise ratio is arterial spin labeling (ASL) perfusion imaging. ASL has been used increasingly in neuroimaging because of its non-invasive and repeatable advantages in quantification and labeling. Repeated measurements of control/spin-labeled paired can lead to a fair image quality, but with the risk of motion artifacts. Ultas et al. (62) followed a mixed modeling approach including incorporting a Buxton kinetic model for CBF estimation, and training a deep fully CNN to learn a mapping from noisy image and its subtraction from the clean images. This approach produced high quality ASL images by denoising images without estimating its noise level. Due to a lower number of subtracted control/label pairs, this method also reduced ASL scan and reconstruction times, which makes ASL even more applicable in clinical protocols. Similarly, Kim et al. demonstrated image quality improvement using pseudocontinous ASL using data with 2 signal averages to predict images acquired with 6 signal averages, a roughly 3-fold speedup in imaging time (63). They also demonstrated that it was possible to reconstruct Hadamard-encoded ASL imaging from a subset of the reconstructed post-label delay images (though this does not allow for any speed-up in image acquisition). Owen et al. used a convolutional joint filter to exploit spatio-temporal properties of the ASL signal. This filter could reduce artifacts and improve the peak signal-to-noise ratio of ASL by up to 50% (64). Finally, Gong et al. demonstrated the benefits of including multi-contrast approaches (i.e., proton-density images along with ASL difference images) with multi-lateral guided filters and deep networks to boost the SNR and resolution of ASL (65). They also showed that the network could be trained with a relatively small number of studies and that it generalized to stroke patients (**Figure 5**).

# Spurious Noise

Proton MR spectroscopic imaging can measure endogenous metabolite concentration in vivo. The Cho/NAA ratio has been used to characterize brain tumors, such as glioblastoma. One challenge is the poor spectral quality, because of the artifacts caused by magnetic field inhomogeneities, subject movement, and improper water or lipid suppression. Gurbani et al. (66) applied a tiled CNN tuned by Bayesian optimization technique to analyze frequency-domain spectra to detect artifacts. This CNN algorithm achieved high sensitivity and specificity with an AUC of 0.951, while compared with the consensus decision of MRS experts. One particular type of MRS artifact is ghost or spurious echo artifact, due to insufficient spoiling gradient power. Kyathanahally et al. (67) implemented multiple deep learning algorithms, including fully connected neural networks, deep CNN, and stacked what-where auto encoders, to detect and correct spurious echo signals. After training on a large dataset with and without spurious echoes, the accuracy of the algorithm was almost 100%.

# Motion Artifact

MRI is susceptible to image artifacts, including motion artifacts due to the relatively long acquisition time. Küstner et al. (68) proposed a non-reference approach to automatically detect the presence of motion artifacts on MRI images. A CNN classifier was trained to assess the motion artifacts on a per-patch basis, and then used to localize and quantify the motion artifacts on a test data set. The accuracy of motion detection reached 97/100% in the head and 75/100% in the abdomen. There are several other studies on the detection or reducing of motion artifacts (69–71). Automating the process of motion detection can lead to more efficient scanner use, where corrupted images are re-acquired without relying on the subjective judgement of technologists.

# Metal Artifact

Artifacts resulting from metallic objects have been a persistent problem in computed tomography (CT) images over the last four decades. Gjesteby et al. (72) combined a CNN with the NMAR algorithm to reduce metal streaks in critical image regions. The strategy is able to map metal-corrupted images to artifact-free monoenergetic images.

# Crosstalk Noise

Attenuation correction is a critical procedure in PET imaging for accurate quantification of radiotracer distribution. For PET/CT, the attenuation coefficients (µ) are derived from the CT Hounsfield units from the CT portion of the examination. For PET/MRI, attenuation coefficient (µ) has been estimated from segmentation- and atlas-based algorithms. Maximum-likelihood reconstruction of activity and attenuation (MLAA) is a new method for generating activity images. It can produce attenuation coefficients simultaneously from emission data only, without the need of a concurrent CT or MRI. However, MLAA suffers from crosstalk artifacts. Hwang et al. (73) tested three different CNN architectures, such as convolutional autoencoder (CAE), U-net, and hybrid of CAE to mitigate the crosstalk problem in the MLAA reconstruction. Their CNNs generated less noisy and more uniform µ-maps. The CNNs also better resolved the air cavities, bones, and even the crosstalk problem.

Other studies have used deep learning to create CT-like images from MRI, often but not always for the purposes of PET/MRI attenuation correction. Nie et al. (74) applied an autocontext model to implement a context-aware deep convolutional GAN. It can generate a target image from a source image, demonstrating its use in predicting head CT images from T1 weighted MR images. This CT could be used for radiation planning or attenuation correction. Han (75) proposed a deep

(TGV). Such methods could speed up MRI acquisition, enabling more functional imaging and perhaps reducing the cost of scanning.

CNN with 27 convolutional layers interleaved with pooling and unpooling layers. Similar to Nie et al., the network was trained to learn a direct end-to-end mapping from MR images to their corresponding CTs. This method produced accurate synthetic CT results in near real time (9 s) from conventional, singlesequence MR images. Other deep learning networks, such as deep embedding CNN by Xiang et al. (76), Dixon-VIBE deep learning by Torrado-Carvajal et al. (77), GAN with two synthesis CNNs and two discriminator CNNs by Wolterink et al. (78), as well as deep CNN based on U-net architecture by Leynes et al. (79) and Roy et al. (80), were also proposed to generate pseudo CT from MRI.

Liu et al tried to train a network to transform T1-weighted head images into "pseudo-CT" images, which could be used for attenuate calculations (81). The errors in PET SUV could be reduced to less than 1% for most areas of the brain, about a 5-fold improvement over existing techniques such as atlas-based and 2 point Dixon methods. More recently, the same group has shown that it is possible to take non-attenuation correction PET brain images and using attenuation corrected images as the ground truth, to directly predict one from the other, without the need to calculate an attenuation map (82). This latter method could enable the development of new PET scanners that do not require either CT or MR imaging to be acquired, and which might be cheaper to site and operate.

### Random Noise

Medical fluoroscopy video is also sensitive to noise. Angiography is one medical procedure using live video, and the video quality is highly important. Speed is the main limitation of conventional denoising algorithms such as BM3D. Praneeth Sadda et al. (83) applied a deep neural network to remove Gaussian noise, speckle noise, salt and pepper noise from fluoroscopy images. The final output live video could meet and even exceed the efficacy of BM3D with a 20-fold speedup.

# SYNTHETIC IMAGE PRODUCTION

Each imaging modality (X-ray, CT, MRI, ultrasound) as well as different MR sequences have different contrast and noise mechanisms and hence captures different characteristics of the underlying anatomy. The intensity transformation between any two modalities/sequences is highly non-linear. For example, Vemulapalli et al. (84) used a deep network to predict T1 images from T2 images. With deep learning, medical image synthesis can produce images of a desired modality without preforming an actual scan, such as creating CT images from MRI data. This can be of benefit because radiation can be avoided.

Ben-Cohen et al. (85) explored the use of full CNN and conditional GAN to reconstruct PET images from CT images. The deep learning system was tested for detection of malignant tumors in the live region. The results suggested a true positive ratio of 92.3% (24/26) and false positive ratio of 25% (2/8). This is surprising because no metabolic activity is expected to be present on CT images. It must be assumed that the CT features somehow contain information about tumor metabolism. In a reverse strategy, Choi and Lee (86) generated structural MR images from amyloid PET images using generative adversarial networks. Finally, Li et al. (87) used a 3D CNN architecture to predict missing PET data from MRI, using the ADNI study, and found it to be a better way of estimating missing data than currently existing methods.

### High-Field MRI

More recently, AI based methods, such as deep CNN's, can take a low-resolution image as the input and then output a highresolution image (88), with three operations, "patch extraction and representation," "non-linear mapping," and "reconstruction" (89). Higher (or super-) resolution MRI can be implemented using MRI scanners with higher magnetic field, such as advanced 7-T MRI scanners, which involves much higher instrumentation and operational costs. As an alternative, many studies have attempted to achieve super-resolution MRI images from lowresolution MRI images. Bahrami et al. (90) trained a deep learning architecture based CNN, inputting the appearance and anatomical features of 3T MRI images and outputting as the corresponding 7T MRI patch to reconstruct 7T-like MRI images. Lyu et al. (91) adapted two neural networks based on deep learning, conveying path-based convolutional encoderdecoder with VGG (GAN-CPCE) and GAN constrained by the identical, residual, and cycle learning ensemble (GAN-CIRCLE), for super-resolution MRI from low-resolution MRI. Both neural networks had a 2-fold resolution improvement. Chaudhari et al. (92) implemented a 3-D CNN entitled DeepResolve to learn residual-based transformations between high-resolution and lower-resolution thick-slice images of musculoskeletal MRI. This algorithm can maintain the resolution as diagnostic image quality with a 3-fold down-sampling. Similar methods have recently been applied to T1-weighted brain imaging, which requires a long acquisition time to obtain adequate resolution for cortical thickness mapping (**Figure 6**).

### Synthetic FLAIR

Synthetic MRI imaging has become more and more clinically feasible, but synthetic FLAIR images are usually of lower quality than conventional FLAIR images (93). Using conventional FLAIR images as target, Hagiwara et al. (94) applied a conditional GAN to generate improved FLAIR images from raw synthetic MRI imaging data. This work created improved synthetic FLAIR imaging with reduced swelling artifacts and granular artifacts in the CSF, while preserving lesion contrast. More recently, Wang et al. showed that improvements in image quality for all synthetic MR sequences could be obtained using a single model for multicontrast synthesis along with a GAN discriminator, which was dubbed "OneforAll" (95). This offered superior performance to a standard U-net architecture trained on only one image contrast at a time. Readers scored equivalent image quality between the deep learning-based images and the conventional MR sequences for all except proton-density images. The deep learning based T2 FLAIR images were superior to the conventional images, due to the inherent noise suppression aspects of the training process.

# IMAGE REGISTRATION

Deformable image registration is critical in clinical studies. Image registration is necessary to establish accurate anatomical correspondences. Intensity-based feature selection methods are widely used in medical image registration, but do not guarantee the exact correspondence of anatomic sites. Hand-engineered features, such as Gabor filters and geometric moment invariants, are also widely used, but do not work well for all types of image data. Recently, many AI-based methods have been used to perform image registration. Deep learning may be more promising when compared to other learning-based methods, because it does not require prior knowledge or hand-crafted features. It uses a hierarchical deep architecture to infer complex non-linear relationships quickly and efficiently (96).

Wu et al. (96) applied a convolutional stacked auto-encoder to identify compact and highly discriminative features in observed imaging data. They used a stacked two-layer CNN to directly learn the hierarchical basis filters from a number of image patches on the MR brain images. Then the coefficients can be applied as the morphological signature for correspondence detection to achieve promising registration results (97). Registration for 2D/3D image is one of the keys to enable image-guided procedures, including advanced image-guided radiation therapy. Slow computation and small capture range, which is defined as the distance at which 10% of the registrations fail, are the two major limitations of existing intensity-based 2D/3D registration approaches. Miao et al. (98) proposed a CNN regression approach, referred to as Pose Estimation via Hierarchical Learning (PEHL), to achieve real-time 2D/3D registration with large capture range and high accuracy. Their results showed an increased capture range of 99–306% and a success rate of 5–27.8%. The running time was ∼0.1 s, about one tenth of the time consumption other intensity-based methods have. This CNN regression approach achieved significantly higher computational efficiency such that it is capable of real-time 2D/3D registration. Neylon et al. (99) presented a method based on deep neural network for automated quantification of deformable image registration. This neural network was able to quantify deformable image registration error to within a single voxel for 95% of the sub-volumes examined. Other studies also include fast predictive image registration with deep encoderdecoder network based on a Large Deformation Diffeomorphic Metric Mapping model (100).

# QUALITY ANALYSIS

Quality control is crucial for accurate medical imaging measurement. However, it is a time-consuming process. Deep learning-based automatic assessment may be more objective and efficient. Lee et al. (101) applied a CNN to predict whether CT scans meet the minimal image quality threshold for diagnosis. Due to the relatively small number of cases, this deep learning network had a fair performance with an accuracy of 0.76 and an AUC of 0.78. Wu et al. (102) designed a computerized fetal ultrasound quality assessment (FUIQA) scheme with two deep CNNs (L-CNN and C-CNN). The L-CNN finds the region of interest, while the C-CNN evaluates the image quality.

## CHALLENGES OF DEEP LEARNING APPLIED TO NEUROIMAGING TECHNIQUES

In summary, deep learning is a machine learning method based on artificial neural networks (ANN), and encompasses supervised, unsupervised, and semi-supervised learning. Despite the promises made by many studies, reliable application of deep

learning for neuroimaging still remains in its infancy and many challenges remain.

First of them is overfitting. Training a complex classifier with a small dataset always carries the risk of overfitting. Deep learning models tend to fit the data exceptionally well, but it doesn't mean that they generalize well. There are many studies that used different strategies to reduce overfitting, including regularization (103), early stopping (104), and drop out (105). While overfitting can be evaluated by performance of the algorithm on a separate test data set, the algorithm may not perform well on similar images acquired in different centers, on different scanners, or with different patient demographics. Larger data sets from different centers are typically acquired in different ways using different scanners and protocols, with subtly different image features, leading to poor performance (21). According to those, data augmentation without standard criteria cannot appropriately address issues encountered with small datasets. Overcoming this problem, known as "brittle AI," is an important area of research if these methods are to be used widely. Deep learning is also an intensely data hungry technology. It requires a very large number of well labeled examples to achieve accurate classification and validate its performance for clinical implementation. Because upstream applications such as image quality improvement are essentially learning from many predictions in each image, this means that the requirements for large datasets are not as severe as for classification algorithms (where only one learning data point is available per person). Nonetheless, building large, public, labeled medical image datasets is important, while privacy concerns, costs, assessment of ground truth, and the accuracy of the

### REFERENCES

1. Jiang F, Jiang Y, Zhi H, Dong Y, Li H, Ma S, et al. Artificial intelligence in healthcare: past, present and future. Stroke Vasc Neurol. (2017) 2:230–43. doi: 10.1136/svn-2017-000101

labels remain stumbling blocks (18). One advantage of image acquisition applications is that the data is in some sense already labeled, with the fully sampled or high dose images playing the role of labels in classification tasks. Besides the ethical and legal challenges, the difficulty of physiologically or mechanistically interpreting the results of deep learning are unsettling to some. Deep networks are "black boxes" where data is input and an output prediction, whether classification or image, is produced (106). All deep learning algorithms operate in higher dimensions than what can be directly visualized by the human mind, which has been coined as "The Mythos of Model Interpretability" (107). Some estimates of the network uncertainly in prediction would be helpful to better interpret the images produced.

### CONCLUSION

Although deep learning techniques in medical imaging are still in their initial stages, they have been enthusiastically applied to imaging techniques with many inspired advancements. Deep learning algorithms have revolutionized computer vision research and driven advances in the analysis of radiologic images. Upstream applications to image quality and value improvement are just beginning to enter into the consciousness of radiologists, and will have a big impact on making imaging faster, safer, and more accessible for our patients.

# AUTHOR CONTRIBUTIONS

GuZ: drafting the review. BJ, LT, YX, and GrZ: revising the review. MW: conception and design and revising the review.


International Conference on Medical Image Computing and Computer-Assisted Intervention. Nagoya: Springer (2013). p. 649–56.


**Conflict of Interest Statement:** The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

Copyright © 2019 Zhu, Jiang, Tong, Xie, Zaharchuk and Wintermark. This is an open-access article distributed under the terms of the Creative Commons Attribution License (CC BY). The use, distribution or reproduction in other forums is permitted, provided the original author(s) and the copyright owner(s) are credited and that the original publication in this journal is cited, in accordance with accepted academic practice. No use, distribution or reproduction is permitted which does not comply with these terms.

# History of Hypertension Is Associated With MR Hypoperfusion in Chinese Inpatients With DWI-Negative TIA

Yue Wang1,2†, Huazheng Liang1,2†, Yu Luo2,3, Yuan Zhou1,2, Lingjing Jin<sup>4</sup> , Shaoshi Wang1,2 \* and Yong Bi 1,2 \*

*<sup>1</sup> Department of Neurology, Shanghai Fourth People's Hospital Affiliated to Tongji University School of Medicine, Shanghai, China, <sup>2</sup> Department of Neurology, Translational Research Institute of Brain and Brain-Like Intelligence, Shanghai Fourth People's Hospital Affiliated to Tongji University School of Medicine, Shanghai, China, <sup>3</sup> Department of Radiology, Shanghai Fourth People's Hospital Affiliated to Tongji University School of Medicine, Shanghai, China, <sup>4</sup> Department of Neurology, Tongji Hospital, Tongji University, Shanghai, China*

### Edited by:

*Hongyu An, Washington University in St. Louis, United States*

### Reviewed by:

*Mohamed Al-Khaled, DOC Medical Center, Qatar Jue Zhang, Peking University, China*

### \*Correspondence:

*Yong Bi drbiyong@126.com Shaoshi Wang wangshaoshi@126.com*

*†These authors have contributed equally to this work*

### Specialty section:

*This article was submitted to Applied Neuroimaging, a section of the journal Frontiers in Neurology*

Received: *13 March 2019* Accepted: *26 July 2019* Published: *14 August 2019*

### Citation:

*Wang Y, Liang H, Luo Y, Zhou Y, Jin L, Wang S and Bi Y (2019) History of Hypertension Is Associated With MR Hypoperfusion in Chinese Inpatients With DWI-Negative TIA. Front. Neurol. 10:867. doi: 10.3389/fneur.2019.00867* Objectives: The present study aimed to examine the prevalence of and risk factors for magnetic resonance (MR) perfusion abnormality in a Chinese population with transient ischemic attack (TIA) and normal diffusion-weighted imaging (DWI) findings.

Methods: Patients with TIA admitted to our stroke center between January 2015 and October 2017 were recruited to the present study. MRI, including both DWI and perfusion-weighted imaging (PWI), was performed within 7 days of symptom onset. Time to maximum of the residue function (Tmax) maps were evaluated using the RAPID software (Ischemaview USA, Version 4.9) to determine hypoperfusion. Multivariate analysis was used to assess perfusion findings, clinical variables, medical history, cardio-metabolic, and the ABCD2 scores (age, blood pressure, clinical features, symptom duration, and diabetes).

Results: Fifty-nine patients met the inclusion criteria. The prevalence of MR perfusion Tmax ≥ 4 s ≥ 0 ml and ≥ 10 mL were 72.9% (43/59) and 42.4% (25/59), respectively. Multivariate analyses revealed that history of hypertension is an independent factor associated with MR perfusion abnormality (Tmax ≥ 4 s ≥ 10 mL) for Chinese patients with TIA (*P* = 0.033, adjusted OR = 4.11, 95% CI = 1.12–15.11). Proximal artery stenosis (>50%) tended to lead to a larger PW lesion on MRI (*p* = 0.067, adjusted OR = 3.60, 95% CI = 0.91–14.20).

Conclusion: Our results suggest that the prevalence of perfusion abnormality is high as assessed by RAPID using the parametric Tmax ≥ 4 s. History of hypertension is a strong predictor of focal perfusion abnormality as calculated by RAPID on Tmax map of TIA patients with negative DWI findings.

Keywords: DWI, PWI, transient ischemic attack, risk factors, hypertension

# INTRODUCTION

Transient ischemic attack (TIA) has been redefined as a transient episode of neurological dysfunction caused by focal brain, spinal cord, or retinal ischemia without evidence of acute infarction (1). According to a multicenter, community-based study, the population of TIA survivors at any given time in China is as large as 10–12 million (2). TIA is associated with high risk of early subsequent stroke up to 20% of patients (3). TIA has been evaluated as a major risk factor for future recurrent ischemic attacks, and emergent diagnosis of the cause is needed to ensure timely treatment and to dramatically reduce the risk of developing strokes (4–6).

Prognosis of TIA depends on not only its pathological basis, but also early identification of high-risk patients with TIA and timely treatment. Usually, TIA diagnosis relies primarily on the reported history. The ABCD2 prediction score (range 0–7, age, blood pressure, clinical symptoms, duration, and diabetes) was originally intended to aid non-specialists in community and emergency department settings to improve risk stratification of patients with transient neurological symptoms, and had little specificity between hospital-based neurologists (7). Therefore, the diagnosis of TIA based on symptoms alone is challenging (8). Moreover, agreement on the vascular origin of transient neurologic symptoms can be low, even among experienced neurologists (9, 10). Early evaluation using imaging techniques is essential for administering the proper medications to treat or prevent TIA and the consequent stroke, which will refine the clinical diagnosis of TIA.

Based on the current diagnostic criteria, TIA is defined as a condition in which transient episode of neurological dysfunction exists without lesions on DWI. However, imaging results of TIA patients show diverse pictures. For example, perfusion-weighted imaging (PWI) shows either positive or negative findings in DWI negative patients. It is estimated that 23–42% of patients with TIA who have a negative DWI show PWI positive lesions (11–15). Acute PWI abnormality is associated with recurrent attacks and even infarct progression (13, 15–18). Therefore, low perfusion may be one of the pathological mechanisms of TIA recurrence. However, little research has been done on the relationship between TIA with negative DWI and perfusion abnormality in Chinese populations. The aim of the present study, therefore, was to assess the prevalence of MR perfusion abnormality and its risk factors in Chinese patients with TIA and negative DWI.

### METHODS

### Subjects

We retrospectively identified patients with TIA admitted to our stroke center between January 2015 and October 2017. The inclusion criteria for this study: (a) patients presented with TIA and evaluated by a certified stroke neurologist at the time of admission and discharge, diagnosis of TIA was confirmed by two certified stroke neurologists; (b) MRI including both DWI and PWI within 7 days of symptom onset, and no DWI evidence of restricted diffusion; (c) Time to maximum of the residue function (Tmax) maps were assessed independently using the RAPID software (Ischemaview USA, Version 4.9). The exclusion criteria: (a) patients with TIA did not have perfusion status assessed, or had DWI showing a lesion; (b) Patients received revascularization therapy (thrombolysis/thrombectomy). Radiologists blinded to clinical information independently evaluated the presence of acute ischemic lesions detected on DWI/PWI. Demographic data, clinical variables, risk factors, ABCD2 scores, neurologic deficits, duration of TIA, number of TIA attacks, time between MRI and onset were documented for each patient. Ethical approval for this study (2018011) was obtained from Human Research Ethics Committee of Shanghai Fourth People's Hospital Affiliated to Tongji University School of Medicine.

### Imaging

MRI was performed using a 1.5-T Avanto scanner (Siemens, Erlangen, Germany). The imaging protocol included DWI, FLAIR, PWI, and MR angiography (MRA). Imaging parameters were listed below. The head coil is an-8-channel phased-array coil. Axial EPI-DWI: 19 slices, slice thickness = 5.5 mm; TR/TE, 3,600/102 ms; FOV = 230 mm<sup>2</sup> , b = 0 and 1,000 s/mm<sup>2</sup> ; EPI factor = 192; matrix = 192 × 192; bandwidth = 964 Hz/pixel. Axial FLAIR: 18 slices, slice thickness = 5.5 mm; TR/TE, 4,000/92 ms; FOV = 230 mm<sup>2</sup> ; TI = 1,532.6 ms; Matrix = 256 × 190; bandwidth = 190 Hz/Px; flip angle = 150◦ . Axial EPI-PWI: 19 slices, slice thickness = 5, 1.5 mm spacing; TR/TE, 1,590/32 ms; measurements = 50; FOV = 230 mm<sup>2</sup> ; matrix size = 128 × 128; band width = 1,346 Hz/pixel; flip angle = 90◦ . Gd-DTPA contrast agent (gadopentetate dimeglumine; Shanghai Pharmaceutical Corporation, Shanghai, China) was intravenously injected (0.2 mmol/kg body weight) at a rate of 4 mL/s after flushing with 30 ml saline. Time-of-flight MR angiography: slice thickness = 0.7 mm; TR/TE, 25/7 ms; FOV = 180 mm<sup>2</sup> ; Matrix = 241 × 256; Bandwidth = 100 Hz/PX; flip angle = 25◦ .

Based on the clinical manifestation of TIA patients, the ischemic lesion site was localized.

Estimates of the volume of hypoperfusion from MRI perfusion scans were performed using the RAPID software, which is an automated image post-processing system (19). We used RAPID in our trial to measure the volume of hypoperfusion (20). Lesion volumes of Tmax ≥ 4 s were used for determining perfusion deficits in TIA patients with negative DWI findings (13, 15).

### Statistical Analysis

Continuous variables were presented with mean ± standard deviation (SD) or median with interquartile range (IQR); categorical variables were summarized as percentages. The normality of distribution for continuous variables was checked with the one-sample Kolmogorov–Smirnov test. Baseline information of patients with and without MR perfusion abnormality was compared using the independent sample t-test or Mann-Whitney U-test for continuous variables and Pearson chi-square or Fisher's exact tests for categorical variables. Binary logistic regression was used to assess the independent association between perfusion abnormality and risk factors. Univariate binary logistic regression analysis was used to screen for possible risk factors using P < 0.1. We assessed odd ratios (OR) of two patterns of perfusion abnormality for categorical variables (MR perfusion Tmax ≥ 4 s < 10 ml as no abnormality, and Tmax ≥ 4 s ≥ 10 ml as abnormality) with MRI perfusion normality being used as the reference. Correlation between TIA patients clinical information and perfusion abnormality with respect to MRI perfusion was tested using the multiple logistic regression analysis modeling with the "Enter" method. The multivariate regression model included history of hypertension and stenosis (50%) with a univariate P < 0.1 as independent variables. Meanwhile, the ABCD2 score, which is known to be correlated with perfusion abnormality, was also included for further analysis though its P > 0.1.

All association data were expressed as OR with corresponding 95% confidence intervals (CI) and P-values. Two-tailed tests were used for all analyses, with the statistical significance level set at 0.05. The data were analyzed with SPSS (version 20.0) for Windows (SPSS Inc., Chicago, IL, USA).

# RESULTS

A total of 154 patients records were evaluated for probable TIA at the Stroke Center of Shanghai Fourth People's Hospital Affiliated to Tongji University School of Medicine between January 2015 and October 2017. Fifty nine patients (24 women, 35 men; age range, 49–86 years; mean, 69 years) met the inclusion criteria. Sixty three patients were excluded because perfusion weighted images were not available (n = 63) after a TIA. Another 12 patients were excluded because they were not given a discharge diagnosis of tissue-negative TIA. Eighteen patients had DWI positive lesions, and another two had inadequate information.

### Patient Baseline Characteristics

A total of 59 subjects, including 35 males and 24 females, were included in the study. The median age of patients was 69 [interquartile range (IQR): 63–78]. Median (IQR) ABCD2 score was 4 (2–4). Baseline perfusion scans were performed after a median (IQR) delay of 5 (4–9) days from symptom onset or five (IQR 3–8) days from last attack. The median (IQR) symptom duration was 15 (5–60) min and the median frequency of TIA attacks at baseline was one (IQR 1–2). The average total cholesterol of patients was 4.24 ± 1.15 mmol/L, ranging from 1.94 to 8 mmol/L. The mean low-density lipoprotein (LDL) cholesterol level was 2.22 ± 0.84 mmol/L, ranging from 0.65 to 4.42 mmol/L. The average fasting blood-glucose (FBG) was 5.65 ± 1.40 mmol/L, ranging from 4.4 to 12.2 mmol/L. A history of hypertension was present in 67.8% (40/59) of patients, diabetes mellitus in 27.1% (16/59), and atrial fibrillation in 3.4% (2/59), history of stroke in 28.8% (17/59), smoking in 30.5% (18/59), and anterior circulation symptoms in 54.2% (32/59) (**Table 1**).

## Comparison of Demographic and Clinical Variables Between Patients With and Without MR Perfusion Abnormality

The prevalence of MR perfusion Tmax ≥ 4 s > 0 mL and Tmax ≥ 4 s ≥10 mL was 72.9% (43/59) and 42.4% (25/59), respectively. **Figure 1** showed typical images of an 84 year old female whose DWI showed negative findings of strokes, but PWI showed a focal lesion on Tmax.

**Table 1** presented the socio-demographic characteristics and clinical risk factors associated with MRI perfusion abnormality. Comparisons of these variables between patients with and without PWI abnormalities (Tmax ≥ 4 s ≥ 10 mL) showed no significant difference in these variables between the two groups except in history of hypertension (χ <sup>2</sup> = 5.22; p = 0.022). Surprisingly, there was no significant difference in the baseline ABCD2 score between these two groups, ABCD2 score has a strong predictive value of early neurological deterioration. Patients with atrial fibrillation tended to have a larger volume of lesions on PW images (8% compared with 0% of patients with no PWI abnormalities, p = 0.094). Patients with focal perfusion abnormalities tended to show more severe stenosis of responsible vessels (p = 0.056).

## Prediction of MRI Perfusion Abnormality

In univariate binary logistic regression analysis, history of hypertension (p = 0.028, OR = 4.15, 95% CI = 1.17–14.69) was independently associated with MR perfusion deficit. Stenosis (50%) (p = 0.065, OR = 3.53, 95% CI = 0.93–13.47) and systolic blood pressure (sBP) (p = 0.198, OR = 1.02, 95% CI = 0.99–1.05) on admission tended to be related to perfusion abnormality after a TIA. ABCD2 score (p = 0.959, OR = 0.99, 95% CI = 0.70– 1.41) was not associated with perfusion abnormality (**Table 2**). Multivariate regression modeling was performed for predictors with p < 0.20.

The multivariate logistic regression of associations between history of hypertension, stenosis (50%), sBP, and MR perfusion abnormality was shown in **Table 3**. It was clear that patients with a history of hypertension had a significantly higher risk of PWI abnormality (Tmax ≥ 4 s ≥ 10 mL) after a TIA. After adjusting potential confounding factors (age, sex, ABCD2), the odds ratios were 3.89 (95% CI, 1.08–13.96, p = 0.037, model 1), 4.33 (95% CI, 1.20–15.65, p = 0.025, model 2), and 4.11 (95% CI, 1.12– 15.11, p = 0.033, model 3), respectively. Stenosis (50%) and sBP on admission were not independently associated with perfusion abnormality after adjusting potential confounders.

# DISCUSSION

To the best of our knowledge, this is the first report that presented the prevalence and clinical risk factors for MRI perfusion abnormality in TIA patients of a Chinese population. The prevalence of MR perfusion Tmax ≥ 4 s ≥ 10 mL was 42.4% (25/59). Meanwhile, we found that among Chinese patients with acute TIA, history of hypertension is an independent factor associated with MR perfusion abnormality (Tmax ≥ 4 s ≥ 10 mL).

### Prevalence of MR Perfusion Abnormality

Our study showed a 72.9% (43/59) prevalence of MR perfusion (Tmax ≥ 4 s > 0 mL) in patients with DWI-negative TIA and 42.4% (25/59) (Tmax ≥ 4 s ≥ 10 mL) had an acute focal PWI lesion without showing a DWI lesion, which is similar to the previous report in Canada which showed a prevalence



*Tmax* ≥ *4 s, Time to maximum of the residue function* ≥ *4 s; IQR, interquartile range; Ghb, hemoglobin; NEU%, neutrophil percentage; FBG, fasting blood-glucose; LDL, low-density lipoprotein cholesterol; ESR, erythrocyte sedimentation rate; sBP, systolic blood pressure; Days\_inhos, Days in hospital.*

*ABCD2: age 60 (1 point), SBP 140, or DBP 90 mm Hg (1 point), clinical features as unilateral weakness (2 points) or speech impairment without weakness (1 point), symptom duration 60 min (2 points), or 10–59 min (1 point), diabetes (1 point).*

*Significant difference when P* < *0.05.*

FIGURE 1 | Typical images of an 84-year old female with a history of hypertension who presented with right upper limb paresis three times within 4 h. It lasted 2 min each time. DWI showed negative findings of strokes (A), but Tmax showed focal hypoperfusion areas in the left frontal and parietal lobes (B, green areas).

of 42% (57/137) (13), but higher than the prevalence of 25% (16/64) in South Korea (15) and 23% (57/137) in the United States (14). There are a few possible reasons for the higher prevalence. Firstly, the variability of findings in these studies is likely due to the inconsistent definition of perfusion. A study reported that a regional PWI lesion was detected on time-to-peak (TTP) and Mean transit time (MTT) maps, which were produced by the standard software bound to the scanner (15). Another study showed that a focal perfusion abnormality was identified on either time to maximum of the residue function (Tmax) or Cerebral blood flow (CBF) maps (14). In the present study, Tmax ≥ 4 s was used to define the regional perfusion abnormality. Secondly, different algorithms used for discrete platforms might be responsible for the discrepancy. Focal perfusion abnormalities were evaluated independently by two observers in some studies (14, 15) or PWI source images were analyzed by a customized Matlab 7.4 (The Mathworks) software (13). However, in this present study we used RAPID to calculate the volume of perfusion. Therefore, the prevalence of MR perfusion abnormality in the present study is higher than that of previous reports. Thirdly, participants in the present study were all inpatients of our stroke center, who were more likely to have perfusion lesions than outpatients because their conditions were more serious.

In the present study, 25 of 59 patients had Tmax ≥ 4 s ≥ 10 mL. Tmax delay threshold 4 s seems to be optimal for early assessment of critically hypoperfused tissue (21). Tmax volume is a good predictor for clinical outcome in MCA occlusions (22). The threshold (Tmax ≥ 4 s) at a volume of 10 mL is optimal



*CI, confidence interval; OR, Odds ratio.*

*ABCD2: age 60 (1 point), SBP 140 or DBP 90 mm Hg (1 point), clinical features as unilateral weakness (2 points), or speech impairment without weakness (1 point), symptom duration 60 min (2 points), or 10–59 min (1 point), diabetes (1 point).*

for predicting infarct growth with the maximal sensitivity and specificity (13).

### Risk Factors Associated With MR Perfusion Abnormality

There are multiple possible clinical risk factors for MR perfusion abnormality in the context of TIA. 67.8% of the 59 TIA patients included in this study had a history of hypertension, which is similar to that of previous studies (14, 23, 24). In the present study, 84% of 25 patients with Tmax ≥ 4 s ≥ 10 mL after TIA onset had a history of hypertension. We found that the increased prevalence of MR perfusion lesions occurred in patients with a history of hypertension, which was further confirmed in the stepwise multiple logistic regression analysis, suggesting that history of hypertension is an independent risk factor for MR perfusion abnormality in patients with TIA. A previous study showed that hypertension could lead to morphological impairment of the cerebral microvasculature, blood-brain barrier disruption, and neuroinflammation (25). Previous findings suggest that acute PWI lesions may be due to a persistent microvascular injury that results in hypoperfusion (15, 26). However, we found that sBP at admission is not a stronger predictor of MR perfusion abnormality after TIA than a history of treated hypertension, which is inconsistent with previous reports on Western populations (23, 24). In their reports, elevated SBP at presentation is more predictive of stroke after a TIA than a history of hypertension (23, 24). There are a couple of possible reasons for this discrepancy. Firstly, Median sBP at admission was measured 2 days after the acute TIA period (>24 h after symptom onset), therefore, it is less likely to reflect the real sBP when TIA occurred and therefore, less predictive for poor short-term prognosis (27). Secondly, the fluctuation of sBP (130– 150 mmHg) in the early course of TIA is minimal, which is not associated with poor 90-day survival (28). Together, our findings suggest that the history of hypertension, but not sBP at admission, is significantly associated with local PWI lesions after a TIA.

In subset analysis of our participants with MR perfusion abnormalities, one-thirds (8/25) of the patients had evidence of proximal artery stenosis or occlusion, which is consistent with previous reports (11, 14, 29). In our study, proximal artery stenosis (>50%) tended to have a larger PW lesion on MRI



*CI, confidence interval; OR, Odds ratio; sBP, Systolic blood pressure at admission.*

*ABCD2: age 60 (1 point), SBP 140, or DBP 90 mm Hg (1 point), clinical features as unilateral weakness (2 points) or speech impairment without weakness (1 point), symptom duration 60 min (2 points), or 10–59 min (1 point), diabetes (1 point).*

*Model 1: adjusted for age and sex.*

*Model 2: adjusted for ABCD2.*

*Model 3: adjusted for age, sex, and ABCD2.*

scans (adjusted OR = 3.60, 95% CI (0.91–14.20), p = 0.067). This finding suggests the added diagnostic value of MR perfusion imaging with MRA for detection of hemodynamic abnormality within the microvasculature (13, 30). In the present study, the widely used ABCD2 score was not associated with perfusion deficit, which is similar to what was reported by a previous study (30). The possible explanation for this might be that ABCD2 score is based on patients' clinical factors and does not include information about brain hemodynamics.

This study has a number of limitations. Firstly, it is a crosssectional study design and cannot demonstrate direct causality between MR perfusion and the risk factors in subjects with TIA. A longitudinal design can help to investigate the direct causality of MR perfusion in future studies. Secondly, we had a relatively small sample size, possibly introducing unknown patient selection bias. Therefore, a large sample size would be optimal for confirming our findings. Thirdly, all patients were recruited from inpatients admitted to one local hospital. Hence, conclusions and observations should be treated with caution. However, our hospital is the first and the only one that can use RAPID to calculate the volume of Tmax ≥ 4 s within the first 7 days after a TIA attack in China. Fourthly, the present study lacks imaging and clinical follow-up. It is unknown whether perfusion abnormalities observed were reversible or progressed to infarction after initial imaging. Therefore, the findings in this study should be considered as preliminary and should be confirmed in future studies. Fifthly, in this study we used Tmax ≥ 4 s for defining perfusion deficits (21), and volume of Tmax ≥ 4 s > 10 mL for defining perfusion abnormality (13). Although our method is based on a previous study, whether this method has better accuracy and applicability needs further prospective, large-scale studies to verify.

In conclusion, history of hypertension is a strong predictor of focal perfusion abnormality calculated by RAPID on Tmax maps in DWI-negative TIA patients. However, further prospective studies including a larger number of patients are needed to confirm this finding.

### REFERENCES


### DATA AVAILABILITY

All datasets generated for this study are included in the manuscript and/or the supplementary files.

### ETHICS STATEMENT

Ethical approval for this study was obtained from Human Research Ethics Committee of Shanghai Fourth People's Hospital Affiliated to Tongji University School of Medicine. Written informed consent was obtained from all subjects.

## AUTHOR CONTRIBUTIONS

YW, HL, and SW contributed to design and conceptualization of the study, data collection, analysis, and interpretation of the data, and drafting of the original manuscript. YL contributed to data collection and revision of the manuscript. YZ contributed to data collection and revision of the manuscript. SW, YB, and LJ contributed to data interpretation and revision of the manuscript.

### FUNDING

The present study was supported by a grant from Shanghai Municipal Commission of Health and Family Planning awarded to YB (No. 201840244); a grant from Commission of Health and Family Planning, Hongkou District, to YZ (No. 1802-07v), a grant from Shanghai Health Bureau Science and Research Projects Foundation (grant number 201740137 to YL) and a grant from Fundamental Research Funds for the Central Universities awarded to YW (No. 22120180281).

### ACKNOWLEDGMENTS

We acknowledge Dr. Xi Zhang, Yangyang Huang, and Linglei Meng for their support in providing relevant cases.

access (SOS-TIA): feasibility and effects. Lancet Neurol. (2007) 6:953–60. doi: 10.1016/s1474-4422(07)70248-x


**Conflict of Interest Statement:** The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

Copyright © 2019 Wang, Liang, Luo, Zhou, Jin, Wang and Bi. This is an open-access article distributed under the terms of the Creative Commons Attribution License (CC BY). The use, distribution or reproduction in other forums is permitted, provided the original author(s) and the copyright owner(s) are credited and that the original publication in this journal is cited, in accordance with accepted academic practice. No use, distribution or reproduction is permitted which does not comply with these terms.

# Ten Years of BrainAGE as a Neuroimaging Biomarker of Brain Aging: What Insights Have We Gained?

Katja Franke<sup>1</sup> \* and Christian Gaser 1,2 \*

<sup>1</sup> Structural Brain Mapping Group, Department of Neurology, University Hospital Jena, Jena, Germany, <sup>2</sup> Department of Psychiatry, University Hospital Jena, Jena, Germany

With the aging population, prevalence of neurodegenerative diseases is increasing, thus placing a growing burden on individuals and the whole society. However, individual rates of aging are shaped by a great variety of and the interactions between environmental, genetic, and epigenetic factors. Establishing biomarkers of the neuroanatomical aging processes exemplifies a new trend in neuroscience in order to provide risk-assessments and predictions for age-associated neurodegenerative and neuropsychiatric diseases at a single-subject level. The "Brain Age Gap Estimation (BrainAGE)" method constitutes the first and actually most widely applied concept for predicting and evaluating individual brain age based on structural MRI. This review summarizes all studies published within the last 10 years that have established and utilized the BrainAGE method to evaluate the effects of interaction of genes, environment, life burden, diseases, or life time on individual neuroanatomical aging. In future, BrainAGE and other brain age prediction approaches based on structural or functional markers may improve the assessment of individual risks for neurological, neuropsychiatric and neurodegenerative diseases as well as aid in developing personalized neuroprotective treatments and interventions.

Keywords: brain age estimation, biomarker, intervention, metabolic health, MRI, neurodegeneration, neurodevelopment, psychiatric disorders

### INTRODUCTION

With population growth and prolonged lifespan, the numbers of individuals with a range of (non-fatal, but) disabling disorders, including neurodegenerative diseases such as cognitive decline and dementia, are rising (1). Understanding the links between brain aging processes and neurodegenerative disease mechanisms is an urgent priority for health systems in order to establish effective strategies to deal with the rising burden. Aging is broadly defined as a time-dependent functional decline, driven by a progressive accumulation of cellular damage throughout life (2) and changes in intercellular communication (3–6). Aging is also a vastly complex process, which is individually modified by manifold genetic and environmental influences (5).

The assessment of the individual's "biological age" was recently promoted, resulting from the interaction of genes, environment, lifestyle, health, and life time, in order (i) to identify subject-specific health characteristics as well as subject-specific risk patterns for various age-related diseases based on pre-established reference curves for healthy aging, and (ii) to develop and monitor (clinical) interventions that are personally tailored based on "biological age" instead of chronological age (7). Several cell-, tissue- or function-based biomarkers that measure differences in the individual aging processes have been developed recently in order to identify and predict

### Edited by:

Brad Manor, Institute for Aging Research, United States

### Reviewed by:

Martin Gorges, University of Ulm, Germany Yenisel Cruz-Almeida, University of Florida, United States

### \*Correspondence:

Katja Franke katja.franke@uni-jena.de Christian Gaser christian.gaser@uni-jena.de

### Specialty section:

This article was submitted to Applied Neuroimaging, a section of the journal Frontiers in Neurology

Received: 13 February 2019 Accepted: 09 July 2019 Published: 14 August 2019

### Citation:

Franke K and Gaser C (2019) Ten Years of BrainAGE as a Neuroimaging Biomarker of Brain Aging: What Insights Have We Gained? Front. Neurol. 10:789. doi: 10.3389/fneur.2019.00789

**143**

individual risks for age-associated diseases and mortality [for recent reviews see (8, 9)], as well as to improve intervention and treatment strategies (2, 5), including DNA methylation status, measuring the accumulation of genetic damage (7, 10, 11), telomere length, assessing telomere attrition (12–16), physical fitness, and allostatic load as a measure for physical, physiological, and metabolic health etc. (17, 18).

Structural brain maturation/aging in humans is characterized by region-specific, non-linear patterns of very well-coordinated and sequenced occurrences of progressive and regressive processes (19) / atrophy (20, 21), respectively, demonstrating robust patterns of alterations (22, 23), where some brain regions are showing greater alterations than others. With the advent of non-invasive methods of in vivo brain imaging, especially magnetic resonance imaging (MRI), and the availability of sophisticated computational methods for processing and analyzing MRI data, cross-sectional as well as longitudinal neuroimaging studies on brain structure and function are increasingly contributing to a more profound understanding of healthy as well as diseased structural brain maturation and aging for recent reviews see (8, 9).

As research increasingly focuses on the interplay between aging and disease, a growing body of research utilizes neuroimaging to develop a biomarker of individual brain health, so-called "brain age." Lately, data-driven learning methods, including cross-validation, pattern classification, and regression-based predictive analyses, exemplify a new trend in biomedical and neuroscientific research, allowing measurements and predictions even at the single subject level (24). To determine the individual trajectory of brain maturation and aging as well as the risks for cognitive dysfunction and age-associated brain diseases, a number of structural and functional brain-based prediction methods for age or cognitive state enjoy increasing popularity in (cognitive) neuroscience, providing personalized biomarkers of brain structure, and function by identifying deviations from pre-established reference curves or automatically discriminating patients with brain disorders from healthy controls (25–30). Most of these studies are using state-of-the-art machine learning techniques to make predictions at the singlesubject level. Especially pattern recognition and regression-based computational modeling methods aim to predict the values of continuous variables, like structural brain age, cognitive states, or neuropsychological characteristics (27). These new brain-based biomarkers offer a powerful strategy for using neuroscience in clinical practice and have a wide range of implementations, such as providing reference curves for healthy brain maturation/aging, predicting personalized brain maturation/aging trajectories, discovering protective, and harmful environmental influences on brain health, disentangling age-related from disease-specific changes in individual brain structure, aiding in the riskassessment, and early detection of certain neurodegenerative diseases, tracking individual disease progression, as well as determining the individual relationship of structural brain aging to cognitive performance and neuropsychiatric symptoms (8).

The "brain age gap estimation (BrainAGE)" method, which utilizes structural MRI data to directly quantify acceleration or deceleration of individual brain aging, was the first brain aging estimation approach that (1) established reference curves for healthy brain maturation during childhood into young adulthood and for healthy brain aging during adulthood into senescence, (2) examined deviations of individual brain aging from the established reference curve of healthy brain aging in neurodegenerative diseases, (3) analyzed longitudinal changes of individual brain aging in several samples, (4) used deviations of individual brain age predictions from the established reference curve of healthy brain aging to predict worsening of cognitive functions and conversion to Alzheimer's disease (AD), (5) studied the effects of a number of several health- and lifestyle-related factors on individual brain aging, (6) monitored the effects of protective interventions on individual brain aging, and (7) was adapted to be also applied in experimental studies with rodents and nun-human primates. This review firstly describes the generation of the BrainAGE model and secondly recapitulates and integrates all studies predicting individual brain age with the innovative BrainAGE method in healthy and diseased populations. Wherever possible, studies applying other brain age prediction approaches to examine the very issue are additionally included in this review. A short summary of all BrainAGE studies summarized here can be found in **Table 1**.

## GENERATION OF THE BRAINAGE MODEL

A growing body of research is using high-dimensional neuroimaging data, i.e., often including several hundred (multi-modal) parameters per individual, and employing supervised, linear, or non-linear pattern recognition techniques in order to depict and quantify structural brain development and aging across the lifespan. In contrast to univariate approaches, multivariate analyses of individual brain structure are able to detect and quantify subtle, but widespread deviations in regionor voxelwise brain structure within the whole brain for the individual's age.

In general, the brain age prediction model needs to be trained first in order to subsequently assess a person's individual brain age. The brain age prediction model is generated by recognizing multivariate patterns of age-typical brain structure and parameters, utilizing MRI data of a large sample of (cognitively) healthy subjects. Subsequently, the age prediction model is applied in previously unseen test subjects, i.e., estimating the subject-specific brain ages utilizing their individual MRI data. The difference between a person's estimated brain age and its chronological age finally identifies the individual deviation from the typical maturation/aging trajectory.

## Pipeline for the Generation of Brain Age Estimations

In general, the workflow of our innovative BrainAGE model includes several processing steps (**Figure 1**). Firstly, the raw T1 weighted image data are preprocessed with a standardized voxelbased morphometry (VBM) pipeline, resulting in comparable as well as more easily processible data to be utilized in the following analysis steps (see Preprocessing of raw MRI data). Secondly, automated data reduction of the preprocessed

Franke and Gaser


Frontiers in Neurology | www.frontiersin.org

(Continued) BrainAGE

### TABLE 1 | Continued


(Continued) BrainAGE

### TABLE 1 | Continued


(Continued)

BrainAGE

### TABLE 1 | Continued


(Continued)

BrainAGE

### TABLE1| Continued


(Continued)

BrainAGE


MRI data is performed in order to reduce computational costs, avoid method-typical over-fitting of pattern recognition, as well as to provide a robust and widely applicable age estimation model (see Data reduction). Thirdly, relevance vector regression (RVR) is performed, capturing the multidimensional maturation/aging patterns throughout the whole brain and thus modeling structural brain maturation/aging. Subsequently, individual brain ages can be estimated (see Training of the BrainAGE algorithm).

### Preprocessing of Raw MRI Data

Preprocessing of the raw MRI data is done using SPM including the VBM8/CAT12 toolbox, running under MATLAB. More specifically, T1-weighted images are corrected for bias-field inhomogeneities (46, 47). Following, the images are spatially normalized. Afterwards, the images are segmented into the tree brain tissue types, i.e., gray matter (GM), white matter (WM), and cerebro-spinal fluid (CSF), within the same generative model (48). Furthermore, adaptive maximum a posteriori estimations (49) and a hidden Markov random field model (50) are applied in order to account for partial volume effects (51). Finally, image preprocessing includes affine registration.

### Data Reduction

Preprocessed MRI data are smoothed with 4 or 8 mm fullwidth-at-half-maximum (FWHM) Gaussian kernels. Thereafter, data are re-sampled to 4 or 8 mm spatial resolution, resulting in 29,852 or 3,747 voxels per subject after masking out nonbrain areas, respectively. Finally, principal component analysis (PCA) is applied to further reduce data dimensionality. As a great portion of the resulting voxels are still sharing much of its variances with their neighboring voxels, PCA is mathematically allowed to be performed although the numbers of data sets in the training sample is lower than the number of voxels, given the numbers of data sets in the training sample is sufficient (see Performance of the BrainAGE model for brain aging from early into late adulthood). The PCA model is calculated within the training data only and subsequently the resulting transformation parameters are utilized to reduce data dimensionality within the independent test samples.

### Training of the BrainAGE Algorithm

The BrainAGE framework utilizes RVR (52, 53) with a linear kernel. Importantly, RVR does not require additional (manual) parameter optimization during the training procedure, which is advantageous over the commonly used support vector machines with regards to computational costs and robust model fitting.

In general, the age regression model is calculated within the training sample, utilizing the preprocessed structural MRI data as independent variables and the chronological ages as dependent variables, resulting in a complex model of healthy brain maturation/aging (**Figure 1A**, left panel). Within this specified regression task (i.e., healthy brain maturation/aging), voxel-specific weights are calculated, representing the voxelspecific importance within this regression task (for illustrations of the resulting voxel-specific weights see **Figure S1** for the brain maturation model & **Figure S2** for the brain aging model).

MCI to AD during follow-up); pMCI\_early,

MCI (i.e., diagnosis was MCI at baseline and conversion to AD was reported after the first 12 months of follow-up, without reversion to MCI or CTR at any available follow-up); sMCI: stable MCI (i.e., diagnosis is MCI at all available time

points, but at least for 36 months); SZ, schizophrenia;

et al. (37);

iNenadic et al. (38);

kHajek et al. (39);

 early converting pMCI (i.e., diagnosis was MCI at baseline but converted to AD within the first 12 months, without reversion to MCI or CTR at any available follow-up); pMCI\_late, late converting

 T-Tau, total tau, WM: white matter

lKolenic et al. (40);

mFranke et al. (41);

aFranke et al. (31);

nFranke et al. (42);

bFranke et al. (32);

oLuders et al. (43);

cFranke et al. (33);

pRogenmoser

 et al. (44);

qFranke et al. (33).

dFranke et al. (34);

eFranke and Gaser (31);

f Franke et al. (35);

gLöwe et al. (36);

hGaser

FIGURE 1 | Depiction of the BrainAGE concept. All MRI data are automatically preprocessed via VBM. (A) The model of healthy brain aging is trained with the chronological age and preprocessed structural MRI data of a training sample (left; with an illustration of the most important voxel locations that were used by the age regression model). Subsequently, the individual brain ages of previously unseen test subjects are estimated, based on their MRI data. (B) The difference between the estimated and chronological age results in the BrainAGE score, with positive BrainAGE scores indicating advanced brain aging (orange line), increasing BrainAGE scores indicating accelerating brain aging (red line), and negative BrainAGE scores indicating delayed brain aging (green line). [Figure and legend adapted from Franke et al. (45), with permission from Hogrefe Publishing, Bern].

Subsequently, the brain maturation/aging model is applied to aggregate the complex, multidimensional maturation/aging pattern throughout the whole brain of a new test subject, resulting in one single value, i.e., the estimated brain age (**Figure 1A**, right panel).

Finally, the difference between estimated brain age and chronological age reveals the individual brain age gap estimation (BrainAGE) score. For BrainAGE, positive values are indicating advanced structural brain maturation/aging, whereas negative values are indicating delayed structural brain maturation/aging. In longitudinal studies, increasing BrainAGE scores are indicating accelerating brain aging over the time. Thus, the individual BrainAGE score is directly quantifying the amount of acceleration or deceleration of brain maturation/aging in terms of years (**Figure 1B**). For example, if a 70 years old individual shows a BrainAGE score of +5 years, the typical atrophy pattern of this individual resembles the brain structure of a 75 years old individual.

### Cross-Validation of the BrainAGE Model in Reference Samples

In order to generate and validate the brain age model, most studies are employing a so-called "cross-validation" approach, i.e., the neuroimaging parameters of a large portion of the reference sample of healthy individuals are used to generate the brain age model. The generated brain age model is then applied to the smaller portion of the reference sample that was not included in the model generation step (i.e., "leftout"), in order to predict individual brain ages based on the identified neuroimaging parameters within the actual training sample. This procedure is repeated multiple times, until an individual brain age is provided for each subject in the whole reference sample.

To measure the accuracy of age estimation, Pearson's correlation coefficient (r), mean absolute error (MAE), and root mean squared error (RMSE) between individual estimated brain ages and chronological ages are calculated:

$$\text{MAE} = 1/\text{n}^\* \sum\_{\mathbf{i}} |\text{BAi} - \text{CA}\_{\mathbf{i}}|,\tag{1}$$

$$\text{RMSE} = \left[1/\text{n}^\* \sum\_{\text{i}} (\text{BA}\_{\text{i}} - \text{CA}\_{\text{i}})^2\right]^{1/2},\tag{2}$$

with n being the number of subjects in the test sample, BA<sup>i</sup> being the estimated subjects-specific brain ages, and CA<sup>i</sup> being the subject-specific chronological ages. Additionally, F statistics of the regression model is used to analyze the fit between BA and CA.

### Application of the Generated BrainAGE Model in Independent Test Samples

Additionally to the cross-validation in the reference samples, the brain age model is further validated in independent test samples of healthy and clinical subjects, in order to prove the generalizability of the pre-established brain age model across different samples and even MRI scanners, which is crucial for broad application in a clinical context, as well as to investigate the power of the brain age models as a diagnostic and prediction tool at a single-subject level, for monitoring individual changes in brain aging during treatment studies, or to explore the effects of various health characteristics, diseases, and life experiences on individual brain aging.

### Species-Specific Adaptations of the BrainAGE Model for Experimental Animal Studies

### Species-Specific BrainAGE Model for Baboons

Within the species-specific BrainAGE model for baboons, we used a customized preprocessing pipeline as described in Franke et al. (33). To further reduce high-frequency noise, a spatial adaptive non-local means (SANLM) filter (54) is applied. The segmentation and spatial registration step requires a baboon-specific tissue probability map (TPM) as well as a "Diffeomorphic Anatomical Registration using Exponentiated

Lie algebra" (DARTEL) template (55), which is estimated during an iterative process based on a rescaled human template. More specifically, affine transformation is initially used to scale the human SPM12 TPM and the CAT12 Dartel template map onto the brain size of baboons. Image resolution of this template is set to isotropic voxel size of 0.75 mm. For each of the performed iteration steps, the resulting tissue maps are averaged and subsequently smoothed with a 2 mm FWHM kernel to estimate an affine registration, finally resulting in a new TPM, a T1-average map, as well as a baboon-specific brain mask. To achieve averaged data, a median function is used in order to reduce distortions by outliers or failed processing. The iteration process is stopped when the actually accomplished change is below a pre-defined threshold as compared to the previous template, resulting in the final segmentation.

After Segmentation and Registration, Data are Smoothed With a 3 mm FWHM Gaussian Smoothing Kernel and resampled to 3 mm. Finally, PCA is Applied to Further Reduce Data Complexity (as Described in Data Reduction).

### Species-Specific BrainAGE Model for Rodents

As described in Franke et al. (34), a preprocessing framework for automatically preprocessing and analyzing MRI data of rodents is providing analyses in the space of a Paxinos atlas (56), including several realignment and normalization steps. First, affine co-registration to the Paxinos template is applied utilizing normalized mutual information. In the next step, a deformation based morphometry (DBM) approach is utilized to analyze positional differences between every voxel within the actual brain data and a reference brain in order to detect structural differences over the entire brain. Thus, all measured time points of the data set of one animal are registered to the individual baseline scan. Afterwards, the deformations between all-time points and the subject-specific baseline measures are being estimated. Minimizing the morphological differences between the baseline and the follow-up brain scans, the deformation maps now encode the information about these differences. Subsequently, the Jacobian determinant of the deformations can be used to calculate local volume changes. Finally, the resulting Jacobian determinants in each voxel are filtered with a 0.4 mm FWHM Gaussian smoothing kernel.

### Technical Notes

The BrainAGE framework is fully automatic. All steps, including MRI preprocessing, data reduction, model training, and brain age estimation, are executed within MATLAB (www.mathworks.com). For preprocessing the T1-weighted images, SPM8 is utilized (www.fil.ion.ucl.ac.uk/spm), integrating the VBM8 toolbox (http://dbm.neuro.uni-jena.de). For the generation of brain age models in baboons and rodents our new CAT12 toolbox (http://dbm.neuro.uni-jena.de) is utilized. For PCA, the "Matlab Toolbox for Dimensionality Reduction" (https://lvdmaaten.github.io/drtoolbox/) is applied. RVR analyzes are performed utilizing the toolbox "The Spider" (http://people.kyb.tuebingen.mpg.de/spider/).

Preprocessing the human MRI data takes about 20–30 min per MRI data set on a MAC OS X, Version 10.12, 2.2 GHz Intel Core i7. The whole process of training the BrainAGE model and estimating brain ages takes between 1 and 5 min in total, depending on the number of features, training, and test subjects.

Baboon TPM and template generation needs about 30 min per subject and iteration, summing up in about 48 h for the whole sample of 29 control subjects. The whole process of training the baboon-specific BrainAGE model and estimating the individual brain ages takes about 1 min in total.

Preprocessing MRI data of rodents takes about 10–15 min per MRI data set on MAC OS X, Version 10.6.3, 2.8 GHz Intel Core 2 Duo, resulting in about 5–6 h for a sample of 24 rats with up to 13 MRI data sets per subject. Within this sample, the whole process of training the rodent-specific BrainAGE model and estimating the individual brain ages is performed within about 5 min.

### EVALUATION OF BRAINAGE PREDICTION PERFORMANCE IN REFERENCE SAMPLES

### Performance of the BrainAGE Model for Brain Maturation During Childhood and Adolescence

For generating the BrainAGE model during childhood and adolescence (31), GM and WM images of a cross-sectional reference sample of 394 healthy children and adolescents from the Pediatric MRI Data Repository [NIH MRI Study of Normal Brain Development; (57)] were utilized, aged 5–18 years (mean age = 10.7 years; SD = 3.9 years), with structural data acquired on six different MRI scanners (1.5T). Using leave-one-out cross-validation, the MAE between estimated brain age and chronological age was 1.1 years. Between estimated brain age and chronological age 87% of the variance were explained (r = 0.93; p < 0.001), with the 95% confidence interval being stable across the age range (± 2.6 years; **Figure 2A**).

Additionally, training the BrainAGE model with the data from only five of the six MRI scanner sites included in the study, and then applying to data from the left-out MRI scanner, estimation accuracy proved to remain stable across all scanner sites. Prediction accuracy ranged between r = 0.90–0.95 and MAE = 1.1–1.3 years, which proved stability of brain age estimation even across scanners (31).

A number of other studies establishing models for brain maturation including age ranges from early childhood to young adulthood have been published so far (58–63). Accuracies for brain age predictions derived from cross-validation in the reference sample ranged from r = 0.43–0.96 and MAEs from 1.0 to 1.9 years. The most accurate model for brain age prediction during development in healthy individuals aged 3–20 years used a number of parameters derived from different MRI modalities (i.e., T1, T2, DTI), including cortical thickness, cortical surface area, subcortical volumes, apparent diffusion coefficient, fractional anisotropy, and T2 signal intensities in predefined

scans. The overall correlation between chronological age and estimated brain age is r = 0.80 (p < 0.001), with an overall MAE of 2.1 years. [Figure and legend reproduced from Franke et al. (33), permitted under the Creative Commons Attribution License.] (D) (a) Chronological and estimated brain age are shown for a sample of untreated control rats, including the 95% confidence interval (gray lines). The overall correlation between chronological and estimated brain age was r = 0.95 (p < 0.0001). [Figure and legend reproduced from Franke et al. (34), with permission from IEEE.] (E) Longitudinal brain aging trajectories for the individual rats. [Figure and legend reproduced from Franke et al. (34), with permission from IEEE].

subcortical regions, applying a regularized multivariate nonlinear regression-like approach, resulting in r = 0.96 and MAE = 1.0 years (59). Although each single MRI modality showed similar predictive power (r ≈ 0.9) across the full age range (i.e., 3– 20 years), modality-specific contributions to the generation of the brain age model differed across neuroanatomical structures and age sub-ranges, with measures of T2 signal intensity being the strongest predictors in age 3–11 years and diffusivity measures being the strongest predictors in the ages 17–20 years (59). Additionally, modality-specific subsets showed worse prediction accuracies compared to the combined model (T1 subset: r = 0.91, MAE = 1.7 years; T2 subset: r = 0.91, MAE = 1.6 years; DTI subset: r = 0.90, MAE = 1.7 years). However, the BrainAGE method (31) outperformed all other brain age models using only a single MRI modality or single-modality subsets, and additionally proved sufficient generalizability across different scanners and even across studies.

# Performance of the BrainAGE Model for Brain Aging From Early Into Late Adulthood

In our first study introducing the BrainAGE model (32), two different samples were used to assess the brain age, i.e., the reference sample from the IXI database (www.brain-development.org; n = 550, aged 19–86 years, collected on three MRI scanners) and another independent test sample of healthy subjects (n = 108, aged 20–59 years, collected on a fourth scanner). The brain age of healthy subjects in both validation samples was accurately estimated, resulting in a MAE of 5 years and an overall correlation of r = 0.92, with the 95% confidence interval for the prediction of age being stable across the age range (**Figure 2B**). The BrainAGE model showed no systematical bias in MAE of brain age estimation as a function of chronological age (r = – 0.01). Furthermore, brain age estimation did not differ between genders (r = 0.92 for both genders; MAE = 5.0 years for males, MAE = 4.9 years for females).

Additional analyses showed that the number of subjects in the reference sample has the strongest influence on brain age prediction accuracy, even though the choice of the preprocessing approach and model-training algorithm would also influence model performance as well as generalizability (32). In detail, the accuracy of brain age estimation worsened with reducing the size of the training/reference sample (full data set for training the BrainAGE model [n = 410]: MAE = 5 years; ½ data set [n = 205]: MAE = 5.2 years; ¼ data set [n = 103]: MAE = 5.6 years). The results further recommend a fairly rapid preprocessing of the T1-weighted MRI images with affine registration and a rather broad smoothing kernel. Dimensionality reduction of the data via PCA moderately improved brain age estimation accuracy and generalizability, while at the same time speeding up the computing time for generating the BrainAGE model as well as estimating the individual brain age values of the independent test subjects (**Figure 3**).

A number of other studies establishing models for brain aging have been published so far (55, 60, 64–79). Accuracies for brain age predictions derived from cross-validation in the whole reference sample of healthy subjects ranged from r = 0.43–0.97, MAEs from 4.3 to 13.5 years, and RMSEs from 5.1 to 21.0 years. In general, studies mathematically modeling healthy brain aging, which use a number of parameters derived from different MRI modalities, tended to provide more accurate brain age predictions. The best performing model in a sample of healthy participants aged 8–85 years was based on a number of T1- and DTI-derived parameters, utilizing linked independent component analysis (ICA), resulted in an overall prediction accuracy of r = 0.97 and MAE = 5.9 years (67). Another study also used a number of parameters derived from different MRI modalities (i.e., T1, T2, T2<sup>∗</sup> , DTI), generating and testing their brain age model by utilizing multiple linear regression in a sample of healthy individuals aged 20–74 years, resulting in an overall age prediction accuracy of r = 0.96 (74). Additionally, this study found voxel-wise mean diffusivity to be the main predictor of the brain age model (i.e., explaining 62.4% of intra-individual variance), followed by GM volume (18.3%), R2<sup>∗</sup> (14.2%) and fractional anisotropy (3%). However, although DTI is a powerful tool offering unique information on tissue microstructure and neural fiber connections that cannot be obtained from standard structural MRI, parameters derived from DTI can differ significantly depending on the type of scanner, field strength, gradient strength, number of gradient orientations, preprocessing, fitting procedure, tractography algorithm etc. (80–83). Unfortunately, all studies including DTI failed to prove generalizability of the established brain age model in independent test samples and across scanners.

Another very recent study used a number of parameters derived from T1 and T2<sup>∗</sup> , including cortical and subcortical measures as well as connectivity data, generating and testing the brain age model by utilizing linear support vector regression (SVR) (79). This approach showed very good performance during cross-validation within the reference sample (combined model: r = 0.93, MAE = 4.3 years), but a rather fair generalizability when validating the brain age model in an independent sample of healthy subjects, with data acquired on a different scanner (combined model: r = 0.86, MAE = 8.0 years).

Aside from the BrainAGE approach, best prediction accuracies during cross-validation in the reference samples as well as during validation of the brain age model in independent test samples were achieved utilizing linear SVR (reference sample: r = 0.89, MAE = 4.3 years; independent test sample: MAE = 3.9 years; (76)], and Gaussian process regression [reference sample: r = 0.92, MAE = 6.2 years; independent test sample: r = 0.93, MAE = 5.8 years; (73)].

### Performance of the BrainAGE Model in Baboons

For establishing the baboon-specific brain aging model, only GM images were used. The baboon-specific brain age estimation model was trained and tested via leave-one-out cross-validation, utilizing one MRI scan per subject. Within each cross-validation loop, PCA was calculated separately in the training set and subsequently applied to the test data before performing RVR. The baboon-specific BrainAGE model showed very good accuracy (r = 0.80), with the linear regression model showing the best fit (R 2 = 0.64; p < 0.0001; **Figure 2C**). Calculation of MAE resulted in 2.1 years, equating to an age estimation error of 11% in relation to the age ranged included (33, 34).

### Performance of the BrainAGE Model in Rodents

As described in Franke et al. (34), training and testing of the rodent-specific BrainAGE model was performed with subjectspecific leave-one-out cross-validation processing, utilizing data sets of 24 rats, repeatedly scanned with up to 13 time points between 97 and 846 days after birth. In detail, to model the rodent-specific aging process, RVR was performed with the preprocessed structural MRI data of all scanning time points of 23 out of the total of 24 subjects. Subsequently, individual brain ages for each scanning time point of the left-out test subject were estimated, repeating the whole procedure for all 24 subjects. Brain age estimation was highly accurate (r = 0.95; p < 0.0001), with the linear regression model showing the best fit between chronological and estimated age (R <sup>2</sup> = 0.91; F = 2622.3; p < 0.0001; **Figure 2D**). Mean MAE was 49 days, which equates to an error of 6% in relation to the age range within this study. Mean RMSE was 71 days. Additionally, longitudinal analyses of subject-specific brain aging trajectories revealed increasing variance between subjects in old age (**Figure 2E**).

## RELIABILITY OF BRAINAGE ESTIMATIONS IN HEALTHY ADULTS

## Scan-Rescan-Stability of BrainAGE Estimations (Same Scanner)

To analyze stability and reliability of BrainAGE estimations, T1 weighted MRI data of 20 healthy subjects were utilized, applying the BrainAGE method to two MRI scans per subject, which were acquired on the same MRI scanner (1.5T) within a time period of max. 90 days. The results showed a strong scan-rescan-stability of BrainAGE estimations based on MRI data acquired on the same scanner, with mean BrainAGE scores between 1st and 2nd scan not differing among each other (p = 0.60) and the intra-class correlation coefficient (ICC; two-way random single measures) between BrainAGE scores calculated from the 1st and 2nd scan resulting in 0.93 [95% confidence interval [CI]: 0.83–0.97; (45)].

# Effect of Different MRI Field Strengths on BrainAGE Estimations

To analyze estimation stability across different scanners and field strengths, T1-weighted MRI data of 60 healthy subjects (aged 60– 87 years) were utilized, applying the BrainAGE method to two MRI scans per subject, acquired on two different MRI scanners (1.5T & 3T) within a short period of time. The results suggest that the field strength affects BrainAGE estimations, which should be corrected for by shifting the BrainAGE scores to a zero group mean with a linear term in both data sets in order to gain interpretability of the results (**Figure S3**). After linearly adjusting for the scanner-specific offset, Student's t-test did not show any difference between the BrainAGE scores calculated from the 1.5T and 3T scans (p = 1.00). ICC between the BrainAGE scores calculated from the 1.5T and 3T scans resulted in 0.90 (CI: 0.84– 0.94), demonstrating strong reliability and generalizability of the BrainAGE model, even with data from different scanners and field strengths (45).

### Sensitivity to Hormone-Related Short-Term Changes of BrainAGE in Women

In order to establish the BrainAGE model as an innovative tool to monitor and evaluate short-term changes in individual brain aging induced by treatments and interventions, we explored its potential to recognize short-term changes in brain structure occurring during the menstrual cycle due to varying hormonal influences (35). A total of 7 young, healthy, naturally cycling women (age range 21–31 years) were scanned on a 1.5T MRI scanner (t1) during menses, (t2) at time of ovulation, (t3) in the midluteal phase, and (t4) at their next menses. During menstrual cycle BrainAGE scores significantly differed (p < 0.05), with BrainAGE scores decreasing by −1.3 years from menses to ovulation (SD = 1.2 years; p < 0.05) and after ovulation slowly increasing (**Figure 4**). Additionally, estradiol levels did negatively correlate with BrainAGE scores (r = −0.42, p < 0.05), but progesterone levels did not (r = 0.08, p = 0.71).

Another study by Luders et al. (84) explored the changes in BrainAGE after pregnancy. A total of 14 healthy women (aged 25–38 years) were scanned on a 3T MRI scanner within the first two after childbirth (early postpartum) as well as 4–6 weeks after childbirth (late postpartum). BrainAGE scores were significantly decreased by an average of −5.4 years from early to late postpartum (SD = 2.4 years; p < 0.001). Additional analyzes of hormone levels also showed a profound postpartum decrease in estradiol (p < 0.001) and progesterone (p < 0.001).

Taken together, these results provide strong evidence that hormonal changes during the course of the menstrual cycle have significant effects on the individual brain structure. Furthermore, the BrainAGE method demonstrated its potential to capture and identify subtle short-term changes in individual brain structure.

FIGURE 4 | Change in BrainAGE scores during the menstrual cycle. BrainAGE scores significantly decreased by −1.3 years (SD = 1.2) at time of ovulation (i.e., t2-t1; \*p < 0.05). The data are displayed as boxplots, containing the values between the 25th and 75th percentiles of the samples, including the median (red lines). Lines extending above and below each box symbolize data within 1.5 times the interquartile range. The width of the boxes depends on the sample size. Note: reduced sample size at t4. [Figure and legend reproduced from Franke et al. (35), with permission from Elsevier, Amsterdam].

# APPLICATIONS OF BRAINAGE MODEL FOR BRAIN MATURATION DURING CHILDHOOD AND ADOLESCENCE

### Effects of Being Born Preterm on Individual Brain Maturation

In a study with pre-term born adolescents, individual BrainAGE scores of subjects being born before the end of the 27th week of gestation (i.e., GA < 27; n = 10) were compared to those being born after the end of the 29th week of gestation (i.e., GA > 29; n = 15), applying the pre-established BrainAGE model for brain maturation during childhood and adolescence (31). At MRI scanning (1.5T), subjects were aged between 12 and 16 years. The results show significantly lower BrainAGE scores by 1.6 years in the group of adolescents being born GA < 27 (−1.96 ± 0.68 years) as compared to subjects being born GA > 29 (−0.40 ± 1.50 years), although the mean difference in gestation age was only 5 weeks, thus probably implying delayed structural brain maturation.

# BRAINAGE IN MILD COGNITIVE IMPAIRMENT AND ALZHEIMER'S DISEASE

### Premature Brain Aging in AD

In a first proof-of-concept application, individual brain ages was studied in a group of cognitively healthy control subjects (CTR; n = 232) and a group of patients suffering from early Alzheimer's disease (AD; n = 102), applying the pre-established BrainAGE model for brain aging during adulthood (32). For the AD group, the mean BrainAGE score was +10 years (p < 0.001), implying systematically advanced brain aging.

In another study that applied the pre-established BrainAGE model for brain aging during adulthood to data from the Alzheimer's Disease Neuroimaging Initiative (ADNI) database, baseline BrainAGE scores resulted in the following group means: (1) −0.3 years in CTR (i.e., being stable in the diagnosis of CTR during 36-months follow-up; n = 108), (2) −0.5 years in sMCI (i.e., stable MCI; being stable in the diagnosis of mild cognitive impairment (MCI) during 36-months follow-up; n = 36), (3) 6.2 years in pMCI (i.e., progressive MCI; changing diagnosis from MCI at baseline to AD during 36-months followup; n = 112), and (4) 6.7 years in AD (i.e., being stable in the diagnosis of AD during 36-months follow-up or until death; n = 150). Post-hoc t-tests resulted in significant BrainAGE differences between CTR/sMCI vs. pMCI/AD groups (p < 0.05), suggesting strong evidence for structural brain changes that show the pattern of advanced brain aging in the pMCI and AD groups (**Figure 5A**) (45).

### Longitudinal Changes of Individual Brain Aging in CTR, MCI, AD

Further analyses explored the individual brain aging trajectories in CTR, sMCI, pMCI, and AD during a follow-up period of up to 36 months (45). BrainAGE scores in pMCI and AD significantly increased by 1.0 additional year in brain aging per follow-up (chronological age) year in pMCI and 1.5 additional years in brain aging per follow-up (chronological age) year in AD, suggesting acceleration of individual brain aging during the course of disease (**Figure 5C**). With pMCI and AD subjects already showing advanced BrainAGE scores of about 6 to 7 years at baseline assessment and mean follow-up durations of 2.6 years for pMCI and 1.7 years for AD, mean BrainAGE scores at last follow-up MRI scan accumulated to about 9 years at the last MRI scan in both diagnostic groups (**Figure 5B**). In contrast, mean BrainAGE scores in CTR and sMCI subjects did not change during follow-up, thus suggesting no deviations from healthy brain aging in both groups.

Additionally, advanced structural brain aging was related to worse cognitive functioning and more severe clinical symptoms during the 36 months follow-up period (baseline BrainAGE scores: r = 0.39–0.46; BrainAGE scores at last follow-up visit: r = 0.46–0.55). Moreover, individual changes in BrainAGE scores were correlated with individual changes in cognitive test scores and clinical severity (r = 0.27–0.33), denoting a significant relationship between acceleration in individual brain aging and prospective worsening of cognitive functioning, being most pronounced in pMCI and AD subjects (45).

### Effects of APOE-Genotype on Longitudinal Changes in CTR, MCI, AD

Studying the effects of Apolipoprotein E (APOE) on individual brain aging trajectories during a 36 months follow-up period, neither APOE ε4-status, nor particular allelic isoforms had a significant effect on baseline BrainAGE scores in the four diagnostic groups (36). However, individual brain aging accelerated significantly faster in APOE ε4-carriers as compared to APOE ε4-non-carriers in the pMCI and AD groups. More specifically, in pMCI ε4-carriers individual brain aging accelerated with the speed of 1.1 additional years per follow-up year, whereas in pMCI ε4-non-carriers individual brain aging accelerated with the speed of only about 0.6 years. Likewise, in AD ε4-carriers individual brain aging accelerated with the speed of 1.7 additional years per follow-up year, whereas in AD ε4 non-carriers individual brain aging accelerated with the speed of only about 0.9 years per follow-up year. In line with previous results, deviations from normal brain aging trajectories were not observed in healthy controls or sMCI subjects, neither in ε4-carriers nor ε4-non-carriers (**Figure 6**).

FIGURE 5 | Longitudinal BrainAGE. Box plots of (A) baseline BrainAGE scores and (B) BrainAGE scores of last MRI scans for all diagnostic groups. Post-hoc t-tests showed significant differences between NO/sMCI vs. pMCI/AD (\*p < 0.05) at both time measurements. (C) Longitudinal changes in BrainAGE scores for NO, sMCI, pMCI, and AD. Thin lines represent individual changes in BrainAGE over time; thick lines indicate estimated average changes for each group. Post-hoc t-tests showed significant differences in the longitudinal BrainAGE changes between NO/sMCI vs. pMCI/AD (\*p < 0.05). [Figures and legend reproduced from Franke et al. (45), with permission from Hogrefe Publishing, Bern].

FIGURE 6 | Longitudinal BrainAGE in APOE ε4-carriers and ε4-non-carriers. BrainAGE scores at (A) baseline for APOE ε4-carriers [C] and non-carriers [NC] in the 4 diagnostic groups NO, sMCI, pMCI, and AD. BrainAGE scores differed significantly between diagnostic groups (p < 0.001). Post-hoc tests showed significant differences between BrainAGE scores in NO as well as sMCI from BrainAGE scores in pMCI as well as AD (p < 0.05). (B) Estimated longitudinal changes in BrainAGE scores for the 4 diagnostic groups: NO (light blue), sMCI (green), pMCI (red) and AD (blue), subdivided into APOE ε4 carriers and non-carriers. Post-hoc t-tests resulted in significant differences for ε4 carriers and non-carriers as well as for NO/sMCI vs. pMCI/AD (p < 0.05). [Figures and legend reproduced from Loewe et al. (36), permitted under the Creative Commons Attribution License].

### BRAINAGE-BASED PREDICTION OF CONVERSION TO ALZHEIMER'S DISEASE

### BrainAGE-Based Prediction of Conversion From MCI to AD

In a study by Gaser et al. (37), the BrainAGE approach was implemented to predict future conversion to AD at a singlesubject level up to 36 months in advance, based on structural MRI. The sample included 195 participants diagnosed with MCI at baseline, of whom 133 participants were diagnosed with AD during 36 months of follow-up. The BrainAGE scores at baseline examination differed significantly between the participants, who did not convert to AD (i.e., sMCI; 0.7 years) and those, who converted to AD within the 1st follow-up year (i.e., pMCI\_fast; 8.7 years) as well as in 2nd or 3rd follow-up year (i.e., pMCI\_slow; 5.6 year). A close relationship was shown between advanced brain aging, prospective worsening of cognitive functioning, and clinical disease severity. Predicting conversion from MCI to AD by using baseline BrainAGE scores, post-test probability increased to 90%. This gain in certainty based on the baseline BrainAGE score was 22%, being the highest as compared to baseline hippocampus volumes (right/left: 16%/17%), cognitive scores (MMSE: 11%; CDR-SB: 0%; ADAS: 18%), and even stateof-the-art CSF biomarkers (T-Tau: 4%, P-Tau: 0%, Aβ42: 0%, Aβ42/P-Tau: 8%). Predicting future conversion to AD during the 1st follow-up year based on baseline BrainAGE scores showed an accuracy of 81% (area under curve (AUC) in receiver-operating characteristic (ROC) analysis = 0.83), being significantly more accurate than conversion predictions based on chronological age, hippocampus volumes, cognitive scores, and CSF biomarkers (for exact numbers see **Table 1**). Furthermore, higher BrainAGE scores were related to a higher risk of developing AD, i.e., each additional year in BrainAGE score induced a 10% greater risk of developing AD (hazard rate: 1.1, p < 0.001). More specifically, as compared with participants in the lowest quartile of BrainAGE scores, participants in the 2nd quartile had about the same risk of developing AD (hazard ratio [HR]: 1.1; p = 0.68), those in the 3rd quartile had a three times greater risk (HR: 3.1; p < 0.001), and those in the 4th quartile had a more than four times greater risk (HR: 4.7; p < 0.001) of developing AD (**Figure 7A**). BrainAGE outperformed all other baseline measures.

## Effects of APOE-Genotype on BrainAGE-Based Prediction of Conversion From MCI to AD

A study by Loewe et al. (36) additionally explored the effects of the APOE-genotype on BrainAGE-based prediction of conversion from MCI to AD during the 36 months of follow-up period. Independent of APOE status, higher baseline BrainAGE scores were associated with a higher risk of converting to AD, with BrainAGE scores above median of 4.5 years resulting in a nearly 4 times greater risk of converting to AD as compared to BrainAGE scores below the median (HR: 3.8, p < 0.001). Again, the Cox regression model based on baseline BrainAGE scores outperformed all other models based on cognitive scores, even when including the APOE ε4-status into the models (**Figure 7B**). Also, predictions based on baseline BrainAGE scores were significantly more accurate than predictions based on chronological age or cognitive test scores (for exact numbers see **Table 1**), especially in APOE ε4-carriers.

### EFFECTS OF PSYCHIATRIC DISORDERS ON BRAIN AGING

A recent study on the effects of psychiatric disorders on individual brain aging analyzed data from schizophrenia (SZ)

Interestingly, another study by Hajek et al. (39) in young adult patients with early SZ as well as young adult patients with early BD and young adults with familial risk for BD, aged 15– 35 years, resulted in comparable results. Specifically, participants with first-episode SZ showed advanced BrainAGE of 2.6 years as compared to their chronological age (p < 0.001), whereas participants at familial risk for or in the early stages of BD showed no differences between brain age and chronological age as well as compared to controls (p = 0.70). Post-hoc analyses additionally showed that BrainAGE was negatively associated with GM volume diffusely throughout the brain (**Figure 8C**). The authors concluded that the greater presence of neurostructural antecedents may differentiate SZ from BD and that BrainAGE could consequently aid in early differential diagnosis between BD and SZ.

A third study in first-episode SZ investigated whether comorbid obesity or dyslipidemia additionally contributes to brain alterations (40). Comparable to previous studies, young adult participants with first-episode SZ (n = 120; 18–35 years) showed neurostructural alterations, which resulted in their brain age exceeding their chronological age by 2.6 years (p < 0.001). Furthermore, the diagnosis of first-episode SZ and obesity were each additively associated with BrainAGE (p < 0.001), resulting in BrainAGE scores being highest in obese participants with firstepisode SZ (3.8 years) and lowest in normal weight controls (−0.3 years; **Figure 8B**). However, neither dyslipidemia nor medical treatment was associated with BrainAGE. In conclusion, this study suggests obesity being an independent risk factor for diffuse brain alterations, manifesting as advanced brain aging already in the early course of SZ. Thus, targeting metabolic health and intervening at the BMI level might potentially slow brain aging in schizophrenic and psychotic patients.

# EFFECTS OF INDIVIDUAL HEALTH ON BRAIN AGING

## Effects of Type 2 Diabetes Mellitus on Brain Aging

In the study by Franke et al. (41), the BrainAGE method was applied to a sample of participants with type 2 diabetes mellitus (DM2) and CTR participants (mean age: 65 ± 8 years) in order to quantify the effects of DM2 on individual brain aging in cognitively healthy older adults. Participants with DM2 showed significantly increased BrainAGE by 4.6 years as compared to agematched healthy CTRs (p < 0.001). Moreover, longer diabetes duration was correlated to higher BrainAGE scores (r = 0.31, p < 0.05). Additionally, BrainAGE scores were also positively related to fasting blood glucose (r = 0.34, p < 0.05), with a difference of 5.5 years (p < 0.05) between participants with the lowest vs. highest values.

# Longitudinal Effects of Type 2 Diabetes Mellitus on Brain Aging

Additionally, Franke et al. (41) further analyzed a small subsample of DM2 and CTR participants that completed a follow-up MRI scan 3.8 ± 1.5 years after their baseline

patients, bipolar disorder (BD) patients (mostly with previous psychotic symptoms or episodes), as well as CTR participants, aged 21–65 years. Significantly higher BrainAGE scores by 2.6 years were found in SZ, but not BD patients, indicating advanced structural brain aging in SZ (**Figure 8A**). This study thus suggested, that there might be an additional progressive pathogenic component despite the conceptualization of SZ as a neurodevelopmental disorder (38).

below the median (light lines) and above the median (dark lines). Duration of follow-up is truncated at 1,250 days. [Figure and legend reproduced from Loewe et al. (36), permitted under the Creative Commons Attribution License].

assessment. GM and WM volumes did not differ between both groups or between time points. However, BrainAGE scores were increasing by 0.2 years per follow-up year in participants with DM2, but did not change in CTRs during follow-up. Specifically, baseline BrainAGE scores in DM2 patients were increased by 5.1 years as compared to CTR (p < 0.05), they even increased by 0.8 years during follow-up (p < 0.05). Thus, brain aging in DM2 did even more accelerate during follow-up.

### Individual Health and Brain Aging

In addition to the effects of DM2 on individual brain aging in non-demented older adults, the study by Franke et al. (41) also explored the (additional) effects of lifestyle risk factors (i.e., smoking duration, alcohol intake), individual health marker (i.e., hypertension, TNFα), and common clinical outcomes (i.e., cognition, depression). The results revealed BrainAGE being also correlated to smoking duration (r = 0.20, p < 0.01), alcohol consumption (r = 0.24, p < 0.001), TNFα levels (r = 0.29, p < 0.01), verbal fluency (r = −0.25, p < 0.01), and depression (r = 0.23, p < 0.05), but not to hypertension (p = 0.9). Furthermore, contrasting individuals with the lowest values (i.e., 1st quartile) vs. those with the highest values in these measures (i.e., 4th quartile) resulted in BrainAGE differences of 3.4 years for smoking duration (p < 0.01), 4.1 years for alcohol intake (p < 0.01), 5.4 years for TNFα (p < 0.01), 5.6 years for verbal fluency (p < 0.001), and 5.4 years for depression (p < 0.01; **Figure 9A**), with all results being independent of diabetes duration, gender, and age (41).

### Gender-Specific Effects of Health Characteristics on Brain Aging

In a study by Franke et al. (42), the effects of various physiological and clinical markers of personal health on individual BrainAGE scores were further explored and quantified, utilizing a sample of cognitively unimpaired participants, aged 60–90 years.

In the male sample, the included health parameters explained 39% of the observed variance in BrainAGE (p < 0.001), with body mass index (BMI), uric acid, γ-glutamyl-transferase (GGT), and diastolic blood pressure (DBP) contributing most. Additional quartile analyses revealed significant differences in BrainAGE

FIGURE 9 | The effects of low vs. high levels in distinguished variables on BrainAGE. (A) Mean BrainAGE scores in participants with values in the 1st (plain squares) and 4th (filled squares) quartiles of distinguished variables from the diabetes study. [Figure and legend reproduced from Franke et al. (41), permitted under the Creative Commons Attribution License.] (B) Mean BrainAGE scores of cognitively healthy CTR men in the 1st vs. 4th quartiles of the most significant physiological and clinical chemistry parameters (left panel). BrainAGE scores of cognitively healthy CTR men with "healthy" markers (i.e., values below the medians of BMI, DBP, GGT, and uric acid; n = 9) vs. "risky" markers (i.e., values above the medians of BMI, DBP, GGT, and uric acid; n = 14; p < 0.05; right panel). [Figures and legend modified from Franke et al. (42), permitted under the Creative Commons Attribution License.] (C) Mean BrainAGE scores of cognitively healthy CTR women in the 1st vs. 4th quartiles of the most significant physiological and clinical chemistry parameters (left panel). BrainAGE scores of cognitively healthy CTR women with "healthy" markers (i.e., values below the medians of GGT, ALT, AST, and values above the median of vitamin B12; n = 14) vs. "risky" clinical markers (i.e., values above the medians of GGT, ALT, AST, and values below the median of vitamin B12; n = 13; p < 0.05; right panel). [Figures and legend modified from Franke et al. (42), permitted under the Creative Commons Attribution License]. \*p < 0.05; \*\*p < 0.01.

scores between the 1st vs. 4th quartile groups (**Figure 9B**, left panel), resulting in 7.5 years for BMI (p < 0.001), 6.6 years for DBP (p < 0.01), 7.5 years for GGT (p < 0.01), and 5.6 years for uric acid (p < 0.05). When combining these four health markers, the effects on individual BrainAGE even were compounded. In detail, comparing individual brain ages of male subjects with values below the medians vs. those with values above the medians of BMI, DBP, GGT, and uric acid resulted in BrainAGE scores of −8.0 vs. 6.7 years (p < 0.05; **Figure 9B**, right panel), thus suggesting a strong relationship between individual health and neurostructural aging in men.

In the female sample, the included health parameters explained 32% of the observed variance in BrainAGE (p < 0.01), with GGT, aspartat-amino-transferase (AST), alaninamino-transferase (ALT), and vitamin B<sup>12</sup> contributing most. In addition, 1st vs. 4th quartile analyses resulted in differences in BrainAGE (**Figure 9C**, left panel) of 6.6 years for GGT (p < 0.01), 3.1 years for AST (p < 0.10), 5.1 years for ALT (p < 0.05), and 4.8 years for vitamin B<sup>12</sup> (p < 0.05). Again, when combining these four health markers, the effects on individual BrainAGE were compounded, resulting in mean BrainAGE scores of −1.0 vs. 3.8 years (p < 0.05; **Figure 9C**, right panel), thus suggesting a mediocre relationship between individual health and neurostructural aging in women.

### PROTECTING INTERVENTIONS FOR BRAIN AGING

### Effects of Long-Term Meditation Practice on Brain Aging

Exploring the effects of long-term meditation practice, the study by Luders et al. (43) included 50 meditation practitioners with 4– 46 years of meditation experience (mean: 20 ± 11 years) and 50 non-meditating, age-matched CTRs. At age 50 years, BrainAGE in meditation practitioners was about 7.5 years lower than in CTRs (p < 0.05). Additionally, gender exerted a main effect, with BrainAGE in females being lower by 3.4 years as compared to males (p < 0.01). Furthermore, age-by-group interaction was significant (p < 0.05), with follow-up analyses revealing significant effects for BrainAGE in meditation practitioners. In detail, for each year in chronological age over the age of 50 years, there was a significant decrease of 1 month and 22 days in BrainAGE in the meditation practitioners (**Figure 10**).

### Effects of Making Music on Brain Aging

Another study investigated the impact of music-making on brain aging, including non-musicians, amateur musicians, and professional musicians, aged 25 ± 4 years (44). All three groups were closely matched regarding age, gender, education, and other leisure activities. The "musician status" had a significant effect on BrainAGE (p < 0.05; non-musicians: −0.5 ± 6.8 years; amateur musicians: −4.5 ± 5.6 years; professional musicians: −3.7 ± 6.6 years), suggesting a decelerating effect of making music on individual brain aging. Post-hoc comparisons revealed lower BrainAGE scores in amateur musicians (p < 0.05) and professional musicians (p = 0.07) as compared to non-musicians. While no significant correlation between years involved in musical activities and BrainAGE score was found in amateur musicians (r = −0.1, n.s.), a small correlation was found in professional musicians (r = 0.3, p < 0.05). Thus, making music seems to have a slowing effect on the aging of the brain, especially for amateur musicians, while professional musicians revealed a lower effect probably due to stress-related interferences.

### GENDER-SPECIFIC EFFECTS OF PRENATAL UNDERNUTRITION ON INDIVIDUAL BRAIN AGING

### Results From Studies in Humans

Utilizing a subsample of the Dutch famine birth cohort, a recent study investigated the effects of fetal undernutrition during early gestation on individual brain aging in late-life (85). The participants of the MRI subsample were aged about 67 years at the time of MRI acquisition, including individuals being born before the famine in Winter 1944/45, individuals being prenatally exposed to the famine during early gestation, and individuals being conceived after the famine. In females, 28% of the observed variance BrainAGE at age 67 years was explained by birth characteristics, chronological age at MRI data acquisition, and famine exposure (p < 0.05), whereas in males, 76% the observed variance in BrainAGE was explained by the combination of birth characteristics, late-life health characteristics, chronological age, and famine exposure (p < 0.05). In the male sample, BrainAGE scores differed significantly between the three groups (p < 0.05). In the female sample, BrainAGE scores did not differ between the groups. Post-hoc tests in the male sample showed advanced brain aging by 2.5 years (p < 0.05) in those who had been prenatally exposed to the famine during early gestation, whereas those who had been born before the famine showed delayed brain aging by −1.8 years, resulting in a difference of about 4 years (p < 0.05; **Figure 11A**). With regard to BrainAGE scores there were no significant differences between males and females (85).

## Results From Studies in Non-human Primates

An experimental study of maternal nutrient restriction (MNR) in baboons also studied the effects of prenatal undernutrition on structural brain aging based on the baboon-specific BrainAGE model [see Species-specific BrainAGE model for baboons; (33)]. The experimental group included 11 subjects [5 females], with prenatal undernutrition being induced by MNR of 30% during the whole gestation. The CTR group included 12 same-aged subjects [5 females]. Subjects were aged 4–7 years [human equivalent to 14–24 years] at time of MRI data acquisition. In the female MNR offspring, baboon-specific BrainAGE scores were increased by 2.7 years, as compared to female CTR offspring (p = 0.01; **Figure 11B**), strongly suggesting premature brain aging resulting from prenatal undernutrition during the whole gestation. There were no differences in BrainAGE scores between the male MNR and CTR offspring (33).

FIGURE 10 | Group-specific links between age-related measures. Scatterplots and regression lines were generated separately for (A) controls (circles) and (B) meditation practitioners (triangles). The x-axes display the chronological age; the y-axes display the BrainAGE index (negative values indicate that participants' brains were estimated as younger than their chronological age, positive values indicate that participants' brains were estimated as older). [Figures and legend reproduced from Luders et al. (43), with permission from Elsevier].

SUMMARY

least for 36 months); SZ, schizophrenia].

In this review, we recapitulated studies that utilized the innovative BrainAGE biomarker to capture individual agerelated brain structure, covering age ranges from childhood until late adulthood (**Figure 12** for a graphic summary of all results in human studies). This predictive analytical method provides a personalized biomarker of brain structure that can help to elucidate und further examine the patterns and mechanisms underlying individual differences in brain structure and disease states. Because brain-age estimation is done on an individual level, the BrainAGE biomarker might be very well-suited for clinical use. The method is deriving individual predictions from multivariate patterns and interactions between voxels across the whole brain. In contrast to other structural measures, such as regional or global volumes, cortical thickness, or fractional anisotropy, BrainAGE scores are preserving the complex patterns of subtle variations in brain structure and their regional interactions. Additionally, reducing the complex multivariate structural information from the whole brain into a single metric resolves the problem of multiple comparisons and enables a better detection of effects (7, 24).

According to the American Federation of Aging Research (86), markers of aging should possess certain characteristics: They should be able to determine biological aging, predict the rate of aging, monitor the fundamental processes underlying aging, and be measured accurately, efficiently, and repeatedly, without harming the subject. Further, the markers need to be applicable across the species for mechanistic examinations. However, reproducibility and accuracy of some widely used biomarkers of aging, like telomere length, vary widely due to differences in extraction methods, laboratory-dependent methodological details, and measurement methods (87–89). Thus, accuracy is sometimes so low that measurement errors impede detection of differences in telomere length (88). Although biomarkers of aging should preferably be closely related to the mechanistic aging process, development of markers of brain aging that are related to brain function and structure is much more advanced and provide a considerably higher degree of correlation to age and diagnostic specificity. Moreover, brain-aging markers based on structural MRI show less inter-individual variability and methodological variations of measurements across labs or study sites. The superiority of phenotype-related markers may be explained by a number of reasons: At present, it is easier to determine phenotype because the processes underlying brain aging are complex and not yet well-understood. This is all the more so for the many compensatory pathways in the biological environment by which the organism modulates or responds to the process of aging. Aside from the complexity being present at the cellular level, the organism can respond to an infinite number of biological and environmental influences with only limited changes to the phenotype. Consequently, establishing phenotyperelated biomarkers for structural brain maturation and aging (e.g., BrainAGE) might probably be a better approach to assess and longitudinally track individual brain aging trajectories.

In general, cognitive impairment is not due to just one disease. Cognitive impairment could be caused by AD and other forms of dementia, as well as several disease conditions, e.g., traumatic brain injury, stroke, depression, or developmental disabilities. Age-related cognitive decline is a growing concern in modern societies since mental health is perceived as a major determinant limiting quality of life during aging (90). Thus, biomarkers measuring individual brain age and predicting individual trajectory of cognitive decline are highly desirable. Approaches to determine brain age based on structural neuroimaging data are designed to indicate deviations in age-related changes in brain structure by establishing reliable reference curves for healthy brain aging and providing individual brain age measures, while accounting for the multidimensional atrophy patterns in the brain. Although multiple factors affect and modify individual brain aging trajectories, normal brain aging follows coordinated and sequenced patterns of GM and WM loss as well as CSF expansion (21, 91, 92). Several studies applying the MRI-based models for structural brain aging, have already demonstrated profound relationships between premature brain aging and AD disease severity and prospective decline of cognitive functions (45), MCI and AD (93), conversion to AD (37), SZ (76, 94), traumatic brain injury (73), HIV (95), chronic pain (96), DM2 (41), and elderly people suffering from undernutrition during gestation (85), as well as being indicative of poorer physical and mental fitness, higher allostatic load, as well as increased mortality (97). Furthermore, significant associations between individual brain aging and various health parameters, personal lifestyle, or drug use (42, 98), levels of education and physical activity (77), and meditation practice (43) have been shown. However, although Brown et al. (59) showed a relation between increased premature brain maturation and increased executive

### REFERENCES

1. Vos T, Flaxman AD, Naghavi M, Lozano R, Michaud C, Ezzati M, et al. Years lived with disability (YLDs) for 1160 sequelae of 289 diseases and injuries 1990-2010: a systematic analysis for the Global Burden of Disease Study 2010. Lancet. (2012) 380:2163–96. doi: 10.1016/S0140-6736(12)61729-2

intelligence measures in adolescents as well as Steffener et al. (77) showing a correlation between delayed brain aging and higher education levels in adults, this issue has to be explored in more depth with well-characterized and well-tested samples with regards to cognitive reserve and IQ levels.

In conclusion, the phenotypic approach presented here has already established and validated reference curves for age-related changes in brain structure. Furthermore, it also showed great potential for easy application in multi-center studies. Thus, this predictive analytical method provides an individualized biomarker for determining the biological age of brain structure, which also relates to cognitive function. This MRI-based marker is able to predict individual aberrations in brain maturation and aging as well as the occurrence of age-related cognitive decline and age-related neurodegenerative diseases. This review has recapitulated evidence that neuroimaging data can be used to establish biomarkers for brain aging, which has already been confirmed as providing vital prognostic information. In future, combining different biomarkers of structural and functional brain age, like the assessment of age-related changes of parameter estimates based on the "theory of visual attention" (99–103), may enhance sensitivity and specificity for detecting aberrations in biological age compared to the chronological age in various neurological and psychiatric conditions and in neurodegenerative diseases. The important prognostic information included in the estimation of the structural and functional brain age may aid in developing personalized neuroprotective treatments and interventions.

### AUTHOR CONTRIBUTIONS

All authors listed have made a substantial, direct and intellectual contribution to the work, and approved it for publication.

### FUNDING

This work was supported by the European Community [FP7 HEALTH, Project 279281 (BrainAGE) to KF] and the German Research Foundation [DFG; Project FR 3709/1-1 to KF]. The sponsors had no role in the design and conduct of the study; collection, management, analysis, and interpretation of the data; and preparation, review, or approval of the manuscript.

### SUPPLEMENTARY MATERIAL

The Supplementary Material for this article can be found online at: https://www.frontiersin.org/articles/10.3389/fneur. 2019.00789/full#supplementary-material


and white matter microstructure. Neuroimage. (2012) 63:365–80. doi: 10.1016/j.neuroimage.2012.06.038


for epidemiologic research. Cancer Epidemiol Biomarkers Prev. (2013) 22:2047–54. doi: 10.1158/1055-9965.EPI-13-0409


**Conflict of Interest Statement:** The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

Copyright © 2019 Franke and Gaser. This is an open-access article distributed under the terms of the Creative Commons Attribution License (CC BY). The use, distribution or reproduction in other forums is permitted, provided the original author(s) and the copyright owner(s) are credited and that the original publication in this journal is cited, in accordance with accepted academic practice. No use, distribution or reproduction is permitted which does not comply with these terms.

# Hemodynamic Surveillance of Unilateral Carotid Artery Stenting in Patients With or Without Contralateral Carotid Occlusion by TCD/TCCD in the Early Stage Following Procedure

### Edited by:

*Hongyu An, Washington University in St. Louis, United States*

### Reviewed by:

*Marek Czosnyka, University of Cambridge, United Kingdom Georgios Tsivgoulis, National and Kapodistrian University of Athens, Greece*

> \*Correspondence: *Min Yang dryangmin@gmail.com*

*†These authors have contributed equally to this work*

### Specialty section:

*This article was submitted to Applied Neuroimaging, a section of the journal Frontiers in Neurology*

Received: *27 February 2019* Accepted: *20 August 2019* Published: *04 September 2019*

### Citation:

*Yan Z, Yang M, Niu G, Zhang B, Tong X, Guo H and Zou Y (2019) Hemodynamic Surveillance of Unilateral Carotid Artery Stenting in Patients With or Without Contralateral Carotid Occlusion by TCD/TCCD in the Early Stage Following Procedure. Front. Neurol. 10:958. doi: 10.3389/fneur.2019.00958* Ziguang Yan† , Min Yang\* † , Guochen Niu, Bihui Zhang, Xiaoqiang Tong, Hongjie Guo and Yinghua Zou

*Department of Interventional Radiology and Vascular Surgery, Peking University First Hospital, Beijing, China*

Objective: To evaluate the cerebral hemodynamic variations in patients with unilateral carotid artery stenosis and contralateral carotid occlusion (CCO) in hours following carotid artery stenting (CAS) by transcranial Doppler (TCD) or transcranial color-code Doppler (TCCD).

Methods: Sixty-five consecutive patients who underwent unilateral CAS were enrolled. Among them, 14 patients had ipsilateral severe stenosis and CCO (CCO group) while the other 51 patients had only unilateral severe carotid stenosis without CCO (UCS group). All patients underwent TCD or TCCD monitoring before, at 1 and 3 h after CAS. We monitored bilateral middle cerebral artery (MCA) peak systolic velocity (PSV), pulsatility index (PI), and blood pressure (BP), and compared that data between two groups.

Results: In UCS group, ipsilateral MCA PSV increased relative to baseline at 1 h (96 ± 30 vs. 85 ± 26 cm/s, 15%, *P* < 0.001) and 3 h (97 ± 29 vs. 85 ± 26 cm/s, 17%, *P* < 0.001) following CAS. Significant PI increases were observed at 1 and 3 h following CAS on the ipsilateral side. In CCO group, ipsilateral MCA PSV increased relative to baseline at 1 h (111 ± 30 vs. 83 ± 26 cm/s, 35%, *P* < 0.001) and 3 h (107 ± 28 vs. 83 ± 26 cm/s, 32%, *P* <0.001) following CAS. The magnitude of ipsilateral MCA PSV increase was significantly higher in CCO group compared with UCS group at 1 h (*P* = 0.002) and 3 h (*P* = 0.024) following CAS, while BP similarly decreased between the two groups. On the contralateral side, significant MCA PSV increases were observed following CAS in CCO group but not in UCS group. Bilateral MCA PSV increases were higher in patients with a stenosis degree of ≥90% than in patients with stenosis degree of 70–89% only in CCO group.

Conclusion: The ipsilateral MCA PSV and PI increase moderately in the initial hours after unilateral CAS in patients without CCO. In patients with CCO, the ipsilateral, and contralateral MCA PSV increase significantly in the early stage following CAS. CCO is a factor of the increased blood flow velocity in ipsilateral MCA after unilateral CAS.

Keywords: carotid artery stenosis, contralateral carotid occlusion, carotid artery stenting, transcranial Doppler, transcranial color-code Doppler, cerebral hemodynamics, early stage

### INTRODUCTION

Contralateral carotid artery occlusion (CCO) was found in 5– 15% of carotid artery stenosis (CS) patients (1–4). According to the North American Symptomatic Carotid Endarterectomy Trial (NASCET), CCO has been demonstrated as an independent risk factor for carotid endarterectomy CEA (1, 2, 5, 6). While, carotid artery stenting (CAS) is suggested as an alternative for the treatment of patients with CS and CCO (2, 3). A recent meta-analysis about cerebral hyperperfusion syndrome (CHS) encouraged further investigation on cerebral hemodynamic monitoring (7). Besides, a crucial risk factor of periprocedural stroke following CAS is hemodynamic disturbance (HD), which often occurs within 6 h after CAS (8–11). However, only a few studies have evaluated cerebral hemodynamic changes in the early stage following CAS in patients with CCO. Transcranial Doppler (TCD) and transcranial color-code Doppler (TCCD) are bedside examinations and can be used for routine clinical monitoring of cerebral hemodynamic changes immediately after CAS (12). Our study used TCD and TCCD to assess the immediate effect on cerebral hemodynamics after CAS in patients with and without CCO.

### MATERIALS AND METHODS

### Subjects

All patients who underwent CAS in Department of Interventional Radiology and Vascular Surgery at Peking University First Hospital from Jan, 2013 to Dec, 2018 were enrolled in this study. One hundred forty-eight patients underwent CAS, of whom 27 patients had no bone window. TCD were performed in 121 patients and 56 of them were excluded because of simultaneous bilateral carotid stenting (nine patients), simultaneous vertebral or subclavian artery stenting (16 patients), carotid artery near occlusion (20 patients), or moderate-severe contralateral carotid artery stenosis (11 patients). Carotid stenosis was diagnosed using ultrasound and computed tomography angiography (CTA), and finally Four-vessel angiography. Among all the remaining 65 patients, 14 patients were diagnosed severe CS with CCO, 51 patients had severe unilateral CS.

### CAS Protocol

CAS was performed in symptomatic (at least 2 weeks after onset of symptom) or asymptomatic patients with >70% stenosis (NASCET criteria). Written informed consent was obtained from all of the patients that underwent CAS. At least 72 h before the procedure, all patients received antithrombotic premedication (100 mg aspirin and 75 mg clopidogrel). Transbrachial approach was used in one patient because of aortic-iliac artery occlusion. Transfemoral approach with local anesthesia using 2% lidocaine was used in all the other cases. Distal embolic protection device was used in all the patients. We routinely applied pre-dilation with a 4.0–5.0 mm balloon catheter (Boston Scientific, Natick, MA), and selected the appropriate stent device (Precise RX, Cordis Endovascular; Acculink, Abbott Vascular; and Carotid Wallstent, Boston Scientific) according to the anatomic location and the diameter of the artery at the operater's discretion. We would not perform post-dilation unless the residual stenosis was more than 30%. The completion angiogram of carotid artery and distal cerebral vasculature was performed after stent deployment (**Figure 1**).

### Transcranial Doppler

Examination was performed using a 2-MHz probe connected to a TCD machine (TC2021, EME, Companion III, Germany) or a transcranial color-code Doppler (TCCD) machine (GE LOGIOe) fitted with 2.0-MHz sector array transducer. The ipsilateral and/or contralateral middle cerebral artery (MCA) was insonated through the temporal window at a depth of 46–60 mm. We recorded peak systolic velocity (PSV) and pulsatility index (PI) at baseline on the day before CAS, and again at about 1 and 3 h following the CAS procedure. To maintain a constant depth, angle of insonation, and an original probe-skin contact point (**Figure 1**), all TCD or TCCD examinations in the patients were performed by an identical physician. Post-CAS hyperperfusion was defined as the MCA-PSV exceeded 2-fold of the pre-CAS TCD measurement (13, 14).

## Blood Pressure Control

Blood pressure (BP) was monitored and controlled throughout the periprocedure period. Before balloon predilation, systolic BP was controlled below 160 mmHg. After predilation and stent deployment, systolic blood pressure was preliminarily controlled between 90 and 140 mmHg for unilateral CAS patients. If potential hyperperfusion or hypoperfusion were detected by the first TCD, BP would be further adjusted. Hemodynamic depression (HD) was defined as periprocedural hypotension (BP < 90/60 mmHg) or bradycardia (heart rate < 50 beats/min). Persistant HD was defined as HD persisted for at least 1 h. Dopamine or/and atropine were used for HD patients. Urapidil or/and nicardipine were administered intravenously to lower BP, which was measured during the examination using a standard BP cuff.

FIGURE 1 | A patient with carotid artery stenosis and contralateral carotid occlusion (CCO) underwent carotid artery stenosis (CAS), and periprocedural transcranial color-code Doppler (TCCD) monitoring. (A) Digital subtraction angiography (DSA) showing right internal carotid artery (ICA) occlusion (black arrow) and right external carotid artery supplying right middle cerebral artery (MCA) via collateral circulation of ophthalmic artery. (B) DSA showing left ICA severe stenosis (arrow). (C) DSA showing left ICA supplying right anterior cerebral artery (ACA) via anterior communicating artery. (D) TCCD before CAS showing left MCA peak systolic velocity (PSV) was 94 cm/s, while systolic blood pressure (SBP) was 155 mmHg. (E) DSA showing left ICA following CAS. (F) DSA showing left ICA supplying right ACA and MCA via anterior communicating artery following CAS. (G) TCCD at 1 h after CAS maintained a constant depth, angle of insonation, and an original probe-skin contact point, showing left MCA PSV was 133 cm/s, while SBP was 130 mmHg.

## Statistical Analysis

We performed all statistical analyses using IBM SPSS software (version 23.0). TCD data are presented as mean ± standard deviation (SD). PSV and PI values the day before CAS, and at both 1 and 3 h following CAS were evaluated using paired t-test, after repeated measure ANOVA. Bonferroni correction was used, and statistical significance was considered to be P < 0.05/3 (=0.0167). Variations between groups were compared using independent t-test and P < 0.05 was considered statistically significant.

### Study Approval

The protocol for this study was approved by the institutional review board at the Peking University First Hospital in accordance with the Chinese clinical research ethics guidelines. All data were obtained from the Peking University First Hospital, Department of Interventional Radiology and Vascular Surgery, after anonymization.

### RESULTS

All CAS procedures were successful and without adverse events. Among the 65 patients enrolled, 14 patients had ipsilateral severe stenosis and CCO (CCO group), the other 51 patients had only unilateral severe carotid stenosis without CCO (UCS group). The mean (±SD) age of UCS group was 66 ± 8 years. Of these patients, 24 (47%) were symptomatic, while the remaining 27 patients (53%) were asymptomatic. Forty-two patients (82%) of UCS group were male. The average degree of ICA stenosis of UCS group was 82 ± 8%. The mean (±SD) age of CCO group was 67 ± 7 years. Of these patients, 10 (71%) were symptomatic, while the remaining four patients (29%) were asymptomatic. Twelve patients (86%) of CCO group were male. The average degree of ICA stenosis of CCO group was 81 ± 11%. Angiography showed opened anterior communicating branch in all the CCO patients. Contralateral MCA was supplied by anterior communicating branch in four patients before CAS and in six patients after CAS.

The demographic data are shown in **Table 1**. Three different types of stent were used in both groups. There were no instances of severe hyperperfusion syndrome, renal failure, deaths or disabling strokes in any of the participants in the month following CAS. Three patients in UCS group had minor stroke in the early phase following CAS. Four patients in UCS group and two patients in CCO group had persistent HD, which we treated with dopamine during the 24-h period following CAS (**Table 1**). In both groups, the mean BP decreased after CAS. The mean BP values did not significantly differ between the two groups, either at baseline or post-CAS.

TCD examinations were performed in all the 65 patients before CAS, and at 1 and 3 h after CAS. Among them, three patients in UCS group and two patients in CCO group received only ipsilateral TCD examination because of unilateral absence of bone window, or contralateral MCA occlusion. In UCS group, at 1 h after CAS, TCD showed a significant PSV increase in the



*NO, near occlusion; ICA, internal carotid artery; iMCA, ipsilateral middle cerebral artery; PSV, peak systolic velocity; CAS, carotid artery stenting; HD, hemodynamic depression. P* < *0.05 was considered statistically significant.*

TABLE 2 | Parameters of hemodynamic changes in UCS group.


*CAS, carotid artery stenting; BP, blood pressure; iMCA, ipsilateral middle cerebral artery; PSV, peak systolic velocity; PI, pulsatility index; cMCA, contralateral middle cerebral artery. P* < *0.017 (after Bonferroni correction) was considered statistically significant.*

ipsilateral MCA (from 85 ± 26 to 96 ± 30 cm/s, 15%, P < 0.001). The average PI also increased in the ipsilateral MCA (from 0.85 ± 0.16 to 0.94 ± 0.24, P = 0.003). At 3 h after CAS, the PSV in the ipsilateral MCA was also significantly increased compared to the value before CAS (from 85 ± 26 to 97 ± 29 cm/s, 17%, P <0.001), but similar to the value 1 h after CAS (P = 0.514). A significant PI increase was observed 3 h after CAS (from 0.85 ± 0.16 to 1.0 ± 0.25, P < 0.001). On the contralateral side, there was no significant PSV or PI increase in the MCA for either 1 or 3 h after CAS (**Table 2**).

In CCO group, at 1 h after CAS, TCD showed a significant PSV increase in the ipsilateral MCA (from 83 ± 26 to 111 ± 30 cm/s, 35%, P < 0.001). At 3 h after CAS, the PSV value in the ipsilateral MCA was also significantly increased compared to prior CAS (from 83 ± 26 to 107 ± 28 cm/s, 32%, P < 0.001), TABLE 3 | Parameters of hemodynamic changes in CCO group.


*CAS, carotid artery stenting; BP, blood pressure; iMCA, ipsilateral middle cerebral artery; PSV, peak systolic velocity; PI, pulsatility index; cMCA, contralateral middle cerebral artery. P* < *0.017 (after Bonferroni correction) was considered statistically significant.*

TABLE 4 | Increase rate of ipsilateral MCA PSV following CAS in UCS group and CCO group.


*CAS, carotid artery stenting; BP, blood pressure; iMCA, ipsilateral middle cerebral artery; PSV, peak systolic velocity; PI, pulsatility index.*

*P* < *0.05 was considered statistically significant.*

but similar to the value at 1 h after CAS (P = 0.144). There was no significant PI increase in the ipsilateral MCA for either 1 or 3 h after CAS. On the contralateral side, the MCA PSV increased in 1 h after CAS (69 ± 16 vs. 90 ± 29, 28%, P = 0.001) and 3 h after CAS (69 ± 16 vs. 86 ± 29, 22%, P = 0.005) compared with the value before CAS. There was no significant PI increase in the contralateral MCA for either 1 or 3 h after CAS (**Table 3**).

The increase rate of BP had no significant difference at 1 or 3 h after CAS between the two groups. There was no significant difference of the average pre-CAS ipsilateral MCA PSV between the two groups (P = 0. 829). The magnitude of ipsilateral MCA PSV increases in CCO group significantly exceeded that observed in UCS group at both 1 h after CAS (35 vs. 15%, P = 0.002), and 3 h after CAS (32 vs. 17%, P = 0.024; **Table 4**). In CCO group, five patients had a ≥90% stenosis degree. In these patients, the magnitude of ipsilateral MCA PSV increase was 53 ± 17% at 1 h and 52 ± 21% at 3 h after CAS, significantly higher than the magnitude of 26 ± 11% (P = 0.004) at 1 h and 21 ± 19% (P = 0.018) at 3 h in the other nine patients. In UCS group, at 1 or 3 h after CAS, the magnitude of ipsilateral MCA PSV increase had no statistically significant difference whether stenosis degree was ≥90% (**Table 4**). In both groups, the magnitude of ipsilateral MCA PSV increase had no significant difference with the varied type of Willis circle, whether the patients were ≥70 years old or whether the patients were asymptomatic (data not shown).

# DISCUSSION

Patients with CS and CCO carry a higher incidence of complication following CEA and CAS (15). A previous metaanalysis recommended CAS, rather than CEA in patients with CCO (2). HD and CHS are two different complications of CAS related to cerebral hemodynamic changes, both may occur within 6 h following CAS (16–19). However, only few studies have focused on cerebral hemodynamic changes in the early stage following CAS, especially in the patients with CCO. The present research clarified the changes of bilateral MCA PSV in the early stage after unilateral CAS in patients with or without CCO.

A previous research demonstrated an about 20% increase of the ipsilateral MCA PSV in the early stage following CAS (12). However, sample in that research had some extent heterogeneity. The present research excluded several potential risk factors, such as simultaneous bilateral carotid stenting, simultaneous vertebral or subclavian artery stenting, carotid artery near occlusion, or contralateral carotid artery stenosis (12, 20). Therefore, the variation of cerebral blood flow velocity after CAS in patients with simple unilateral carotid artery stenosis could be observed for the first time. Meanwhile, the influence of CCO on MCA PSV change after unilateral CAS could be demonstrated more clearly. Concerning the changes of PSV, the previous research stated that there were no significant differences between patients with ≥90% stenosis and those with 70–89% stenosis. The present research shows that although in UCS group the increment in ipsilateral MCA PSV in patients with ≥90% stenosis is greater, there is still no statistical significance. In CCO group, however, it is observed that ipsilateral MCA PSV increased significantly higher in patients with a ≥90% stenosis, which might be attributed to the impaired cerebral hemodynamic autoregulation.

Following CAS, there is a 3.1–6.8% risk of CHS, that most likely occurs in the early post-procedural period (7). Abou-Chebl et al. (11) has suggested that patients with severe bilateral carotid stenosis were predisposed to CHS, and patients with CCO should require more intensive hemodynamic monitoring after CAS. However, in the present study, no patient had more than 100% increase of the MCA PSV following the procedure and none CHS occurred. The increase of ipsilateral MCA PSV was at an average of 35 and 32% at 1 and 3 h following CAS, respectively. The maximum magnitude of MCA PSV increase was 84% in the ipsilateral side and 67% in the contralateral side. These results suggest that for patients with CCO, under a strict BP control and cerebral hemodynamic monitoring after CAS, the risk for CHS can be reduced.

Regional cerebral blood flow is proportional to blood flow velocity in the MCA (21, 22). A previous research measured cerebral blood flow by SPECT within 2 h following CAS in patients with CCO (23). In that research, no significant difference was found in resting cerebral blood flow in both hemispheres immediately after CAS, which differed from the present research. Besides, the previous research did not include comparisons with a control group. To our knowledge, there are no other study focus on the immediate cerebral hemodynamic changes in CCO patients following CAS.

In the control group, there were only a little bit more than 15% average increase of ipsilateral MCA PSV at 1 and 3 h following procedure, perhaps due to a relatively normal cerebral autoregulation (24). In this article, we analyzed not only PSV but PI. Increase of PI indicates that the waveform becomes steeper. The PI is not dependent solely on cerebrovascular resistance but a product of the interplay between cerebral perfusion pressure, pulse amplitude of arterial pressure, cerebrovascular resistance and compliance of the cerebral arterial bed as well as the heart rate (25). Notably, PI increased significantly in the ipsilateral MCA following CAS in UCS group. This finding reveals that vasoconstriction of resistance arterioles can accommodate the substantially increased MCA blood flow that follows CAS (18, 25, 26). It is probably because CCO could reduce the cerebral vascular reactivity and the cerebral perfusion reserve (27–29), no PI changes were found in CCO group. Hence the increases of bilateral MCA PSV as well as the cerebral blood was greater than that of patients without CCO.

The present study did not include some parameters such as intracranial pressure or cerebrovascular reactivity. Only to measure the MCA velocities can facilitate the TCD examination and ensure the data of all the patients could be collected on time. Some medications, such as statins, vasopressor or antihypertensives, may have an impact on cerebral circulation (30). The potential confounding role of these medications will be studied in future researches. There were two limitations in the present research. First was the limited sample size. The present research observed greater increases of ipsilateral MCA PSV in patients with an original stenosis degree of ≥90%. However, it needs further confirmation by future large sample study. The second limitation was the gender imbalance. This was because TCD or TCCD were not feasible in patients with a poor temporal window, and female accounted for a high incidence.

# CONCLUSIONS

In patients with unilateral severe carotid stenosis and without CCO, the ipsilateral MCA PSV and PI increase moderately in the initial hours after unilateral CAS. In patients with CCO, the ipsilateral and contralateral MCA PSV significantly increase in the early stage following CAS. The MCA PSV of both sides may increase more in CCO patients with an original stenosis degree of ≥90%. CCO is a factor of the increased blood flow velocity in ipsilateral MCA after unilateral CAS.

# DATA AVAILABILITY

The datasets generated for this study are available on request to the corresponding author.

### ETHICS STATEMENT

The protocol for this study was approved by the institutional review board at the Peking University First Hospital in accordance with the Chinese clinical research ethics guidelines. All data were obtained from the Peking University First Hospital, Department of Interventional Radiology and Vascular Surgery, after anonymization. Informed consent was obtained from all of the patients that underwent CAS.

### AUTHOR CONTRIBUTIONS

ZY and MY: conception or design of the work, drafting and critical revision of the article, and final approval of the

### REFERENCES


version to be published. GN, BZ, XT, and HG: data collection, data analysis, and interpretation. YZ: conception or design of the work.

### FUNDING

This work was funded by Scientific Research Seed Fund of Peking University First Hospital (2019SF038).


reserve and predicting long-term prognosis in patients with internal carotid artery occlusion. Neurosurgery. (1993) 32:912–18. doi: 10.1097/00006123–199306000–00005


**Conflict of Interest Statement:** The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

Copyright © 2019 Yan, Yang, Niu, Zhang, Tong, Guo and Zou. This is an open-access article distributed under the terms of the Creative Commons Attribution License (CC BY). The use, distribution or reproduction in other forums is permitted, provided the original author(s) and the copyright owner(s) are credited and that the original publication in this journal is cited, in accordance with accepted academic practice. No use, distribution or reproduction is permitted which does not comply with these terms.