MRI Insights Into Adolescent Neurocircuitry—A Vision for the Future

Adolescence is the time of onset of many psychiatric disorders. Half of pediatric patients present with comorbid psychiatric disorders that complicate both their medical and psychiatric care. Currently, diagnosis and treatment decisions are based on symptoms. The field urgently needs brain-based diagnosis and personalized care. Neuroimaging can shed light on how aberrations in brain circuits might underlie psychiatric disorders and their development in adolescents. In this perspective article, we summarize recent MRI literature that provides insights into development of psychiatric disorders in adolescents. We specifically focus on studies of brain structural and functional connectivity. Ninety-six included studies demonstrate the potential of MRI to assess psychiatrically relevant constructs, diagnose psychiatric disorders, predict their development or predict response to treatment. Limitations of the included studies are discussed, and recommendations for future research are offered. We also present a vision for the role that neuroimaging may play in pediatrics and primary care in the future: a routine neuropsychological and neuropsychiatric imaging (NPPI) protocol for adolescent patients, which would include a 30-min brain scan, a quality control and safety read of the scan, followed by computer-based calculation of the structural and functional brain network metrics that can be compared to the normative data by the pediatrician. We also perform a cost-benefit analysis to support this vision and provide a roadmap of the steps required for this vision to be implemented.


INTRODUCTION
Adolescence is the time of onset of many psychiatric disorders. Half of pediatric patients present with comorbid psychiatric disorders that complicate both their medical and psychiatric care. Currently, diagnosis and treatment decisions are based on symptoms. The field urgently needs brain-based diagnosis and personalized care.
Neuroimaging can shed light on how aberrations in brain circuits might underlie psychiatric disorders and their development in adolescents. MRI has become the leading modality for mapping the human brain non-invasively. Apart from mapping individual brain regions, the importance of connections between these regions and their role within the brain network as a whole are becoming increasingly recognized and studied within the framework of connectomics (Sporns et al., 2005). Different MRI techniques can be applied to map the connections of the network and to quantify the connectivity strength. Most commonly, connections are derived from diffusion-weighted images (with tractography used to model white matter pathways) or from the functional MRI (fMRI) signal (using temporal correlation of the signal as a proxy for connectivity). Connections can be compared between subjects individually. However, one can also utilize graph theory, which offers new ways to perform network characterization and comparison. Graph theory operates with an abstracted notion of a graph, which is defined as a set of nodes (in our case, brain regions), connected by a set of edges (e.g., white matter tracts). Important network characteristics can be extracted, such as node degree, characteristic path length, average clustering coefficient and other quantifiable measures of network connectivity (Rubinov and Sporns, 2010). Studying the human connectome using graph theory offers a unique opportunity to better understand inter-individual differences in the neural circuitry.
MRI connectomics has been applied to both the adult and developing brain (Hagmann et al., 2010), including extensive work by our group (Tymofiyeva et al., 2012(Tymofiyeva et al., , 2014Ziv et al., 2013). This framework also has been applied to study the neural signature of psychiatric disorders, for example, adult depression (Bai et al., 2012;Korgaonkar et al., 2014;Qin et al., 2014;Gong and He, 2015;Sacchet et al., 2016), as well as adolescent depression (Ellis et al., 2017;Tymofiyeva et al., 2017) and anxiety (Sharp and Telzer, 2017). In the next section, we systematically review both the structural and functional connectivity literature on psychiatric disorders and related symptoms in adolescents.

REVIEW OF CONNECTIVITY STUDIES IN ADOLESCENTS
To perform a review of MRI literature over the last 5-years that provides insights into development of psychiatric disorders in adolescents, the electronic database PubMed was searched using the following Boolean search term, applied to titles and abstracts: (MRI OR fMRI OR DTI) AND (adolescent OR youth) AND (psychiatric OR neurologic OR mental OR depression OR autism OR anxiety OR PTSD OR psychosis OR ADHD OR attention OR bipolar OR schizophrenia OR OCD) AND (connectivity OR connectome OR network OR circuit).
We thus focused on eight disorders: major depressive disorder (MDD), autism, anxiety, bipolar disorder, attentiondeficit/hyperactivity disorder (ADHD), post-traumatic stress disorder (PTSD), schizophrenia, and obsessive-compulsive disorder (OCD) or their relevant clinical and non-clinical symptoms in youth. We defined the age of adolescence as age between 10 and 19, whereas studies could also include older subjects in addition to those between 10 and 19.
The initial search resulted in 177 entries. After excluding articles that: (a) are not in English (1 article), (b) animal studies (5); (c) review articles (8), (d) focus on non-general populations such as Down syndrome (37), (e) do not include any pathology focus (11), (f) do not analyze brain connectivity (21), the resulting set comprised 94 articles. Two additional articles meeting the eligibility criteria were identified through other sources. The total number of articles included in qualitative synthesis was 96. The search results and main finding of the included articles are summarized in Table 1.
Of the 96 included articles, 19 investigated constructs associated with mental illness such as increased rumination, decreased resilience, sensitivity to loss, increased MDD symptom expression, social anxiety, decreased mindfulness, hyperactivity, inattention, anhedonia, etc., without explicitly studying DSMdiagnoses in adolescents. We find the contribution of these studies relevant to the topic because of the limitation of the current diagnostic system and potential of alternative approaches such as the National Institute of Mental Health (NIMH) Research Domain Criteria (RDoC), as will be discussed in the following section. Arguably the most important implication of these neuroimaging findings is that they provide both the motivation and the rationale to pursue a much more explicitly preventative psychiatry approach in helping atrisk children before they present with a manifest psychiatric disorder (McCrory et al., 2017). Developing a neurocognitively informed screening tool capable of accurately indexing latent vulnerability is essential if we are to identify those children who are not yet overtly symptomatic but who are at most risk for the development of a future psychiatric disorder. More broadly, by establishing a better understanding of the specific neurocognitive mechanisms implicated in the pathogenesis of psychiatric disorders, we will be in a much better position to develop effective preventative interventions that increase the likelihood of resilient outcomes in children and adolescents (McCrory et al., 2017).
To graphically synthesize the findings of the reviewed literature, we have mapped the circuitry implicated in eight psychiatric disorders (MDD, autism, anxiety, bipolar disorder, schizophrenia, ADHD, PTSD, OCD) or their relevant nonclinical symptoms in youth onto a single brain model (Figure 1). Figure 1 illustrates both increased and decreased structural and  Classification between individuals with disorders and healthy controls was achieved with high accuracy. Shared atypical brain connections contributing to classification were mostly present in the DMN and SN. These functional connections were significantly associated with severity of social deficits in ASD. Distinct atypical connections were also more related to the DMN and SN, but showed different atypical connectivity patterns between the two disorders Chuang et al. Group difference Pre-treatment (trauma-focused CBT) individual differences in modularity, assortativity, and global efficiency during covert fear vs. neutral task blocks predicted PTSD symptom reduction. PTSD patients with greater treatment response showed greater network modularity and assortativity but lesser efficiency. At the group level, greater symptom reduction was associated with greater pre-to-post-treatment increases in network assortativity and modularity, and this was more pronounced among participants with less symptom improvement Cisler et al.

Diagnosis
Group difference Whole-brain analyses showed significant interactions of anxiety, age, and attention task (threat appraisal, explicit threat memory, physical discrimination) on left amygdala functional connectivity with the ventral medial PFC and ventral ACC. During threat appraisal and explicit threat memory, anxious youth showed more negative amygdala-PFC coupling, whereas anxious adults showed more positive coupling Green et al. (2017) 38 9-17 Autism Exposure-based fMRI (mildly aversive auditory and tactile stimuli)

Diagnosis
Group difference Compared to healthy controls, ASD subjects showed aberrant modulation of fc between pulvinar and cortex (including sensory-motor and prefrontal regions) during sensory stimulation. Among ASD subjects, pulvinar-amygdala connectivity was correlated with severity of sensory over-responsivity symptoms Green et al.

Diagnosis
Group difference Sensory over-responsivity in youth with ASD was related to increased rsFC between SN nodes and brain regions implicated in primary sensory processing and attention. The strength of this connectivity at rest was related to the extent of brain activity in response to auditory and tactile stimuli Gruner et al.

Diagnosis
Group difference Compared to healthy controls, adolescents with depression showed significantly increased sgACC-amygdala functional connectivity and decreased sgACC-fusiform gyrus, sgACC-precuneus, sgACC-insula, and sgACC-middle frontal gyrus functional connectivity. Among depressed adolescents, sgACC-precuneus fc was significantly negatively correlated with depression severity. Compared to healthy controls, depressed adolescents showed poorer perceptual sensitivity in the task, and individual differences in perceptual sensitivity significantly correlated with sgACC fc and depression scores Hong et al. Group difference Among youth with ADHD, higher emotional lability ratings were associated with greater positive intrinsic fc between the amygdala and rostral ACC, and negatively associated with intrinsic fc between bilateral amygdala and posterior insula/superior temporal gyrus. Patterns of amygdala-cortical intrinsic fc in ADHD youth with low emotional lability did not differ from the comparison group, and the effect sizes for these comparisons were smaller than those for the trend-level differences observed between the high emotional lability group and the healthy control group Hwang et al.

Depression
Resting state fMRI Treatment effects Group difference This study was a pilot randomized control trial investigating the effects of rumination-focused cognitive behavior therapy (RFCBT) on reducing rumination and residual depressive symptoms in adolescents with a history of depression and risk of relapse. Over the course of 8 weeks, adolescents who received RFCBT showed significantly reduced rumination and self-reported depression, as well as significant decreases in connectivity between the left PCC and right inferior frontal gyrus and bilateral inferior temporal gyri. Degree of change in connectivity was correlated with changes in self-reported depression and rumination Jacobs et al. Group difference Compared to healthy controls, patients with adolescent-onset schizophrenia showed generalized cognitive impairment with specific deficits in verbal learning and memory, and in processing speed. These measures correlated positively with dorsolateral PFC connectivity with the striatum, and with IQ. DTI voxel-wise comparisons showed lower connectivity between striatum and the motor and lateral orbitofrontal cortices bilaterally, the left amygdala-hippocampal complex, right ACC, left medial orbitofrontal cortex, and right dorsolateral PFC in adolescent-onset schizophrenia patients relative to healthy controls Jann et al.

Autism
Resting state fMRI Diagnosis Group difference Increased local fc in the anterior module of the DMN was accompanied by decreased cerebral blood flow in the same area. Both alterations were associated with greater social impairments. While fc was correlated with cerebral blood flow in healthy controls, this association was disrupted in ASD youth. Reduced long-range fc between the anterior and posterior modules of the DMN was also found in ASD youth Jarcho et al.  Resting state fMRI Future risk Group difference Trait mindfulness in adolescents relates to dynamic but not static rsFC. More mindful youth transitioned more between brain states over the course of the scan, spent overall less time in a certain connectivity state, and showed a state-specific reduction in connectivity between the SN and CEN. The number of state transitions mediated the link between higher mindfulness and lower anxiety Frontiers in Human Neuroscience | www.frontiersin.org

Diagnosis
Group difference In the normative dataset, good sustained attention was characterized by stronger negative fc between cerebellum and motor networks, while stronger positive fc within the motor network was a signature of poorer sustained attention. In separate samples, relative to controls, adolescents with ADHD symptoms had significantly higher intra-individual response variability (IRV) and stronger positive connectivity within low sustained attention networks associated with high IRV, as well as stronger positive connectivity within good sustained attention networks associated with low IRV. There were no differences between the groups for anti-correlated connections in networks associated with either high or low IRV Ordaz et al.

Diagnosis
Group difference Among clinically anxious adolescents completing a dot-probe ask to assess vigilance to threat, vigilance toward threat was positively associated with self-reported distraction and suppression. Fc between a right amygdala seed region and dorsomedial and right dorsolateral PFC regions was inversely related to self-reported suppression and distraction, and dorsolateral PFC-amygdalar connectivity mediated the relationship between attentional vigilance and real-world distraction Price et al.  In this study, screening for ASD was developed based on characteristics of functional networks. Local and global parameters of the brain functional network were first calculated using graph theory, and network parameters of ASD subjects were statistically compared to those of healthy controls. Significantly altered parameters were used as input features of the screening system, and performance of the system was verified using multiple classification techniques. The support vector machine showed an accuracy of 92% Scheuer et al. Group difference In comparing healthy controls and adolescents with depression prior to group CBT, patients with depression showed stronger amygdala and sgACC connectivity with regions of the DMN, whereas healthy controls showed stronger seed-based connectivity with affective regions and regions processing cognition and salient stimuli. Relative to pre-CBT in depressed adolescents, post-CBT functional connectivity significantly increased between the amygdala and the left dorsolateral PFC, bilateral dorsal ACC, and the left anterior insula. Changes in connectivity correlated with significant pre-to-post CBT symptom improvement, and pre-treatment amygdala connectivity predicted treatment response in depressed adolescents Traynor et al. (2018) 62 10-21 Autism Resting state fMRI Diagnosis Group difference Compared to controls, ASD adolescents showed negative connectivity of the PCC with the angular gyrus, positive connectivity of the PCC with the superior temporal gyrus, over-connectivity of the hippocampus with the associative visual cortex, over-connectivity of the thalamus with multiple sensory processing areas of the cortex, over-connectivity of basal ganglia structures (putamen and globus pallidus) with somatosensory and motor cortices and with the fusiform gyrus, and under-connectivity of the left hippocampus with the right peri-rhinal cortex. Within the ASD group, a significant positive association was found between total RBS-R (Repetitive Behavior Score-Revised) score and connectivity between the left primary visual cortex and right inferior frontal gyrus, pars orbitalis Group difference Compared to healthy controls, depressed subjects showed significantly lower FA-weighted node strength of the right caudate. FA-weighted node strength was correlated positively with age across both groups. Network-Based Statistic analysis showed a cluster of lower FA-based connectivity in depressed subjects centered on the right caudate, including connections to the frontal gyri, insula, and anterior cingulate. Within this cluster, the strongest difference between controls and depressed subjects was the connection between the right caudate and middle frontal gyrus, which showed a significant diagnosis by stress interaction and a negative correlation with total stress in depressed subjects Velasquez et al. (2017) 108 8-19 Autism Task-based fMRI (face processing task)

Diagnosis
Group difference During face processing tasks, ASD youth with low-expressing 5-HTTLPR (serotonin transporter-linked polymorphic region) genotypes showed significantly greater amygdala-sgACC connectivity compared to healthy controls and to ASD youth with higher-expressing genotypes. ASD youth with higher-expressing genotypes also showed a negative relationship between amygdala-sgACC connectivity and social dysfunction

(Continued)
Frontiers in Human Neuroscience | www.frontiersin.org  T) during follow-up showed significant reductions in functional connectivity at baseline, primarily involving the limbic system, in comparison to healthy controls and to ARMS subjects that did not transition (ARMS-NT). ARMS-T subjects also exhibited reduced global efficiency at baseline in comparison to ARMS-NT subjects. In the salience network, the mean nodal efficiency computed over regions that were reduced in the ARMS-T group was associated with baseline positive and negative syndrome scale (PANSS) general scores. At the whole brain level, the ARMS-T group network community structure displayed a distinct pattern from that of the ARMS-NT group and of healthy controls

Diagnosis
Group difference Distant fc of regions in the left frontal lobe, right parietal lobe, and left posterior middle temporal cortex showed a group by state interaction such that in the task state (relative to the resting state), fc was reduced (became focal) in healthy controls but increased (became diffuse) in children with ASD. In ASD children, higher state-related increase in distant connectivity of the left frontal and right angular gyrus predicted worse inattention, and global efficiency and modularity were also sensitive to group by state differences, with the magnitude of state-related change predicting inattention Zhang et al. functional connectivities as well as network properties that differentiate symptomatic subjects from controls. Similar to the model proposed by Williams (2017), connectivity aberrations associated with the disorders can be broadly classified as aberrations in six large-scale brain networks: default mode network (DMN), salience network (SN), threat network, reward network, attention network, and cognitive control network (CCN). Notably, this was observed not only for the mood disorders as proposed by Williams but also for other disorders, such as schizophrenia and autism (specifically associated with aberrations in DMN and SN) (Chen et al., 2017;Joshi et al., 2017;Wang et al., 2018a).

LIMITATIONS OF THE INCLUDED STUDIES Small Sample Sizes
The evidence from the reviewed literature requires further investigation of the predictive and diagnostic potential of MRI connectivity measurements as the number of studies and sample sizes are very small. For example, seven of the reviewed studies included under 30 subjects per group. While small studies may be instrumental in finding new candidate biomarkers, replication of the results is the key. Additional well-planned imaging studies with large sample sizes are required. Yet, the field is developing rapidly. Large-scale collaborative efforts by consortia such as ABCD (Casey et al., 2018), ENIGMA (Thompson et al., 2014(Thompson et al., , 2017, and IMAGEN (O'Halloran et al., 2018) will help identify robust brain connectivity signatures of psychopathologies in adolescents. Among the included study, one study (Kaczkurkin et al., 2018) reports functional connectivity analyses using resting-state functional MRI in 833 participants who received both arterial spin labeling (ASL) and resting-state imaging as part of the Philadelphia Neurodevelopmental Cohort. The results revealed that overall psychopathology was associated with decreased connectivity between the dorsal ACC and bilateral caudate.

Group Differences vs. Single-Subject Prediction
Another limitation of the psychiatric imaging literature in general is a profusion of statistically significant, but minimally differentiating biological findings (Kapur et al., 2012). In other words, thousands of studies are published on different aspects of brain disorders to show aberrations of some features (structural or functional) in a patient group usually in comparison with a healthy cohort. While these studies are valuable in terms of finding relevant disease biomarkers, they are not sufficient for direct clinical diagnostic/predictive adoption (Arbabshirani et al., 2017). The main reason is that many of these findings are statistically significant at the group level, but the individual discrimination ability of the proposed biomarkers is not typically evaluated. Group analysis aims to estimate the probability of a certain biomarker given the group (e.g., healthy controls or a patient group): P[brain biomarker|group], and it is typically performed using general linear modeling (e.g., t-test or ANOVA). Single-subject prediction, on the other hand, predicts belonging to the group given the brain biomarker: P[group|brain biomarker], and it is typically performed using generalized linear model such as logistic regression or artificial intelligence (AI) approaches such as machine learning. As mentioned previously, only group analysis is commonly performed in neuroimaging studies. In such cases, effect size can serve as an important indicator of the individual discrimination ability of the biomarker, but it is often not reported-as it is with the majority of the studies reviewed in Table 1. For continuous data analysis, the correlation coefficient points at high or rather low discrimination ability, as in the example of the large study by Kaczkurkin and colleagues discussed above (Kaczkurkin et al., 2018), where Pearson r = −0.18 and r = −0.15. The effect size is low if the value of r varies around 0.1, medium if r varies around 0.3, and large if r varies more than 0.5 (Rosenthal and Rosnow, 1984;Cohen, 1988). Since classification provides information for each individual subject, it is considered a much harder task than reporting group differences. Nevertheless, recent extensive evidence shows the great potential of neuroimaging data for single subject prediction of various brain disorders in adults (Arbabshirani et al., 2017). Several of the studies in adolescents also performed single-subject prediction analysis. For example, functional brain connectivity developmental patterns were found to be a reliable biomarker of severe attention impairment in youth, with a peak receiver operating characteristic curve of 79.3%, measured by area under the curve (Kessler et al., 2016). In another example, machine learning was applied to structural connectivity data to predict symptom improvement in depressed adolescents in response to cognitive behavioral therapy (CBT), resulting in an accuracy of 83% (Tymofiyeva et al., 2019).

Use of the DSM Classification as Ground Truth
One fundamental challenge is that the reviewed clinical studies used the DSM-based diagnosis as the ground truth. The reliance on the categorical system of diagnosis has clear utility; however, its validity has been questioned and dimensional views of illness that incorporate continua of neurobiology and observable behavior have been proposed-i.e., NIMH RDoC (Insel et al., 2010). If one takes seriously the possibility of 1 day developing a neurobiologically-based diagnostic system that would replace the symptom-based DSM nosology, then one biases the results by using DSM diagnosis as the ground truth in research studies. Thus, the advancement of psychiatry appears to be hindered by circuitous reasoning (i.e., a "Catch-22"), and this circularity impedes the development of a clinically viable alternative system (Kapur et al., 2012). Importantly, a recent paper by Drysdale et al. reported promising results showing that adult depression can be subdivided into biological types (Drysdale et al., 2017). Specifically, functional MRI scans of more than 1,100 patients with clinical depression and healthy individuals enabled researchers to demonstrate that patients with depression can be divided into four subtypes based on distinct patterns of functional connectivity in limbic and fronto-striatal networks and different clinical symptoms (Drysdale et al., 2017). Notably, these four subtypes of depression were also associated with differences in clinical treatment outcome.

VISION: THE ROLE OF NEUROIMAGING IN PEDIATRICS
Here, we would like to present a vision of the role that neuroimaging may play in pediatrics and primary care in the future. We would like to start by presenting a case example of a clinical presentation of an adolescent patient to a primary care outpatient clinic.
A concerned parent brings to the primary care clinic a 15-yearold adolescent male with a recent history of increased mood lability, irritability and changes in his personality that has led to significant difficulties with the patient's relationship with his parents and siblings at home and the patient's teachers and classmates at school. The parent states that the patient's personal hygiene has declined at home and that he has started to withdraw from his relationships with his family and friends. The parent also reports that the patient at times loses his temper and has been oppositional with his family, teachers and friends. During the interview by the primary care provider, the adolescent patient complains of problems with feeling easily irritated by others, difficulties with sleep and concentration, and headaches. There are times that the patient also reports feeling "on top of the world" and that he can accomplish many great things. The parent informs the primary care provider that there is a history of mood disorders in the family including a parent with depression and a relative with possible bipolar disorder. The parent also states that there is a relative who has had significant problems with truancy and breaking rules whose behavioral problems started during adolescence. Finally, the parent reports that a great grandparent was hospitalized with a diagnosis of schizophrenia. The parent is concerned about the patient and has brought him to the primary care clinic for help with the son that the parent loves and cares for very deeply.
At present, psychiatric diagnoses are based on symptoms and classified as per the DSM criteria. One significant complicating factor is that most of the DSM diagnoses have been developed for adults, but these same DSM criteria are being applied to make diagnoses in children and adolescents. Based on the presenting symptoms and DSM criteria, the adolescent patient's differential diagnosis is broad. The patient's presenting symptoms of mood lability, irritability, sleep, and concentration problems, and relationship difficulties with family, friends and teachers could be due to a diagnosis of depression or bipolar disorder. The patient's loss of temper and oppositional behavior could be due to a mood disorder such as depression or oppositional defiant disorder that may be the beginning of a possible conduct disorder. The patient's report of feeling at times "on top of the world" and that he can accomplish many great things could be reflective of normative adolescent development or due to bipolar disorder. The patient's decrease in personal hygiene and withdrawal from his family and friends may be due to a mood disorder such as depression or the beginning of another possible psychiatric disorder such as schizoaffective disorder or schizophrenia. Finally, the patient's changes in personality such as his mood lability and complaints of headaches may be due to a mood disorder such as depression or a brain mass such as a tumor. Each of these possible diagnoses has a different clinical prognosis and treatment. Proper diagnosis is essential to ensure that the optimal clinical treatment is selected and to prevent the further worsening of the patient's condition due to either the incorrect diagnosis and treatment or the potential side effects of the use of the inappropriate treatment due to the incorrect diagnosis. For example, if the patient's symptoms lead the clinician to incorrectly make a diagnosis of bipolar disorder based on DSM criteria, then the patient will often be started on a mood stabilizer (e.g., lithium) or atypical antipsychotic medication (e.g., olanzapine) that have several potentially negative side effects including significant weight gain and increased risk for the development of type 2 diabetes. In addition to the potential negative consequences of an incorrect diagnosis to the patient, an incorrect diagnosis can also lead to great costs to the patient, patient's family, and society. For example, missing the diagnosis of depression can lead to significant present and future consequences since adolescent depression confers a strong risk for adult major depressive disorder, increased cardiovascular risk, medical illnesses, disability, premature death, academic and work problems, relationship problems with family and friends, substance abuse, and suicide (Rao et al., 1993(Rao et al., , 1995(Rao et al., , 1999Lewinsohn et al., 1998;Pine et al., 1998;Armstrong and Costello, 2002;Lehrer et al., 2006;Copeland et al., 2007;Crum et al., 2008;Audrain-McGovern et al., 2009;Rao and Chen, 2009; Maughan FIGURE 2 | A schematic of our vision for a routine neuropsychological and neuropsychiatric imaging (NPPI) protocol for adolescent patients. The standardized MRI protocol could consist, for example, of a 30 min brain scan at a 3T MRI scanner that would include a localizer scan, a T1-weighted sequence, a 55-direction diffusion-weighted sequence and an eyes-closed resting-state fMRI sequence. The MRI scan will be followed by a quality control and safety read of the scan by a radiologist. Next, connectivity matrices will be derived, and structural and functional brain network metrics will be calculated. Platforms like BRIDGE (https://bridge. ucsf.edu) can be used to integrate the obtained data and provide the output to the clinician (pediatrician) by means of organized dashboards. Liu et al., 2017). Depression is a highly prevalent, devastating, costly and frequently re-occurring chronic illness that the World Health Organization (WHO) ranks as the #1 leading cause of disability worldwide, affecting over 300 million people (WHO, 2017) at an estimated cost of over $210.5 billion per year in the U.S. alone (Greenberg et al., 2015). Our vision for preventing many of the drastic consequences described in the case example is the introduction of a routine neuropsychological and neuropsychiatric imaging (NPPI) protocol for adolescent patients (Figure 2). The rationale for the presented vision is rooted in the research successes described above, with the expectation that the outlined limitations can be overcome with dedicated work of researchers around the world. The standardized protocol could consist of the following: -A 30 min brain scan at a 3T MRI scanner that would include a localizer scan, a T1-weighted sequence, a 55-direction diffusion-weighted sequence and an eyes-closed resting-state fMRI sequence. -A quality control and safety read of the scan by a radiologist.
-Network construction and a computer-based calculation of the structural and functional brain network metrics (e.g., using the Brain Connectivity Toolbox brain-connectivity-toolbox.net) (Rubinov and Sporns, 2010) that can be compared to the normative data by the pediatrician (e.g., using a platform like BRIDGE https://bridge.ucsf.edu).
The chosen MRI sequences are an example of an easily standardizable set of sequences that in a short time could provide structural and functional connectivity data of sufficient quality. The choice of the 55-direction diffusion-weighted sequence is based on its suitability for high angular resolution diffusion imaging (HARDI) analyses that can help resolve crossing fibers (Tuch et al., 2002). The resting-state fMRI sequence can realistically be performed at any imaging center, and it can potentially allow for derivation of task-based information (Tavor et al., 2016). The protocol can be offered to all adolescent patients without MRI contraindications (e.g., cardiac pacemakers, braces, etc.) to meet two goals: a) To help generate the differential diagnosis for an individual patient or assess risk. b) To personalize treatment (predict treatment outcome) for an individual patient.
For improved diagnostic/predictive accuracy, the scan can be potentially conducted at two different times or even once per year during the adolescent years: 14y.o., 15y.o., 16y.o., 17y.o. (Kessler et al., 2016), especially since the age of onset of different psychiatric disorders can vary. Normative modeling can provide a way to map deviations from an expected pattern at the individual level (Kessler et al., 2016;Marquand et al., 2019). Whereas identifying MRI-based "biotypes" as an alternative to DSM categories is one approach to make diagnosis using NPPI in the future, another approach would be to measure network aberration along several continuous NIMH RDoC dimensions and creating profiles that would inform risks similar to high blood pressure in cardiology, and inform clinical care by suggesting the optimal treatment choice for an individual patient (Williams, 2017). While there are multiple steps necessary to implement this vision, the common concern that is raised when considering imaging is the cost of the MRI procedure, which we will address in the following section.

COST-BENEFIT ANALYSIS: MDD AS AN EXAMPLE
Here, we present a cost-benefit analysis of conducting routine NPPI in adolescents using only a single DSM diagnosis, MDD, as an example (Supplementary Figure 1).
Adolescent depression can be a significant economic burden to patients, families, and society. A cost-of-illness study on adolescent depression in the U.S. estimates direct costs of $1,120 and indirect costs of $310 per year, when accounting for the effects of school absences and parental lost days at work (Domino et al., 2009;Beecham, 2014). Furthermore, adolescents with depression are likely to have higher healthcare costs in other domains, due to more contacts with other healthcare providers (Lynch and Clarke, 2006). These ramifications often continue into adulthood as individuals who had depression in adolescence continue to be associated with higher healthcare utilization and increased work impairment in young adulthood (Keenan-Miller et al., 2007). Costs of adult depression in the U.S. are over $210.5 billion annually (Greenberg et al., 2015). This suggests an economic argument for timely identification and treatment of adolescent depression and for considering the costs of adult depression when weighing costs of MR imaging against benefits.
We estimate the costs of a NPPI scan at $400 per patient. This estimate is based on the University of California San Francisco (UCSF) 3T MRI external recharge rates: $350/30 min (https:// radiology.ucsf.edu/research/core-services/7T-3T-MB) + $50-75 for reading for incidental findings (estimate provided by Research Radiology, San Francisco, USA). While scheduling short 30 min time slots at an MRI scanner may currently not be possible everywhere, streamlined procedures and even cost-reducing solutions such as dedicated head-only scanners (Foo et al., 2018) can be leveraged to increase throughput in the future. In contrast to a conventional MRI, the scan will not need to be read diagnostically by a neuroradiologist; instead a neuroradiologist will need to perform image quality control and a "safety read" for incidental findings. The brain network metrics will be calculated using a computer algorithm. In case of incidental findings, such as an indication of a tumor, further examination and/or treatment may be indicated. The effectiveness of the NPPI will depend on the sensitivity and specificity of the biomarkers. An accuracy of 80% (both sensitivity and specificity) is generally considered to characterize a clinically useful biomarker (Savitz et al., 2013). The adolescent brain connectivity papers reviewed above where single-subject prediction analysis was performed displayed biomarker accuracy in this range (80-83%) (Kessler et al., 2016;Tymofiyeva et al., 2019). Importantly, several single-subject prediction studies in adults that demonstrated MRI's potential to predict treatment response also studied for comparison the predictive value of demographic and clinical variables-as these variables are more readily available and would be a cheaper solution (Månsson et al., 2015;Thompson et al., 2015;Drysdale et al., 2017). These studies found that these demographic and clinical variables failed as predictors of clinical improvement after treatment (Månsson et al., 2015;Thompson et al., 2015;Drysdale et al., 2017).
Based on U.S. Census Bureau estimates, there were 4.1 million 14-year-olds in the U.S. in 2017 (US Census Bureau, 2017). If we performed a routine NPPI scan on every 14-year-old in the U.S. annually, we estimate a cost of $1.6 billion per year, based on an estimated cost of $400 per MRI scan. The estimated cost of adult MDD in the U.S. is over $210.5 billion annually (Greenberg et al., 2015). This means that NPPI would only have to help prevent (through early detection and personalized treatment) ∼0.8% of adult MDD cases to completely cover its costs. Adolescence is the optimal time to intervene because it is when MDD often begins, and it increases the risk of adult depression by a factor of 2-to 3-fold (Pine et al., 1998).
In the calculation above we assumed that we would routinely scan every 14-year-old. This is an extreme scenario, whereas a more plausible scenario would be to scan only at-risk youth and youth presenting with clinically significant problems. This more targeted approach would reduce the estimated total cost of NPPI that we presented above. Given the significant costs of teen MDD to the patient, family and society, the cost of an MRI scan is well worth the benefit in terms of decreased prolonged emotional pain and suffering, lower risk of adult MDD and healthcare costs, and improved health and productivity of the patient. As discussed above, among many other problems, MDD elevates the risk for cardiovascular disease in both adolescents (Goldstein et al., 2015) and adults (Penninx, 2017), and MDD increases the medical costs of treating and managing primary care illnesses such as diabetes in adolescents (Stewart et al., 2005). Additionally, MRI scans are routinely done for medical conditions (e.g., lower back pain and knee injuries) that do not have the potentially devastating consequences of MDD (e.g., suicide). Another example is routine prenatal ultrasound imaging visits that are done with all, not just at-risk, pregnant women. Thus, once valid and reliable MRI biomarkers have been developed and tested for teen MDD, the cost of an MRI scan should not be a justifiable impediment since MRI scans are frequently performed for much less costly and devastating medical conditions. Medical insurance groups or Health Maintenance Organizations (HMOs) may consider this approach as a means of reducing overall healthcare costs, since early prevention could reduce the incidence of adult MDD, which is very costly to both the patient and healthcare system. We also wish to emphasize that our conservative estimate only focuses on MDD; however, many other psychiatric conditions can be potentially assessed using the same MRI scan.

THE ROADMAP
What is needed for this vision to be implemented? We briefly present below a roadmap of the steps required for this vision to be implemented. a) A large normative database. As discussed above, the field is moving toward collecting big data (that often include a replication sample). The ABCD study has already started to release MRI neuroimaging datasets for 10,000 youth who will be followed longitudinally and scanned every 2-years over a total duration of 10-years. As mentioned above, other initiatives such as ENIGMA and IMAGEN can also provide important big datasets. b) A set of NPPI-derived metrics, their sensitivity/specificity, and guidelines on how to combine them with symptombased information. c) Consensus guidelines for when NPPI is indicated (expert consensus panel). d) Required qualifications and pipeline for NPPI + data quality and safety read + metrics calculation. There is currently a high heterogeneity of analysis pipelines used by researchers to derive brain connectivity matrices, which can lead to large discrepancies between the resulting structural (Qi et al., 2015) and functional (Carp, 2012) network metrics. In addition, choice of brain parcellation and edge weights will also affect test-retest reliability of the resulting metrics (Cammoun et al., 2012;Yuan et al., 2019). Development of a standardized, objective, and publicly available pipeline is a necessary step. Once the connectivity matrices are derived, network metrics can be calculated, e.g., using the Brain Connectivity Toolbox (brainconnectivity-toolbox.net) (Rubinov and Sporns, 2010). The Brain Connectivity Toolbox is a MATLAB toolbox for complex-network analysis of structural and functional brain connectivity datasets, which is widely used by brain-imaging researchers and has been incorporated in many projects, including the Human Connectome Project. Platforms like BRIDGE (https://bridge.ucsf.edu) can be used to integrate the obtained data and provide the output to the clinician by means of organized dashboards. Required qualifications for data quality assessment and safety reads need to be specified. e) CPT code. The Current Procedural Terminology (CPT) code set is a medical code set maintained by the American Medical Association through the CPT Editorial Panel. This step is required to enable insurance companies to pay for adolescent NPPI (e.g., similar to ultrasound for neonates).

DISCUSSION
At present, utilization of neuroimaging biomarkers for clinical practice is restricted to neurological conditions such as presurgical evaluation of epilepsy, differential diagnosis of coma, and brain-computer interfaces for locked-in patients (Arslan, 2018). For psychiatric conditions, this is yet to be established for routine clinical applications. Because adolescence is an especially vulnerable time for the development of many important psychiatric disorders, this period presents an especially important opportunity to clinically intervene. In this perspective article, we assess the MRI-based brain connectivity literature over the last 5-years that provides insights into development of psychiatric disorders in adolescents. While the subject numbers are still small and the focus on group differences as opposed to individual subject-based predictions prevails, the reviewed literature demonstrates the potential of MRI to diagnose psychiatric disorders, predict their development, and predict response to clinical treatment. We believe that the continuous progress in neuroimaging techniques together with the ongoing well-coordinated large-scale neuroimaging studies of adolescent brain development will help identify robust biomarkers for personalized/stratified medicine (Kapur et al., 2012) with a key focus on prevention. Inspired by this significant potential, we offer a vision for the role that neuroimaging may play in pediatrics and primary care in the future: a routine NPPI protocol for adolescent patients that could save significant costs to the patients, their families, and society, and significantly reduce suffering that often results from undiagnosed and misdiagnosed adolescent psychiatric disorders. The proposed vision can also help offer preventative measures to at-risk youth in a targeted manner-e.g., by recommending the Training for Awareness, Resilience and Action (TARA) (Henje Blom et al., 2014 to youth at risk for developing depression. Insurance companies are more likely to reimburse such preventative measures when numerical cut-offs are used, as will be provided by our proposed NPPI. The roadmap we provide in paper can help accomplish this endeavor.

DATA AVAILABILITY STATEMENT
The original contributions presented in the study are included in the article/Supplementary Material, further inquiries can be directed to the corresponding author(s).

AUTHOR CONTRIBUTIONS
OT, DX, CH, and TY conceptualized the article. VZ performed literature search. OT, C-ML, and TY performed cost-benefit analysis. OT created figures. OT, VZ, C-ML, DX, CH, and TY wrote the manuscript. All authors contributed to the article and approved the submitted version.