Automated classification of AD, nonetheless, in the evolving state, has shown promise with AI algorithms to accurately differentiate between individuals who have been clinically diagnosed with AD and MCI from those of healthy controls. Currently, more extensive work for approaching the classification tasks was done using DNN that can automatically learn discriminatory multilevel and multimodal features with high accuracy (15, 16). The advantage of DNN algorithms is that these models do not require segmentation and are robust to scale and rotation variations (15). However, the “black-box” nature of DNNs makes them less trustworthy to physicians, thus hindering their expansion into real clinical settings. On the other side, the attractiveness of radiomics-based algorithms is that they can serve as a potential bridge between brain imaging and personalized medicine, as these algorithms are able to incorporate patient-related information and disease-related biomarkers.

The mainstream pipelines using radiomics, first, segment the PET brain image into various regions of interest (ROIs), extract and select the discriminate features for AD-target ROI regions, and finally, input to state-of-the-art machine learning models trained to support different tasks for clinical decision-making (17). A few studies using ROI-based radiomics and support vector machine (SVM) classifiers with selected radiomic features extracted from [18F]FDG PET images reported stability achieving good classification accuracies for classifying AD vs. HC (91.5 to 92.9%), MCI vs. HC (83.1 to 83.7%), AD vs. MCI (85.9 to 87.9%), and MCI_converter vs. MCI_{non−converter} (88%) (18, 19). Similarly, Ciarmiello and co-authors demonstrated that ROI-based radiomics using gray-level run-length matrix for texture analysis and feed-forward multilayer neural network on Aβ-PET data outperform SUVr-based performance (AUC 0.9 vs. 0.71) in differentiation amnestic MCI subjects from healthy controls (20).

Besides conventional ROI-based radiomics, Ding et al. applied voxel-wise radiomics on standard uptake value ratio (SUVr) intensity data using Aβ-PET (9) as a biomarker to classify healthy controls and AD with a nonlinear SVM model and cross-validation techniques. The classification performance was better than that with only the Aβ scores based on regions (ACC = 0.75, SPE = 0.80, SEN = 0.70, AUC = 0.81) or based on voxels (ACC = 0.84, SPE = 0.86, SEN = 0.81, AUC = 0.90).

It is a significant challenge to evaluate and predict the progression of MCI for early treatment of AD (9, 19). A classic approach for predicting MCI conversion is to train a classification model using the data of MCI converters and non-converters and then use the trained classifier to classify new MCI patients (Liu et al. 2018). In addition, the proportional hazards (Cox) model is commonly constructed as the prediction model for investigating the effect of several variables (predictive features) upon the time (conversion time) (21).

Jiang and co-authors designed different Cox models (radiomics, clinical, SUVr FDG PET) to compare the predictive performance of MCI conversion (21). The experimental results showed that the predictive performance of the radiomics-based Cox model was better than that of other Cox models. In the validation dataset, Harrell's consistency coefficient of the radiomics-based Cox model was 0.703 ± 0.002, while those of the clinical and SUVR models were 0.632 ± 0.006 and 0.683 ± 0.009, respectively.

Alongi et al. propose a radiomics approach on [18F]FDG PET/CT brain images based on Statistical Parametric Mapping (SPM) and Pyradiomics to predict PET-Amyloid positivity (22). The authors found six radiomics features that predict cortical Aβ deposition with high sensitivity, specificity, precision, and accuracy of 84.92%, 75.13%, 73.75%, and 79.56%, respectively. These preliminary results obtained from [18F]FDG PET/CT brain radiomics for predicting the presence of Aβ deposition may serve as a new approach for selecting those patients who can benefit from the diagnostic use of Aβ PET.

The workflow for radiomics-based predictive analysis using Aβ PET brain images was also reported (9). The Aβ PET radiomic features with widely used machine learning and cross-validation techniques are able to predict with good performance (AUC 0.83) the progression of high-risk MCI patients' conversion to AD (9). Interestingly, in the MCI and AD groups, the classification outputs are significantly associated with clinical measures, such as apolipoprotein E genotype, polygenic risk scores, polygenic hazard scores, cerebrospinal fluid Aβ, and Tau, cognitive ability score, the conversion time for progressive MCI subjects and cognitive changes, thus, highlighting the solid biological/clinical basis underlying the progression of AD (9).

A few studies attempt to address the added value of dual-modality imaging for the prediction of MCI conversion to AD by using radiomic analysis (14, 23). Yang and co-authors compared the predictive accuracy of single-modality MRI and [18F]FDG PET and dual-modality [18F]FDG PET/MRI to predict the MCI conversion to AD, demonstrating a large overlap between [18F]FDG PET and MRI radiomics model performance, while the dual-modality model resulted in only a modest improvement over the single-modality models (Harrell's C-index of 0.798 for dual-modality vs. 0.760 for MRI, and 0.734 for PET; both P < 0.001) (14). Even though the incremental benefit of combining [18F]FDG PET and MRI for predicting MCI conversion seems limited according to Yang et. al., the integration of cross-modality approaches using radiomics should be prioritized for their potential additive value allowing the combination of the extracted imaging information for each modality. Furthermore, as radiomics is naturally designed to develop decision-support tools for precision medicine, it should involve the combination of radiomic data with other patient characteristics, as available, to increase the power of the decision-support models (14). Zhou and co-authors showed improved performance constructed similar prognostic Cox models, yet, using the clinical data, MRI images, PET images, fused MRI/PET images, and clinical variables and fused MRI/PET images in combination (Harrell's C-index 0.69, 0.73, 0.73 and 0.75, and 0.78, respectively) (23). Significant enhancement in the prediction of conversion of the imaging models (MRI/PET/fused) compared to clinical models was observed, and the combination of fused-modality imaging and clinical variables resulted in the greatest accuracy of prediction. According to the authors, the combination of radiomic and Cox model analyses could be used successfully in survival analysis and may be a powerful tool for personalized precision medicine patients with the potential to undergo conversion from MCI to AD (23).

As the most urgent in clinical practice, the first application of radiomics in nuclear medicine imaging of PD regarded the possibility of effectively correlating imaging and clinical data after diagnosis and initial clinical assessment.

Rahmim et al. analyzed cross-sectional data from Parkinson's Progressive Marker Initiative (PPMI) and found that striatal dopamine transporter (DAT) single-photon emission computed tomography (SPECT) textural features significantly correlated with the cognitive and motor status of 85 PD patients (24). Cognitive and motor performance were measured through UPRDS (part III – motor) and Montreal Cognitive Assessment (MoCA) scores, respectively. In this work, the most effective textural metrics for clinical correlation were extracted from the caudate by the most affected side. The authors discussed their possible use for the early diagnosis of PD in subjects at increased risk (e.g., mutation carriers or RBD patients).

Salmanpour et al. used cross-sectional and timeless data to identify 3 distinct PD subtypes, defined as mild, intermediate, and severe. In this research, non-imaging, conventional imaging data and radiomics features were combined. Importantly, disease clustering was robust to variations in features and sample size (25). The inclusion of radiomics features (especially structural features derived from MRI) was essential for cluster robustness.

Prompted by the previously cited paper from the same group, a longitudinal investigation by Rahmim et al. found that radiomics features from year-0 and year-1 DAT-SPECT images, together with clinical data and conventional imaging features, contributed to predicting the motor outcome of PD patients (26). In this model, clinical variables (year-0 and year-1 UPDRS-III, the unified Parkinson's disease rating scale) were the most predictive, but the addition of conventional imaging and radiomics features allowed to reduce of the average absolute prediction error of the year-4 UPDRS-III from 8 to 3 points.

The same attempt was made two years later, by using only year-0 clinical and imaging data but using an artificial neural network (ANN) to build a predictive model based on the radiomics features extracted from basal ganglia DAT SPECT (27). Importantly, this study was performed on data extracted from the same database (PPMI) as Rahmim et al. (26). The ANN reached an accuracy of 75% in predicting year-4 UPDRS-III class (<30 vs. ≥30), with a clear benefit from the combination of imaging and non-imaging (year-0 UPDRS-III) features to the model.

In recent work, Salmanpour et al. applied their clustering approach to the outcome prediction problem, and divided PD patients from PPMI database into 3 distinct progression trajectories over a 4-years follow-up, based on the motor, and non-motor clinical features, and radiomics features (28). The authors also used clinical and imaging features to build a model for the prediction of the disease trajectory, reaching an accuracy of 79%. In trajectory I (35% of patients), the disease was stable for the whole follow-up duration; in trajectory II (27%), patients underwent a clinical-imaging improvement for the first 2 years and then worsened to severe disease at 4 years; in trajectory 3 (38%), patients showed a monophasic worsening throughout the follow-up. Importantly, the consistency of the three cross-sectional disease subtypes described in previous work was observed at each timepoint in the follow-up (25). While these clusters remained distinct, longitudinal trajectories allowed patients to move across PD subtypes through time. Considering this finding, timeless disease subtyping can be seen as a multidimensional phenotypic description, which is not necessarily correlated to longitudinal progression.

[18F]FDG-PET radiomics was first tested for the diagnosis of PD vs. HC by Wu et al. (29). They found that, in multiple ROIs, features like LGZE (low gray-level zone emphasis), skewness (a measure of asymmetry) and LRHGE (long-run high gray-level emphasis) were predictive of PD diagnosis with an overall accuracy of 88%. Notably, some of these features were also significantly correlated with UPDRS (part III) and uptake values in the ROIs. Recently, Comte et al. used textural features from FDOPA PET to predict dopaminergic denervation (30). Conventional imaging features performed poorly, but the textural features allowed the model to reach an accuracy of 95.83%. Notably, no model was built here for subtyping the dopaminergic denervation into PD/ atypical-parkinsonisms. Shiiba et al. compared DAT-SPECT conventional features and a radiomics signature for the classification of 413 PPMI patients into PD or HC (31). Radiomics signature included features mostly from putamen and pallidum and was up to 96.8% accurate in the discrimination of PD patients. Nevertheless, it failed to outperform the conventional striatal uptake ratio (SUR).

Although automated image-based classification methods have been investigated for more than a decade in the differential diagnosis of PD and atypical parkinsonism, the first radiomics-based approach to this problem was by Hu et al. (32). Here, a combined clinical and [18F]FDG PET/MRI model with imaging features from PET, T1- and susceptibility-weighted imaging reached an accuracy of 96.3% in the differential diagnosis between PD and multiple system atrophy (MSA).

In a study with DAT PET, Zhao et al., found differences in basal ganglia relative binding ratios among patients with different parkinsonisms but failed to demonstrate differences between PD and MSA (33). Despite the authors extracting conventional radiomics features which were significantly different among atypical parkinsonisms no predictive model was built for the differential diagnosis, because the study was centered on deep learning-based applications.