Assessing AI-augmented training for multiple sclerosis classification in a basal ganglia radiomics model

Multiple sclerosis (MS) radiomics is hindered by multicenter variability and limited sample sizes. We evaluated whether GAN-based augmentation (GBA) improves MS classification versus traditional data augmentation (TDA) under center-wise external testing. A conditional GAN generated T1-weighted brain MRIs conditioned on class labels. Ten subcortical regions (including thalamus, putamen, caudate) were segmented with a 3D U-Net; radiomic features (shape, first-order, and texture families) were extracted and selected with LASSO. We used a leave-one-center-out (LOCO) design. All model development, segmentation, cGAN training, feature engineering, and tuning, were performed within the training centers only using inner 5-fold (subject-level 80/20) splits; the entire held-out center was reserved for a single external test. Across centers, GBA yielded small but consistent gains over TDA and real-only training, most evident for a tabular ResNet (average F1 up to 0.957), while confidence intervals overlapped for some metrics. SHAP analyses preserved the salience of basal-ganglia features, supporting biological plausibility. Limitations include a single-country cohort and no public external validation, which constrains generalizability. AI-augmented training provides incremental improvements for MS radiomics under site-held-out testing and motivates broader, international validation and clinically oriented utility analyses.