Multimodal radiomic and foundation-model MRI features converge on basal ganglia patterns in Parkinson’s motor subtypes

Abstract

Parkinson’s disease (PD) comprises heterogeneous motor phenotypes, most prominently tremor-dominant (TD) and postural instability/gait difficulty (PIGD), which differ in progression, disability burden, and therapeutic response. Despite standardized Movement Disorder Society–Unified Parkinson’s Disease Rating Scale (MDS-UPDRS) criteria, subtype classification remains partly subjective and lacks clear biological anchoring. In this pilot study, we investigated whether multimodal MRI-derived representations can discriminate TD and PIGD phenotypes by directly comparing conventional whole-brain radiomic features with latent embeddings extracted from BrainIAC, a transformer-based MRI foundation model. Sixty-two age- and sex-matched PD patients (31 TD, 31 PIGD) underwent 3T MRI including T1-weighted, FLAIR, SWI, and diffusion imaging (FA and MD). For each modality, 1046 radiomic features and 2048-dimensional embeddings were extracted. Within a nested stratified five-fold cross-validation framework, a group of machine learning models were trained using imaging (radiomics/embeddings) features alone, clinical scores alone, and their combination, with fold-wise standardization and principal component analysis (80% variance retained) for a TD vs. PIGD classification task. Radiomic features consistently outperformed foundation-model embeddings (mean AUC: 0.664 vs. 0.592), with the best performance achieved by combining FLAIR and SWI radiomics with clinical scores (AUC = 0.79 ± 0.10). Interpretability analyses indicated that discriminative components were driven mainly by SWI-derived intensity and texture features, with saliency maps highlighting striatal and pallidal regions within basal ganglia circuitry.