Abstract / Summary
Differentiating Parkinson’s disease (PD) from multiple system atrophy (MSA) represents a major clinical challenge, especially in the early stages of the disease, due to the overlap of motor symptoms. Computational tools based on routinely acquired imaging data may represent a promising approach to support clinicians in these challenging neurological conditions by providing interpretable imaging biomarkers. This study investigated whether radiomics combined with explainable machine learning could provide an interpretable framework to distinguish PD from MSA patients, and the latter from healthy control (HC) subjects, using only conventional T1-weighted MRI images. A total of 150 subjects (55 HC, 50 patients with PD and 45 patients with MSA) were enrolled. T1-weighted MRIs were pre-processed (bias field correction, skull stripping, intensity normalization, spatial resampling) and first- and second-order radiomic features were extracted from several brain regions, known to be relevant for the disease course. Six ML classifiers were evaluated using cross-validation and SHapley Additive exPlanations (SHAP) methodology was applied to assess the most influential radiomics-derived imaging biomarkers contributing to classification performance. The best classification performance was obtained in differentiating between PD and MSA and HC and MSA (up to 90% accuracy) across different brain regions. The most discriminating regions were cerebellar vermal lobules, cerebellum (proper), brain stem, and pons. SHAP analyses consistently highlighted radiomic features from these regions as key contributors to model predictions. These findings suggest that radiomics-based approaches applied to conventional MRI may represent a promising direction for developing future clinical decision-support systems.