A Brainstem Radiomics Framework to Distinguish Progressive Supranuclear Palsy from Parkinson's Disease

Background: Differentiating progressive supranuclear palsy (PSP) from Parkinson's disease (PD) can be clinically challenging. In the neuroimaging field, radiomics has emerged as a promising approach to capture subtle microstructural and textural image alterations, improving differential diagnoses. Objective: To assess the diagnostic value of brainstem radiomic features from T1-weighted magnetic resonance imaging (MRI) in distinguishing PSP from PD patients. Methods: This study included 433 participants from two independent cohorts: an Italian training cohort (84 PSP and 177 PD) and an international validation cohort (68 PSP and 104 PD). Radiomic features including first-order, shape, and texture descriptors were extracted with PyRadiomics from brainstem segmentations generated by the automated deep-learning-based AssemblyNet pipeline. Classification models (Decision Tree, Support Vector Machine, Random Forest, and XGBoost) were trained using nested cross-validation and tested on the independent cohort. Model interpretability was examined with SHapley Additive exPlanations. Results: Radiomics-based models yielded high and consistent performance in distinguishing PSP from PD, higher than brainstem volume. In the validation cohort, Random Forest and XGBoost achieved the best performance (area under the curve [AUC]: 0.93 and 0.94, respectively). Texture- and intensity-based radiomic features emerged as the most informative predictors, while shape descriptors showed lower relevance in discrimination between PSP and PD. Conclusions: Brainstem radiomics extracted from routine T1-weighted MRI demonstrated excellent classification performance in distinguishing PSP from PD patients and generalized robustly across independent datasets. Texture-based features captured microstructural disorganization not reflected by automated volumetry, underscoring the added value of radiomics for differential diagnosis in atypical parkinsonism and for integration in future multimodal biomarker frameworks. © 2026 The Author(s). Movement Disorders published by Wiley Periodicals LLC on behalf of International Parkinson and Movement Disorder Society.