Differentiating early Parkinson’s disease from multiple system atrophy on multimodal PET/MRI using atlas-guided segmentation and vision transformer classification

Introduction Early differentiation between Parkinson’s disease (PD) and multiple system atrophy-parkinsonian type (MSA-P) remains clinically challenging because of overlapping motor symptoms and subtle imaging abnormalities. Methods We developed a fully automated multimodal PET/MRI framework for early PD–MSA-P differentiation that integrates atlas-guided subject-specific label generation, nnU-Net segmentation of the caudate nuclei and putamina, multimodal image fusion, and Vision Transformer classification with multi-slice voting (ViT+MSV). A total of 155 early-stage patients, including 121 with PD and 34 with MSA-P, were evaluated using patient-level stratified 4-fold cross-validation. Results The segmentation model achieved an overall Dice score of 0.853 (95% CI, 0.830–0.877) and an IoU of 0.747 (95% CI, 0.725–0.770). ViT+MSV achieved a cross-validated AUC of 0.947 (95% CI, 0.905–0.981), an accuracy of 0.929 (95% CI, 0.890–0.968), a sensitivity of 0.942 (95% CI, 0.900–0.982), and a specificity of 0.882 (95% CI, 0.765–0.974). The combination of [ 18 F]fluorodeoxyglucose (FDG) PET, [ 11 C]CFT dopamine-transporter PET, and T2-weighted imaging (T2WI) showed the best overall classification performance. Grad-CAM visualizations showed that class-discriminative gradient-based saliency was mainly concentrated on the bilateral putamen. Discussion These findings suggest that anatomically guided multimodal PET/MRI analysis with transformer-based classification may support early PD–MSA-P differentiation while providing spatially interpretable evidence for model decisions.