An explainable attention-based multimodal MRI framework for Parkinson’s disease classification: a pilot feasibility and internal validation study

Background Early diagnosis of Parkinson’s disease (PD) is challenging due to clinical heterogeneity and subtle MRI changes. This study explored the feasibility of an attention-enhanced hybrid deep learning (DL) framework integrating structural MRI, neuromelanin-sensitive MRI (NM-MRI), and arterial spin labeling (ASL) perfusion to diagnose PD. Methods In this retrospective study, 54 participants (29 PD patients, 25 healthy controls) underwent multimodal neuroimaging (T1-weighted, NM-MRI, and ASL). A hybrid framework utilizing a 2.5D ResNet18 backbone augmented with Convolutional Block Attention Modules (CBAM) and a Multi-head Attention tabular branch was developed. Post-hoc explainable AI (XAI) was validated using fidelity and reliability metrics. Results The multimodal model achieved an AUC of 0.91 (95% CI, 0.87–0.96), outperforming unimodal models (T1-only: 0.81, NM-MRI-only: 0.77, ASL-only: 0.79). Feature analysis ranked putaminal and caudate cerebral blood flow and nigral neuromelanin as dominant features. Saliency maps showed high anatomical coverage and reliability in the caudate, pallidum, and thalamus for PD patients. Conclusions This pilot study suggests that an attention-based multimodal DL framework can extract PD-related signals more effectively than unimodal approaches, even in small cohorts. The model provides anatomically consistent explanations, though larger multi-center studies are needed to confirm generalizability.