DECODE-PD: deep explainable composite imaging fusion of MRI, PET, and SPECT for early detection of Parkinson's disease

Parkinson's disease (PD) remains challenging to diagnose at early stages due to subtle neurodegenerative changes and symptom overlap with other neurological disorders i.e., Alzheimer's and corticobasal degeneration. Current imaging-based AI classifiers often rely on single- or dual-modalities, limiting sensitivity, and most lack clinically validated interpretability. This study introduces DECODE-PD, an end-to-end, explainable AI framework for early-stage PD classification using fused MRI, PET, and SPECT imaging data, the first to combine these modalities for PD detection. We propose a robust multimodal medical image fusion (MMIF) algorithm that aligns structural, metabolic, and functional scans through precise co-registration and atlas-based normalization, followed by modality-specific deep feature extraction and an adaptive attention-based fusion layer. The classification backbone is a novel ViNeXt ensemble, integrating Vision Transformer and ConvNeXt architectures with transfer learning to capture both global context and fine-grained spatial features. An explainable prototype-based layer (XProtoPNet) provides intrinsic interpretability by associating decisions with learned PD-specific prototypes, enabling clinically relevant explanations. Tested on the PPMI dataset, DECODE-PD achieved 96.1% accuracy (AUC=0.985), outperforming state-of-the-art multimodal classifiers and demonstrating a 94.8% accuracy for early-stage PD detection, a >10% improvement over MRI-only approaches. Clinical review confirmed the alignment of prototype activations with known PD pathology in 92% of cases. This work establishes a new benchmark for performance and trust in AI-assisted PD diagnosis, with potential for integration into clinical decision support systems to accelerate diagnosis, enable earlier intervention, and improve patient outcomes.