Multimodal Neuroimaging Fusion using 3D Residual Networks and Vision Transformers for Parkinson’s Disease Detection

Abstract: The problem of initial and reliable analysis of Parkinson disease (PD) is caused by the heterogeneous nature of its clinical manifestation and indistinct neurobiological alterations of prodromal and early stages of the disease. The study hypothesizes an explainable Deep Learning (DL) framework of automated analysis and detection of PD through multimodal neuroimaging. The clinical data from the Parkinson’s Progression Markers Initiative (PPMI), Structural Magnetic Resonance Imaging (MRI) (T1-weighted), resting-state functional Magnetic Resonance Imaging (fMRI), and Diffusion Tensor Imaging (DTI) were jointly leveraged to capture complementary morphological, functional, and microstructural biomarkers. Structural MRI was Z-score normalized, fMRI underwent wavelet denoising, and DTI was corrected for eddy currents to reduce noise, standardize intensities, and enhance reliable multimodal feature extraction. 3D Sobel filtering was applied to MRI, fMRI, and DTI to enhance structural edges and anatomical boundaries. A multimodal fusion architecture integrating 3D Residual Networks with vision transformers (3D RN-ViT) was developed to diagnose PD. To ensure clinical transparency, Explainable Artificial Intelligence (XAI) techniques were employed to localize disease-relevant neuroanatomical regions and quantify modality-wise contributions. A Multi-Layer Perceptron (MLP) head was employed for final classification on the extracted multimodal features. Satin Bowerbird Optimization (SBO) was employed for optimal feature selection to classify the best discriminative multimodal structures while reducing redundancy and dimensionality. Experiments were carried out in the Python framework using PyTorch. The proposed 3D RN-ViT attained 99.2% accuracy, 99.2% precision, and a 99.34% sensitivity. Overall, the proposed approach provides a transparent and scalable pathway for the analysis of PD, supporting biomarker discovery and serving as a promising proof-of-concept for AI-assisted neuroimaging systems. Future external cross-cohort validation will be required prior to clinical deployment.