Explainable Graph Neural Networks for Alzheimer's Disease Diagnosis Using Multi Modal MRI Data

Diagnosis of AD from neuroimaging still remains a challenging task due to complex changes in brain structure and connectivity and limited interpretability of deep learning models. This work presents an explainable multi modal Graph Neural Network framework for the diagnosis of AD by fusing structural MRI, functional MRI, and diffusion MRI. Subject specific multi layer brain graphs are constructed to model the anatomical and connectivity information in a complementary manner. Modality specific GNN branches were fused by an intermediate fusion strategy. Classification is done along with intrinsic interpretability using attention based GNN and further enhanced using GNNExplainer to identify disease relevant subgraphs of the brain. Experimental results on benchmark neuroimaging data demonstrate that the proposed model achieves superior performances, reaching <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\mathbf{9 5. 2} \boldsymbol{\%}$</tex> accuracy and 0.98 AUC for AD versus NC classification, and 91.1% accuracy for AD versus MCI. Generated explanations show clinically relevant regions such as the hippocampus, posterior cingulate cortex, and default mode network.