A hybrid CNN-GCN framework for interpretable Alzheimer's disease diagnosis from MRI scans

Medical image analysis for Alzheimer's Disease (AD) diagnosis faces two key challenges: capturing spatial dependencies between anatomically connected brain regions and providing clinically interpretable explanations. While Convolutional Neural Networks (CNNs) excel at local feature extraction and Vision Transformers handle long-range dependencies, neither explicitly models the relational structure between brain regions-critical for understanding disease progression. We propose a hybrid CNN-GCN framework that: (1) extracts patch-level features instead of using the full image, (2) constructs graphs encoding both spatial adjacency and semantic similarity between brain regions, and (3) provides interpretability through GNNExplainer-based heatmaps highlighting diagnostically relevant areas such as the hippocampus. Our key contribution is the integration of complexity-aware node features that explicitly encode regional diagnostic importance within the graph structure. Evaluated on the ADNI dataset with 6,000 MRI slices across AD, Mild Cognitive Impairment (MCI), and Normal Control (NC) classes. The interpretability module aligns predictions with established neuroanatomical patterns, identifying regions like the hippocampus that influence diagnostic decisions. By combining localized feature analysis with graph-based relational reasoning, our approach provides both accurate classification and transparent explanations rooted in clinical knowledge, addressing key obstacles to real-world implementation of AI-assisted neurodegenerative disease assessment.