CSD-GCN: Cross-Modal Self-Distillation Graph Convolutional Network for Alzheimer’s Disease Diagnosis

Abstract

Alzheimer’s disease (AD) is a prevalent neurodegenerative disorder that demands more comprehensive and precise modeling to unravel its complex pathological mechanisms and multidimensional brain network abnormalities. However, current brain network analysis approaches often suffer from limitations such as single-modal data processing and shallow fusion strategies, which hinder their ability to capture the multi-scale connectivity patterns associated with AD effectively. To alleviate these limitations, a novel multi-modal fusion framework is proposed in this paper, namely a cross-modal self-distillation graph convolutional network (CSD-GCN). The framework can deeply integrate fMRI and DTI data to facilitate the systematic extraction of hierarchical features from multimodal brain networks. The architecture of CSD-GCN comprises three progressive layers: edge-to-edge (E2E), edge-to-node (E2N), and node-to-graph (N2G). These layers are designed to perform multi-scale feature extraction, capturing neurodegenerative patterns ranging from local connection strengths to global topological structures. Moreover, the model incorporates a cross-modal self-knowledge distillation strategy and a cross-attention mechanism to enhance inter-modal feature alignment and fusion. These components collectively improve the generalization ability and discriminative capability of the learned representations. Experimental evaluations show that CSD-GCN outperforms existing methods in binary and ternary classification tasks for AD, with ablation studies confirming the individual effect of each component within the proposed framework.