Dynamic Adjacency Matrix Learning and Multi-Order Random Walk Aggregation for Alzheimer's Disease Diagnosis From Resting-State fMRI

Alzheimer's disease (AD) is a progressive, irreversible neurodegenerative disorder, and early diagnosis is critical for timely intervention and slowing down the course of the disease. Functional magnetic resonance imaging (fMRI), a non-invasive neuroimaging technique, can be used to detect functional brain activity. However, most of the existing graph neural network methods based on static adjacency matrices are difficult to capture individual differences and remote brain region interactions, thus limiting their classification performance and interpretability. In this paper, we propose a Dynamic Multi-Scale Brain Network (DMSBN) model based on resting-state fMRI, which improves the discriminative ability and biological interpretability of the model through dynamic adjacency matrix learning and multistage stochastic wandering aggregation mechanism. The experimental results show that the classification accuracy of DMSBN on the Alzheimer's Disease Neuroimaging Initiative (ADNI) database reaches 96.57%, which is significantly better than the existing baseline methods. In addition, the brain region analysis module identified high contributing brain regions associated with Alzheimer's disease pathology, providing new insights for early intervention. This study validates the effectiveness of dynamic neighborhood learning and multi-scale feature fusion in brain network analysis, providing a powerful aid for early diagnosis of Alzheimer's disease.