Attention-Guided Deep Graph Neural Network for Longitudinal Alzheimer’s Disease Analysis

Abstract

Alzheimer’s disease (AD) is the main reason for dementia among aged people. Since AD is less likely reversible and has no cure yet, monitoring its progress is essential for adjusting the therapy plan of the patients to delay its deterioration. The computer-aided longitudinal AD data analysis is helpful to this kind of task, which can be used to evaluate the disease status, identify discriminative brain regions, and reveal the progression of the disease. However, most of the existing methods exist two main issues: i) the graph features are extracted globally from the entire graph, which is very sensitive to the noises; ii) they have difficulties in processing dynamic graphs, whereas the brain networks are highly variable, as they vary from individuals or changes along time or by disease. To address these issues, a novel Attention-Guided Deep Graph Neural (AGDGN) network is proposed in this paper, which utilizes an Attention-Guided Random Walk (AGRW) module to extract the structural graph features from the brain network. Since AGRW only needs the local information around the neighborhood nodes at each step of random walk, it is robust to the graph noise and flexible in dealing with dynamic graphs. Moreover, the global attention mechanism is integrated into the sequence processing module. The two attention mechanisms are jointly trained to reveal the most informative brain regions from both structural and temporal domain for AD analysis. Experimental results and analysis on the Alzheimer’s Disease Neuroimaging Initiative (ADNI) dataset demonstrate the effectiveness and efficiency of the proposed method.