Hierarchical Pooling Graph Convolutional Neural Network for Alzheimer’s Disease Diagnosis

Deep learning techniques have found extensive applications in utilizing magnetic resonance imaging (MRI) to support the diagnosis of Alzheimer’s disease (AD). However, existing research primarily focuses on the pathological changes in brain regions affected by the disease, while overlooking the intrinsic correlations among these regions. This disregard may lead to inaccurate disease predictions. Graph data, represented in the form of nodes and their connecting edges, effectively describe the relationships between nodes. However, constructing representative brain connectivity graphs remains a notable task. To solve the above problems, we come up a layered pooling graph convolutional classification network based on MRI to learn differential features between samples in the MRI data for AD diagnosis. Our innovation lies in utilizing two layers of classification networks, processing brain connectivity graphs with self-attention convolutions to obtain brain feature structure graphs at different granularities. By integrating data at different scales, we can comprehensively and accurately capture feature information in brain magnetic resonance images. Our method further utilizes global pooling to aggregate learned brain structural features and generate gradually evolving topic-level representations for AD diagnosis. Experimental validation on the openly accessible ADNI dataset demonstrates the competitive performance of our method in multiple AD-related classification tasks. Compared to existing methods, our method better captures brain structural features and exhibits stronger generalization capability.