A Multigraph Combination Screening Strategy Enabled Graph Convolutional Network for Alzheimer’s Disease Diagnosis

Abstract

Alzheimer’s disease (AD) is a degenerative disorder that encompasses multiple stages during its onset. There are certain shared characteristics among patients at various stages of AD, which results in the presence of incorrect edges in the graph structure constructed using graph convolutional network (GCN) for AD diagnosis. Due to the presence of incorrect edges, a singular graph structure faces challenges in accurately capturing the relationships between nodes. To tackle such a problem, this article proposes a screening strategy that constructs a large number of graphs, and selects an optimal graph combination. For each graph, the model adaptively aggregates lesion area features of similar nodes. Such a graph-selecting strategy alleviates the impact of incorrect edges and yields better performance. First, a multiscale composition module is designed to find the potential relationship between nodes, and the graph structure at different scales is constructed by extracting the significant pathogenic features from the node features. Second, a multihop node aggregation (MHNA) algorithm is proposed to find the correlation between multihop nodes in the same category, and highly correlated multihop nodes are found by traversing the features of different hop nodes. Third, an optimal multigraph combination screening strategy is proposed to select the optimal multihop graph combinations under the optimal multiscale combinations, and further adaptive fusion by using the multigraph attention mechanism. This enables the whole model to capture the distinctive features of AD while enhancing aggregation among similar nodes. The proposed model achieves an average accuracy of 90.21% and 94.10% on the NACC and Tadpole datasets, respectively, surpassing state-of-the-art results.