Multi-pattern learning and multi-mapping graph convolutional networks for Alzheimer's disease diagnosis

Current evidence indicates brain atrophy and white matter fiber abnormalities in Alzheimer's disease (AD). However, their classification performance is limited by confounding connection features and diverse node-edge information. To address the above limitations, we proposed a Multi-Pattern Learning and Multi-Mapping Graph Convolutional Networks (MPL-MMGCN). First, morphological features and anatomical features are extracted from sMRI and DTI as node attributes, and the structural network is derived from DTI as graph representation to jointly construct the brain attribute network. Then, the constructed brain attribute networks were categorized into whole-brain pattern, intra-hemispheric pattern, and inter-hemispheric pattern based on the strength of structural connections. Next, node attributes and connection features are spatially mapped, and both are further fused with network connections based on GCN to obtain node representations for different brain patterns. Finally, the disease diagnosis results are derived by ensemble learning voting. The experimental results on the ADNI dataset show that our method improves the classification performance of brain diseases, compared with state-of-the-art methods. The study delves deeper into the role of various patterns, highlights the significance of node attributes and connection features in diverse tasks, and validates the effectiveness of edge mapping. This study helps to better understand the pathology of brain diseases and provides new insights into the diagnosis and treatment of brain diseases.