Coupled Multi-view Graph Convolutional Networks Based on Multi-scale Attention and Knowledge Embedding for Parkinson's Disease Diagnosis

Parkinson's disease (PD) is a neurodegenerative disease that frequently occurs in middle-aged and elderly people. Early diagnosis is significant in slowing its progression, and attention mechanisms and graph convolutional networks (GCNs) in the deep learning method can extract effective features for early diagnosis. However, existing attention mechanisms with fixed receptive fields or scales fail to focus on local details and crucial global patterns in parallel, and a conventional single-view GCN is insufficient for revealing graph features in both brain regions and sample relationships. Based on the above, we propose coupled multi-view GCNs based on multi-scale attention and knowledge embedding. First, the brain connectivity networks generated from diffusion tensor imaging (DTI) are input into the proposed multi-scale hybrid attention (MSHA) module to emphasize distinct channel and spatial information at multiple scales. Subsequently, we devise multi-view brain region GCNs (MVBRGs) to extract representations of brain regions from different perspectives. Finally, we design multi-view sample GCNs (MVSGs) that incorporate phenotypic data as knowledge embedding to discover critical features of samples for final dichotomous classifications. The experimental results based on the Parkinson's Progression Markers Initiative (PPMI) data source demonstrate that our approach achieves an average accuracy of 88.74% in PD diagnosis, exhibiting state-of-the-art performance.