A Novel Dual-Level Multi-Graph Convolutional Network for Parkinson's Disease Diagnosis

Parkinson's disease (PD) is a common and complex neurodegenerative disorder. Its early discovery based on brain network analysis is critical for conducting effective clinical intervention. However, most existing methods depend mainly on single-modality information due to the lack of multimodal data. In this work, we utilize dual-level i.e. individual data and population data collected from fMRI and DTI imaging modalities respectively, to design a knowledge-driven dual-level multi-graph convolutional network (K-DMGCN) for early detection of PD. On the one side, we construct a knowledge-guided functional graph based on neuroanatomical priors, a data-guided functional graph via self-attention, and a structural graph from structural connections, aiming to mine brain connectivity relationships at the individual level. On the other side, cross-modal similarity graphs are built by integrating phenotypic data with image features, aiming to find common pathological patterns across different patients through adaptive graph convolution. A Top-K pooling operation is introduced to screen critical brain regions to increase the model's interpretability. In addition, we design a structural-functional discrepancy loss function to reduce multimodal information redundancy. Experimental results on the dataset from the Affiliated Hospital of Qingdao University demonstrated that the proposed K-DMGCN achieved superior classification performance over state-of-the-art methods. It achieved an accuracy of 90.21% for PD classification. Furthermore, we found that the most PD-relevant brain regions can be identified through cross-modal collaborative learning, which may provide a fresh understanding to explore the pathological mechanisms of PD.