Multi-Modal Brain Graph Learning of Shared-Specific Features for Schizophrenia Diagnosis

Recent years have witnessed promising progress in schizophrenia diagnosis with multi-modal brain networks. However, existing works usually focus on either learning the specific characteristics of each modality or exploiting the shared information among modalities. Few methods explicitly explore the shared information as well as preserve the modality-specific characteristics. In this paper, we propose a shared-specific graph learning (S²GL) framework for schizophrenia diagnosis, which benefits disease diagnostic performance by exploiting shared information and characteristics unique to each modality. Specifically, two modality-specific graph learning modules are adopted to learn multi-level specific representations of the structural brain graph and functional brain graph, respectively. Meanwhile, a structural-functional shared graph learning module is designed to mine the shared features between modalities by exploiting their correlations, providing cross-enhanced representations for each modality. These representations can not only provide beneficial information for mining high-order modality-specific features but also be fused to generate shared representations for disease diagnosis. Next, we develop a shared-specific graph feature aggregation module to integrate the multi-level specific and shared representations, which can provide comprehensive multi-modal information to boost disease diagnosis. Finally, we formulate a unified end-to-end framework for schizophrenia diagnosis. Experiments on an open schizophrenia dataset demonstrate that our S²GL is effective and superior to cutting-edge methods.