Self-expression property theory guided multi-modal brain graph learning

Integrating multi-modal brain graphs for brain disorder analysis has attracted increasing attention recently. However, existing multi-modal brain graph learning methods suffer from two important problems: first, they randomly use simple connectivity profiles as attribute features of brain graphs, causing an insufficient description of brain connectivity networks. In addition, unobvious differences between healthy and patient brain graphs require to learn discriminative feature representations of single-modal brain graphs, which is rarely considered in previous methods. Second, mining the complex relationship between modalities is significant for achieving deep fusion, but most of the existing methods adopt a simple fusion way, like the attention mechanism, and cannot recognize important brain regions and connections associated with disorders for interpretable analysis. To address the problems above, in this work, we integrate cross-scale contrastive graph neural networks and self-expression property theory to learn multi-modal and multi-attribute brain graph representations for disorder analysis. Specifically, we extend multi-attribute features for characterizing brain graphs and design cross-scale contrastive graph neural networks to handle multi-attribute brain graphs to enhance discriminative multi-level representations for each modality. Then, we introduce the self-expression property theory to innovatively extract the complementary relationship of multi-modal brain graphs while generating the consensus graph to learn the common feature. The consensus graph is an interpretable brain graph to explore salient brain regions and connections related to disorders. We adaptively fuse the common feature and each modality feature using the attention mechanism to learn optimal representations of multi-modal brain graphs for disorder prediction. Extensive experiments on three disorder datasets verify the effectiveness of our method, especially pathogenic analysis results are in line with medical findings, which has the potential to promote brain disorder research.