MM-TransUNet: Multi-modal and multi-graph feature learning for neurodevelopmental disorder diagnosis

Graph learning advances the feature learning of brain network topology for neurodevelopmental disorder diagnosis. However, existing methods struggle with multi-modal feature heterogeneity and insufficient modeling of complex graph structures, which limits their representation capability and generalizability. We proposed MM-TransUNet, a multi-modal and multi-graph feature learning framework. The Graph Token Statistics Transformer (GTST) module was designed to learn global structural features and explicitly encode graph topology by orthogonal projection and second-order statistics. A TransUNet encoder was designed to learn multi-scale features from the population graph and the correlation synergy graph. It integrated the GTST module with the hierarchical architecture of Graph UNet to improve the interaction between global and local features. Furthermore, a distribution-aware dynamic fusion (DADF) module was designed to adaptively fuse features via a confidence factor-based dynamic weighting strategy. In this module, multi-modal distribution discrepancies are explicitly modeled to regulate modality contributions, which improves the consistency and robustness of multi-modal representations. Evaluations on the ABIDE and ADHD-200 datasets demonstrate that MM-TransUNet outperforms state-of-the-art methods, achieving superior diagnostic accuracy and strong generalization across distinct subject cohorts. The source code is released at https://github.com/Roninddd/MM-TransUNet.