MTIF-GNN: A Multi-Modal Topology and Integrated Focused Network for ASD Diagnosis

The diagnosis of autism spectrum disorder (ASD) has been a longstanding focus in clinical and neuroscience research, particularly with the growing use of multi-site functional MRI (fMRI) data. However, most existing methods often suffer from limitations such as inadequate multi-modal data fusion and insufficient mining of implicit features. To address these challenges, an end-to-end multi-modal topology-integrated graph neural network (MTIF-GNN) is proposed for multi-site ASD diagnosis. MTIF-GNN constructs two complementary population graphs to fully extract and integrate both explicit and implicit features, aiming to achieve more accurate diagnostic results. To further enhance network performance, an adaptive topology update (ATU) module is designed to dynamically adjust and optimize the graph structure within MTIF-GNN, enabling more effective capture of latent relationships among nodes and global topological patterns. Finally, this study achieves deep integration of explicit and implicit features through joint learning optimization of the graph neural network, and obtains the ASD diagnosis results using a multi-layer perceptron. Experimental results demonstrate that the proposed method outperforms existing approaches and also achieves excellent performance in diagnosing major depressive disorder (MDD), indicating its broad applicability across various psychiatric disorders and providing effective support for cross-disorder brain function research. The code is available at https://github.com/cvmdsp/MTIF-GNN.