Dynamic multi-scale topology learning and cross-modal graph fusion for autism spectrum disorder diagnosis

Multi-modal fusion methods have demonstrated potential in integrating complementary information from different data modalities. However, existing methods often rely on static fusion strategies, failing to capture the heterogeneous interactions across modalities. To address this challenge, we propose the dynamic multi-scale topology learning and cross-modal interaction graph fusion network (DMTL-Net) for autism spectrum disorder (ASD) diagnosis. The model is designed to extract distinctive and complementary features from brain functional and structural networks while capturing their dynamic topological organization. Specifically, the structural radiomics similarity network (RSN) and functional connectivity network (FCN) are constructed from structural magnetic resonance imaging (sMRI) and functional magnetic resonance imaging (fMRI), respectively. Multi-modal graph initialization is performed by the cross-modal inclusive integration and consensus filtering to balance sensitivity and specificity. Then we propose the dynamic topology learning mechanism that integrates multi-scale feature extraction with adaptive channel-wise and spatial-wise fusion, effectively capturing complex inter-modal interactions. Additionally, focal loss is used to reduce the contribution of easy-to-classify samples and emphasize hard examples during optimization. Evaluated on the Autism Brain Imaging Data Exchange dataset, the proposed DMTL-Net demonstrates superior performance and consistently outperforms the state-of-the-art methods.