Topology-Enhanced Multi-Modal Residual Graph Convolutional Network for Autism Spectrum Disorder Identification

Functional Brain Networks (FBNs) have shown promise in diagnosing neurological disorders, particularly Autism Spectrum Disorder (ASD). Currently, Graph Convolutional Networks (GCNs) excel in processing FBN data, making them instrumental in learning from functional Magnetic Resonance Imaging (fMRI) and analyzing ASD. Despite their advantages, existing methods primarily concentrate on the structural properties of brain networks, often overlooking the essential topological information. To bridge this gap, we introduce a Topologically Enhanced Multi-Pattern Residual Graph Learning (TEMPRGL) framework. Our approach begins by constructing diverse brain networks from fMRI data, capturing a range of connectivity patterns. We then employ a dual-feature extraction strategy on FBNs: one that utilizes multiple Residual GCN modules to generate multi-modal graph representations, which are subsequently integrated; and another focuses on combining topology information from FBN. To validate our TEMPRGL, we exeperiment on the Autism Brain Imaging Data Exchange (ABIDE) dataset. It attained an impressively high accuracy rate of 92.5%., underscoring its effectiveness in neurological disorder diagnosis.