GP-MR-GCN: A Novel Framework for ASD Diagnosis Using Functional Brain Networks

Functional brain networks analysis is a powerful tool for diagnosing autism spectrum disorder (ASD). The resting-state functional magnetic resonance imaging (rs-fMRI) along with non-image information facilitates ASD classification based on graph neural network (GNN). However, this strategy has two critical limitations. First, paying more attention to aberrant functional connectivity patterns while ignoring local regional activity. Second, it is hard to capture multi-scale features of nodes from classical GNN architectures. To address these challenges, a graph pooling multi-scale residual graph convolutional network (GP-MR-GCN) is proposed in this work. The model involves three key processes. (1) Modeling rs-fMRI data as graph structures, and extracting substructural patterns through unsupervised graph pooling. (2) Incorporating non-image information to adaptively recalibrate and optimize node feature distributions. (3) Learning discriminative graph representations across varying topological resolutions based on the multi-scale residual graph convolutional network. The experimental results on the ABIDE-I dataset demonstrated that the proposed model achieves superior ASD diagnostic accuracy over the baseline methods.