Autism Spectrum Disorder Diagnosis Using Graph Neural Network Based on Graph Pooling and Self-adjust Filter

Currently, accurately identifying autism spectrum disorders (ASD) still have challenges. However, graph neural network used for ASD diagnosis only focus on the part of abnormal brain functional connectivity, ignoring the effects between brain regions and the help of phenotypic information, and the kernel is also pre-defined. To solve the above problems, a graph neural network based on self-attention graph pooling and self-adjust filter is proposed for ASD diagnosis. Specifically, first, self-attention graph pooling is used for feature extraction of the fMRI to account for the influence of nodes with abnormal brain functional connectivity and activity between brain regions in fMRI data. Then, image features were taken as graph nodes and phenotypic information as edges to form a population graph. Finally, to focus on both high and low frequency information in the graph, a graph neural network based on self-adjust filter is used to learn node embedding. Experimental results on ABIDE I showed the effectiveness of the proposed method.