A Graph Attention Neural Network for Diagnosing ASD with fMRI Data

Abstract

Autism Spectrum Disorder (ASD) is a common psychiatric disorder disease that typically causes impaired communication and compromised social interactions. Functional magnetic resonance imaging (fMRI) data is one of the common neuroimaging modalities for understanding human brain functionalities as well as the diagnosis and treatment of brain disorders. There are many successful applications of deep learning to fMRI analytics where fMRI data is mostly considered structured Euclidean grids or time series, and features were usually extracted from the computer vision or time series perspective. Graph neural networks (GNNs) are neural networks that learn the interactions of graphs via message passing between the nodes of graphs. Recently variants of GNNs such as graph attention network (GAT) have demonstrated outstanding performances on many machine learning tasks. In this paper, we proposed a connectivity based graph attention network for autism diagnosis using functional connectivity (FC) patterns obtained from resting-state fMRI (rs-fMRI). We define graphs based on F Cs and statistics of fMRI time series and present a connectivity-based GATs model for fMRI data analysis. The graph-theoretic based approach enables us to pass messages among connectomic (Non-Euclidean) neighborhoods, which is consistent with the brain functional networks. To evaluate the performance of our proposed GAT, we apply the GAT model to the classification of A SD patients from normal controls. Our results show that GAT can effectively capture salient features for ASD classifications in fMRI analysis.