Graph Convolutional Networks and Functional Connectivity for Identification of Autism Spectrum Disorder

Abstract

The purpose of this study is to apply graph convolutional networks (GCNs) for feature extraction and classification of patients with autism spectrum disorder (ASD). The number of people with (ASD) increases every year and poses a threat to the life and future of many children which makes this study very important. We used the resting-state fMRI data from a large multi-site dataset called Autism Brain Imaging Data Exchange I (ABIDE I) to validate our proposed approach. Based on functional connectivity (FC), we represented the brain through a complex network where the regions of the brain represent the nodes in the network and the correlation coefficient between two regions represents the weight of the edge connects them. The data were preprocessed, and we constructed a functional connectivity graph for each subject by parcellation of the whole brain into 392 distinct regions using the (CC400) atlas. The graph measures were then calculated and used as features for both nodes and edges to classify these subjects by graph convolutional networks' classifier which proposed in this study. The results we achieved in our experiments were with accuracy of 70% to identify patients with autism spectrum disorder from healthy individuals, which proved the accuracy and robustness of our approach in classifying brain diseases.