Diagnosis of Autism Spectrum Disorder Based on Contrastive Functional Connectivity Graph Learning Network

Abstract

To reduce the dependence on tagged data, we proposed a Contrastive Functional Connectivity Graph Learning Network (CFCG-Net) for the diagnosis of autism spectrum disorder. CFCG-Net is mainly composed of three parts: construction of contrastive Functional Connection (FC) graphs, learning of contrastive FC graphs, and dynamic graph classification based on population graph. Firstly, we constructed contrastive FC graphs for each subject based on the original brain functional connectivity. Secondly, we constructed a graph convolutional network to learn the contrastive FC graphs to obtain the contrastive embeddings of each subject, and further defined the population graph through the contrastive embeddings. Finally, we trained a dynamic graph classifier to predict the class probability of each subject. CFCG-Net is tested on the ABIDE I dataset, which verified the effectiveness of CFCG-Net.