Universum Inspired Graph Contrast Learning for Autism Spectrum Disorder Identification

The functional brain network (FBN) has recently emerged as a pivotal biomarker in the diagnosis of autism spectrum disorder (ASD). Meanwhile, graph convolutional networks (GCNs) represent a promising strategy for the concurrent extraction of FBN features and the facilitation of ASD identification. Unfortunately, the efficacy of existing GCN methods is often limited by their reliance on extensive annotated datasets. To address this limitation, we propose the Universum Inspired Graph Contrast Learning (Uni-GCL) framework, a novel approach to brain disease analysis that capitalizes on the generation of Universum samples for the pre-training of the GCN model. Specifically, we use a Mixup technique to generate universum samples and design a universum-inspired contrastive loss function. This methodology necessitates only a modest number of samples for finetuning the pre-trained model, effectively reducing data requirements while simultaneously boosting model performance. To systematically assess the performance of Uni-GCL, we conducted rigorous experiments on the Autism Brain Imaging Data Exchange (ABIDE) initiative for ASD identification. The results demonstrate that Uni-GCL, augmented with universum samples, yields compelling outcomes, underscoring its potential as a robust method in the diagnostic arsenal for ASD.