Dynamic Graph Contrastive Learning for Neuro Disease Identification using Resting-State fMRI

Resting-state functional magnetic resonance imaging (rs-fMRI) techniques provide valuable insights into the connectivity of brain functions during the resting state. Current machine learning and deep learning methodologies have been employed to analyze rs-fMRI data for disease diagnosis; however, these existing approaches necessitate substantial amounts of data. To address this challenge, we propose a Dynamic Graph Contrast Learning (DGCL) framework based on functional MRI for analyzing neurological disorders. The DGCL framework consists of a pre-training phase aimed at generating information-rich fMRI representations using unlabeled training data, followed by a fine-tuning phase dedicated to downstream disease recognition tasks. Specifically, in the pre-training phase, we first apply data augmentation techniques to enhance the fMRI dataset by augmenting blood oxygen level dependent (BOLD) signals. Subsequently, we construct dynamic graphs from the augmented data and utilize two graph encoders for feature extraction from these constructed dynamic graphs through comparative learning. The fine-tuning phase involves training a neural network specifically for disease recognition tasks. Experiments conducted on a rs-fMRI dataset comprising 92 subjects demonstrate that the DGCL framework exhibits strong performance in identifying brain diseases.