Exploring graph contrastive learning for brain disorder analysis

<p dir="ltr">Increasingly, deep learning is being integrated into scientific discovery and, as such, this powerful technology is augmenting and accelerating research. One of the most recent applications of deep learning has been neuroscience, which is attracted increasing interest and showing great potential for discovering abnormalities in the brain. However, the scarcity of data and noisy labels are hampering prediction quality with today’s methods. Moreover, explainability, which is a significant limitation of artificial intelligence, is a huge issue in neuroscientific research that is currently resulting in suboptimal analyses of brain graphs. To tackle these challenges, we propose a cutting-edge contrastive learning framework specifically designed for brain graphs that includes an interpretable learning kernel for brain graph mining. This framework departs from traditional graph contrastive learning methods by augmenting the data using a mechanism based on counterfactual thinking. Notably, this method ensures the brain graphs generated will be legal. Moreover, a novel contrastive loss enhances learning by emphasising contrasts both within and between classes, i.e., it emphasises both the intra- and inter-class contrasts. Most importantly, an interpretable brain graph learning kernel accentuates the most important regions and substructures of the brain graphs for subsequent analysis and, based on a comparison with the literature, may even provide novel insights into neuroscience. An evaluation involving disorder prediction and salient region analysis with three real-world brain disorder datasets demonstrates superior performance compared to existing methods, as well as, the method to be highly efficient at identifying critical regions and connections within the brain.</p>