Self-Supervised Graph Contrast Learning Improves Functional Brain Network Accuracy for Depression Identification

A self-supervised graph contrast model for functional brain networks improved major depressive disorder identification accuracy by 3% and 5% at two sites versus prior state of the art. Adapted Grad-CAM flagged key frontal, limbic, striatal, thalamic, and insular regions.

• Uni-GCL uses self-supervised graph contrast learning, intended to reduce reliance on large labeled functional brain network training datasets.
• Reported accuracy gains over prior state-of-the-art methods were 3% at Site 21 and 5% at Site 1.
• The adapted Grad-CAM highlighted frontal opercular, caudate, fusiform, amygdala, putamen, thalamus, triangular inferior frontal, insula, and hippocampus regions as significant contributors to major depressive disorder identification.

Automated summary

RadPigeon summaries are original and for information only. They are not clinical advice.