Graph Self-Supervised Learning With Application to Brain Networks Analysis.

Area of Science:

• Neuroscience
• Machine Learning
• Medical Imaging

Background:

• Deep supervised models for brain disease diagnosis are limited by insufficient training data and supervision.
• Developing robust learning frameworks for brain networks, which are non-Euclidean graph data, is crucial.

Purpose of the Study:

• To address limitations in deep learning for brain disease diagnosis by proposing a self-supervised learning framework.
• To generalize self-supervised learning to non-Euclidean brain network data.

Main Methods:

• Proposed BrainGSLs, an ensemble masked graph self-supervised framework.
• Incorporated a topological-aware encoder, a node-edge bi-decoder, and a signal representation learning module.
• Evaluated on Autism Spectrum Disorder (ASD), Bipolar Disorder (BD), and Major Depressive Disorder (MDD) diagnosis.

Main Results:

• BrainGSLs demonstrated remarkable improvement over state-of-the-art methods in brain disease diagnosis.
• The framework successfully identified disease-associated biomarkers consistent with prior research.
• A strong association was found between Autism Spectrum Disorder (ASD) and Bipolar Disorder (BD).

Conclusions:

• Self-supervised learning, particularly with masked autoencoders on brain networks, offers a promising approach for disease diagnosis.
• BrainGSLs provides an effective method for capturing information from limited data and insufficient supervision in neuroimaging.
• The findings highlight the potential of AI in understanding complex neurological and psychiatric disorders.