BrainMass: Advancing Brain Network Analysis for Diagnosis With Large-Scale Self-Supervised Learning.

Area of Science:

• Neuroscience
• Artificial Intelligence
• Medical Image Analysis

Background:

• Foundation models excel in self-supervised learning for diverse tasks.
• Medical data's heterogeneity and collection challenges benefit from foundation models.
• Limited research exists on brain network foundation models, hindering their application.

Purpose of the Study:

• To address the gap in brain network foundation models.
• To enhance adaptability and generalizability in neuroscience research.
• To develop a versatile framework for analyzing brain networks.

Main Methods:

• Curated a large dataset (70,781 samples, 46,686 participants) from 30 sources.
• Introduced pseudo-functional connectivity (pFC) for data augmentation.
• Proposed the BrainMass framework with Mask-ROI Modeling (MRM) and Latent Representation Alignment (LRA) for self-supervised learning.

Main Results:

• BrainMass achieved superior performance on eight internal and seven external brain disorder diagnosis tasks.
• Demonstrated significant generalizability and adaptability across diverse neuroscience applications.
• Showcased powerful few/zero-shot learning capabilities.

Conclusions:

• BrainMass offers a robust foundation model for brain network analysis.
• The framework exhibits potential for clinical applications due to its interpretability in disease contexts.
• Highlights the efficacy of self-supervised learning for advancing neuroscience research.