Multi-Task Collaborative Pre-Training and Adaptive Token Selection: A Unified Framework for Brain Representation...

Area of Science:

• Neuroscience
• Artificial Intelligence
• Medical Imaging

Background:

• Structural magnetic resonance imaging (sMRI) is crucial for understanding brain structure.
• Existing deep learning models often overlook the brain's cognitive functions, focusing solely on anatomical attributes.
• Representing the brain requires capturing subtle, distributed information linked to individual cognitive differences.

Purpose of the Study:

• To develop a brain representation learning framework, MCPATS, that captures cognition-related information from sMRI.
• To address the limitations of previous models by integrating fine-grained local details with global structural context.
• To improve the accuracy of brain disease diagnosis through enhanced representation learning.

Main Methods:

• MCPATS combines Multi-task Collaborative Pre-training (MCP) and Adaptive Token Selection (ATS).
• MCP utilizes mask-reconstruction, distort-restoration, adversarial learning, and age-prediction for progressive representation learning.
• ATS employs mutual attention to highlight discriminative features during downstream tasks.

Main Results:

• MCPATS demonstrated superior performance in brain disease diagnosis across three public datasets compared to existing methods.
• The framework successfully learned representations that capture cognition-related information, validated through detailed analysis.
• The proposed method effectively integrates local and global brain structure features for improved diagnostic accuracy.

Conclusions:

• MCPATS offers a robust framework for learning brain representations that incorporate cognitive variability.
• The model's ability to capture subtle, distributed information enhances its utility in neuroscience research and clinical applications.
• MCPATS represents a significant advancement in leveraging sMRI data for understanding brain structure and function.