A Cross-Feature Mutual Learning Framework to Integrate Functional Connectivity and Activity for Brain Disorder...

Area of Science:

• Neuroimaging
• Machine Learning
• Brain Disorders

Background:

• Functional magnetic resonance imaging (fMRI) derived time courses (TC) and functional network connectivity (FNC) are valuable for studying brain disorders.
• Current deep learning methods often fail to integrate TC and FNC effectively, lacking end-to-end mixed feature learning frameworks.

Purpose of the Study:

• To introduce a novel cross-feature mutual learning (CFML) framework for enhanced brain disorder classification.
• To enable collaborative learning and mutual knowledge transfer between TC-specific and FNC-specific models.

Main Methods:

• Developed a recurrent neural network-based encoder for TC and a transformer-based encoder for FNC.
• Designed a cross-modal module for adaptive feature integration.
• Implemented CFML strategy with feature-specific, feature-exchange, and joint losses for collaborative training.

Main Results:

• CFML achieved 85.1% accuracy in distinguishing schizophrenia (SZ) patients from healthy controls (HC).
• Outperformed 12 comparative models by 3.0-9.2% accuracy.
• Demonstrated superior performance using combined TC and FNC features.

Conclusions:

• CFML effectively integrates complementary fMRI features for improved brain disorder classification.
• The proposed framework shows significant potential for advancing the diagnosis and understanding of brain disorders.