CognitMoE: A cognition-aware collaborative multi-expert network for bipolar disorder diagnosis

Area of Science:

• Neuroimaging
• Artificial Intelligence
• Psychiatric Disorders

Background:

• Early diagnosis of bipolar disorder (BD) is hindered by unclear pathological features.
• Structural MRI (sMRI) is valuable for BD diagnosis but current methods struggle to capture specific brain region pathology.
• Existing sMRI approaches often lack the granularity to effectively identify subtle structural and functional brain alterations in BD.

Purpose of the Study:

• To develop an advanced classifier, CognitMoE, for accurate early diagnosis of bipolar disorder (BD).
• To leverage brain-region-level imaging features and cognitive function insights for improved BD detection.
• To reduce computational complexity compared to traditional sMRI-based diagnostic methods.

Main Methods:

• Proposed CognitMoE, a cognition-aware collaborative multi-expert network using brain-region-level imaging features.
• Implemented a Cognition-Aware Attention Module (CAAM) to extract and weight brain region features based on atlas knowledge.
• Utilized a Collaborative Multi-Expert Network (CMEN) to simulate brain region interactions and uncover pathological changes.

Main Results:

• CognitMoE demonstrated superior performance over existing methods on two datasets (OpenfMRI and a collected dataset).
• Achieved significant improvements in BACC (7.4%), F1 score (3.4%), and sensitivity (12.5%).
• Effectively distinguished between BD patients and healthy controls, highlighting its potential for early diagnosis and screening.

Conclusions:

• CognitMoE offers a promising approach for early bipolar disorder diagnosis by accurately capturing subtle brain structural and cognitive alterations.
• The brain-region-level analysis reduces computational load while enhancing diagnostic precision.
• This method supports clinical screening and reduces misdiagnosis rates, aiding timely intervention for BD patients.