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

Early diagnosis of bipolar disorder (BD) is challenging due to unclear pathological features. Structural MRI (sMRI) is crucial for BD diagnosis due to its high spatial resolution and stable image quality. However, current sMRI-based methods often fail to capture specific brain region pathology effectively. To address this problem, we proposed CognitMoE, a classifier based on a cognition-aware collaborative multi-expert network, which employs brain-region-level imaging features to accurately capture the pathological changes of BD patients from the perspective of brain region structure and cognitive function. Unlike traditional approaches, our brain-region-level method reduces computational complexity. Specifically, the Cognition-Aware Attention Module (CAAM) leverages brain atlas knowledge to extract and weight brain region features, emphasizing brain regions associated with cognitive dysfunction. The Collaborative Multi-Expert Network (CMEN) is a Mixture of Experts model designed to simulate brain region collaboration during cognitive tasks, uncovering relevant pathological changes. Experiments show CognitMoE outperforms existing methods on OpenfMRI and a collected dataset, with BACC, F1 score, and sensitivity improving by 7.4 %, 3.4 %, and 12.5 %, respectively. These results highlight CognitMoE's ability to better distinguish BD patients from healthy controls and reduce misdiagnosis, supporting early BD diagnosis and clinical screening. The code is available at https://github.com/Xiaotong-Zhu/CognitMoE.