Uncertainty-Aware Evidential Multi-Modal Fusion Network for Large-Scale Multi-Center Major Depressive Disorder...

Area of Science:

• Neuroscience
• Artificial Intelligence
• Medical Imaging

Background:

• Clinical diagnosis of major depressive disorder (MDD) relies on subjective assessments, necessitating objective automated methods.
• Existing neuroimaging fusion techniques struggle with modality reliability, uncertainty quantification, and multi-center site effects.

Purpose of the Study:

• To develop an uncertainty-aware evidential multimodal fusion network for large-scale, multi-center MDD classification.
• To address limitations in dynamic modality assessment, prediction uncertainty, and generalizability across different research sites.

Main Methods:

• Proposed an uncertainty-aware evidential multimodal fusion network incorporating deep functional feature extractors (DFFE) and dual-stream hierarchical feature fusion (DS-HFF) for fMRI and sMRI.
• Utilized Dempster-Shafer theory (DST) for an Evidential Multimodal Fusion (EMF) strategy to represent modality evidence with explicit uncertainty.
• Implemented site-adversarial regularization for learning site-invariant features and evaluated using a leave-one-site-out cross-validation (LOSO-CV) on the REST-meta-MDD dataset (1,601 subjects, 16 sites).

Main Results:

• The proposed method outperformed existing approaches on the REST-meta-MDD dataset.
• Demonstrated superior uncertainty calibration compared to Softmax and MC Dropout baselines.
• An uncertainty-based rejection mechanism enhanced classification performance, and interpretability analysis identified key brain regions consistent with MDD pathophysiology.

Conclusions:

• The developed framework offers a robust, trustworthy, and interpretable solution for multi-center MDD diagnosis.
• The uncertainty-aware approach effectively handles multi-center data challenges and improves diagnostic reliability.
• Findings support the potential of AI-driven neuroimaging analysis for objective MDD assessment.