Classifying Major Depressive Disorder Using Multimodal MRI Data: A Personalized Federated Algorithm.

Area of Science:

• Neuroscience
• Artificial Intelligence
• Medical Imaging

Background:

• Accurate diagnosis of major depressive disorder (MDD) relies on neuroimaging, but multisite data present heterogeneity and privacy challenges.
• Sharing raw MRI data is restricted due to ownership, security, and privacy concerns, hindering robust diagnostic model development.
• Federated learning (FL) provides a privacy-preserving approach for collaborative model training across sites without raw data sharing.

Purpose of the Study:

• To develop a privacy-aware federated learning framework for scalable, multisite diagnosis of major depressive disorder (MDD) using multimodal MRI.
• To address domain shift issues inherent in multisite neuroimaging data.
• To enhance the integration of structural MRI (sMRI) and functional MRI (fMRI) for improved MDD classification.

Main Methods:

• Proposed the personalized Federated Gradient Matching and Contrastive Optimization (pF-GMCO) algorithm.
• Incorporated gradient matching with cosine similarity for adaptive site contribution weighting.
• Utilized contrastive learning for client-specific model optimization and multimodal compact bilinear (MCB) pooling for feature integration.

Main Results:

• Evaluated pF-GMCO on the Rest-Meta-MDD dataset comprising 2293 subjects from 23 sites.
• Achieved a diagnostic accuracy of 79.07% for major depressive disorder (MDD).
• Demonstrated superior performance and interpretability compared to existing methods.

Conclusions:

• pF-GMCO offers an effective and privacy-aware framework for multisite MDD diagnosis using federated learning.
• The approach successfully addresses domain shift and integrates multimodal MRI data.
• This method facilitates collaborative research and development of diagnostic tools for mental health disorders.