Gradient Matching Federated Domain Adaptation for Brain Image Classification.

Area of Science:

• Artificial Intelligence
• Neuroscience
• Medical Imaging

Background:

• Federated learning (FL) enables collaborative model training on decentralized data, crucial for sensitive medical information like brain images.
• Domain shift, variations in data characteristics across different sites, can significantly degrade federated model performance.
• Existing federated learning methods struggle to address domain shift effectively in multisite brain imaging studies.

Purpose of the Study:

• To propose a novel Gradient Matching Federated Domain Adaptation (GM-FedDA) method to address domain shift in federated brain image analysis.
• To enhance the robustness and performance of local federated models trained on multisite neuroimaging data.
• To improve privacy-preserving, collaborative analysis of medical images across different institutions.

Main Methods:

• Developed a two-stage approach: pretraining with One-Common-Source Adversarial Domain Adaptation (OCS-ADA) and fine-tuning with Gradient Matching Federated (GM-Fed) learning.
• OCS-ADA utilizes a public dataset and gradient matching loss to pretrain encoders, reducing domain shift at each private site.
• GM-Fed fine-tuning updates local models by minimizing gradient matching loss, aligning optimization directions across sites.

Main Results:

• The proposed GM-FedDA method demonstrated superior performance compared to existing methods in diagnostic classification tasks.
• Validated on multisite resting-state functional MRI (fMRI) data for schizophrenia and major depressive disorder classification.
• Achieved robust local federated models by effectively reducing domain discrepancy and preserving data privacy.

Conclusions:

• GM-FedDA offers a promising solution for multisite brain imaging analysis, overcoming challenges posed by domain shift.
• The method has significant potential for applications requiring collaborative analysis of sensitive, decentralized data while ensuring privacy.
• Highlights the effectiveness of gradient matching techniques in federated domain adaptation for medical image classification.