Asymmetric Mutual Learning for Decentralized Federated Medical Imaging

Rigorous medical data privacy laws significantly obstruct data collection necessary for centralized machine learning, creating a substantial barrier to medical artificial intelligence (AI) application advancements. Centralized federated learning offers a potential solution by enabling collaborative model training without direct data exchange. However, it encounters notable challenges, including reliance on a trusted central server and the requirement for participants to disclose model information explicitly. These issues can hinder widespread adoption, especially in sensitive domains such as healthcare, where data privacy and security are paramount. To address these challenges, this study introduces a novel decentralized federated learning framework, named FedDAL, designed to enhance privacy and efficiency in collaborative model training. Our framework innovatively maintains local private models and proxy models through asymmetrical mutual learning, tailored to consider the varying sizes and knowledge bases of different models involved. This asymmetrical approach ensures that local models can remain private, while only compact proxy models are shared during the iterative update process, significantly reducing the risk of sensitive data leakage and communication overhead. We validate the efficacy and practicality of our proposed framework through comprehensive experiments on medical image classification and segmentation tasks. Our results demonstrate that FedDAL not only achieves superior performance compared to traditional centralized federated learning methods but also offers a more secure and efficient approach. This framework represents a significant advancement in federated learning, providing a viable pathway for integrating AI into medical practice while upholding stringent privacy standards.