Unified Bias Reduction in Federated Learning: A Two-Stage Feature and Label Space Disentanglement in Medical Image Analysis

Federated learning, as a typical ensemble distributed machine learning framework, enables collaborative model training across decentralized data sources while preserving data privacy, which is crucial in domains like medical images. However, due to discrepancies in imaging equipment, protocols, and patient demographic information, particularly the diversity in manufacturers, models, and imaging parameter configurations of medical electronic devices, client data often exhibits significant heterogeneity, leading to biases in both the feature space and the label space. These biases degrade model performance and generalization. To address this issue, we propose a novel Dual-Space Debiasing Framework (FedDSDF) for medical electronic data, which jointly tackles feature-space and label-space biases by integrating Feature Space Debiasing (FSD) and Label Space Debiasing (LSD). Specifically, FSD is achieved via prototype-guided alignment with personalized feature storage in federated learning, while LSD is realized through distribution-aware knowledge distillation. We theoretically demonstrate the convergence of our framework and validate its effectiveness through extensive experiments on 4 public medical image datasets and a large-scale private fetal ultrasound dataset collected from five hospital centers in a real-world application. Experimental results demonstrate that FedDSDF outperforms state-of-the-art FL methods, enhancing accuracy and robustness across heterogeneous data distributions.