Classifier retraining with decoupled federated learning for imbalanced medical image classification

Class imbalance is a major challenge in federated learning (FL) for medical image classification due to the varying prevalence of disease, which is further exacerbated by data heterogeneity from diverse imaging devices and client distributions. Existing FL methods typically follow a coupled learning paradigm to address this issue, where feature representation and classifier are optimized jointly, inevitably causing a trade-off, i.e., improving minority-class accuracy at the cost of majority-class performance, limiting clinical reliability. To tackle this, we propose Federated Decoupling (FedDec), a novel framework that decouples representation learning from classifier retraining to achieve balanced multi-class performance. Specifically, FedDec first employs a Bi-level Curriculum Learning (BCL) strategy that dynamically prioritizes historically misclassified samples at the sample level and poorly calibrated classes at the class level, yielding more balanced and discriminative feature representations. It then introduces a Class-aware Gaussian Prototype Learning (CGPL) module that models features from BCL as Gaussian distributions, effectively capturing intra-class variation and enabling privacy-preserving synthetic feature generation. Finally, a Personalized Classifier Retraining (PCR) module leverages these prototypes to generate class-balanced features for local classifier retraining, thereby effectively mitigating bias. Experiments on Fed-ISIC2019 and Camelyon17 demonstrate that FedDec outperforms twelve state-of-the-art baselines, improving overall and balanced performance while alleviating the minority–majority trade-off.