Solving Class Imbalance in Medical Image Classification Based on Federated Learning

Federated Learning (FL) has excellent privacy-preserving characteristics and is widely used in the medical field. However, medical data often suffer from class imbalance, which can affect the performance of models. In this paper, we first introduce a novel privacy-preserving federated learning scheme that addresses the class imbalance problem through both feature learning and classifier learning. In the feature learning stage, two feature extraction networks are used in parallel and optimized. Finally, in the classifier learning stage, multiple loss functions are designed to ensure balanced learning across all classes. Experimental results on a publicly available skin dataset with simulated multi-source medical image data demonstrate outstanding performance in handling class imbalance. The model achieves balanced classification accuracy, overall accuracy, and F1 score of 87.02%, 88.27%, and 87.33%, respectively. Compared to other models, the balanced accuracy, overall accuracy, and F1 score are improved by 8.71%, 3.61%, and 8.23%, respectively.