Federated Learning with Imbalanced and Agglomerated Data Distribution for Medical Image Classification

Federated learning (FL), training deep models from decentralized data without privacy leakage, has drawn great attention recently. Two common issues in FL, namely data heterogeneity from the local perspective and class imbalance from the global perspective have limited FL's performance. These two coupling problems are under-explored, and existing few studies may not be sufficiently realistic to model data distributions in practical sceneries (e.g. medical sceneries). One common observation is that the overall class distribution across clients is imbalanced (e.g. common vs. rare diseases) and data tend to be agglomerated to those more advanced clients (i.e., the data agglomeration effect), which cannot be modeled by existing settings. Inspired by real medical imaging datasets, we identify and formulate a new and more realistic data distribution denoted as L2 distribution where global class distribution is highly imbalanced and data distributions across clients are imbalanced but forming a certain degree of data agglomeration. To pursue effective FL under this distribution, we propose a novel privacy-preserving framework named FedIIC that calibrates deep models to alleviate bias caused by imbalanced training. To calibrate the feature extractor part, intra-client contrastive learning with a modified similarity measure and inter-client contrastive learning guided by shared global prototypes are introduced to produce a uniform embedding distribution of all classes across clients. To calibrate the classification heads, a softmax cross entropy loss with difficulty-aware logit adjustment is constructed to ensure balanced decision boundaries of all classes. Experimental results on publicly-available datasets demonstrate the superior performance of FedIIC in dealing with both the proposed realistic modeling and the existing modeling of the two coupling problems.