Federated learning on long-tailed datasets

Federated learning has become a promising approach for decentralized training across different domains. However, when it comes to medical imaging, its application often faces the challenge of dealing with long-tailed distributions. These distributions are characterized by an imbalance between majority and minority classes, which is commonly observed in medical imaging tasks, such as skin lesion classification and gastrointestinal image recognition. Despite their prevalence, the exploration of federated long-tailed learning (Fed-LT) in medical images is still limited. Existing Fed-LT methods in natural imaging primarily focus on optimizing global models, neglecting crucial intra-class variations inherent in medical datasets due to diverse populations, unique findings, and varied scanning equipment. This chapter introduces the challenges of Fed-LT within medical imaging recognition. Following, it presents a self-adaptive aggregation method to effectively solve these challenges through two comprehensive medical imaging case studies.