Addressing multifaceted heterogeneity: a practice of federated learning strategy fusion for medical image segmentation

The emergence of federated learning provides a new solution to the privacy issues of machine learning in medical image processing, enabling various medical institutions to train models without sharing data. However, the various heterogeneous scenarios in reality have made the practice of federated learning more complex. For this reason, researchers have explored a series of solutions based on FedAvg. This study holds that constructing a unified federated learning framework to adaptively address multiple heterogeneity issues is of great value, as it can efficiently utilize valuable distributed medical data to achieve the optimal model performance and privacy considerations. Therefore, in this study, a series of existing federated learning methods for addressing single heterogeneity were practiced in the medical image segmentation task, including federated self-supervised learning, federated distillation, and prototype alignment, and attempts were made to integrate these methods to achieve the fusion of their mechanisms. The results show that the integration of these methods does have higher potential in adapting to heterogeneous environments, but the performance cannot effectively outperform the benchmark FedAvg. Therefore, a more precisely designed scheme still needs to be proposed.