FedPAB: Federated Medical Image Segmentation with Personalized Attention and Boundary-oriented Learning

Federated learning (FL) has attracted significant attention in the field of medical image segmentation. Due to its ability to tackle critical challenges such as insufficient training data and privacy concerns. However, data distribution varies among clients, especially for image features, making it impossible to train a single global model to fit all clients. Additionally, medical images demand precise segmentation boundaries. To address these issues, in this paper, we introduce FedPAB, a novel Federated learning framework with Personalized Attention module and Boundary-oriented Learning for medical image segmentation. Specifically, the personalized attention module alleviate feature shift problem by by selectively emphasizing important features based on the distribution of each client. We further leverage boundary-oriented learning to guide the model to delineate the boundary accurately by distinguishing the boundary and background at the representation level. Extensive experiments have been conducted on two widely used medical image segmentation tasks to demonstrate the effectiveness and superiority of our proposed FedPAB.