Abstract: Federated learning (FL) enables distributed medical institutions to learn a shared model collaboratively with privacy protection. However, the model learned by federated learning still suffers from performance drop if it is deployed to an unseen domain out of the federated institutions. In this article, the authors focus on the federated domain generalization task on medical image segmentation, which aims to learn a model by federated learning across multiple-source domains such that it can generalize to unseen target domains directly.Most previous federated domain generalization methods adopt cross-domain data augmentation to improve generalization. However, they either suffer from the potential risk of privacy leakage when sharing individual samples’ information, or they lack flexibility for target images to leverage source domains’ features adaptively during inference.To address the above two problems, this article proposes FedSA, a Federated domain generalization approach via amplitude sampling and domain attention.