FedHNR: Federated hierarchical resilient learning for echocardiogram segmentation with annotation noise

Echocardiogram segmentation based on federated learning plays a critical role in enhancing diagnostic accuracy and efficiency. However, challenges such as inter-client annotation noise, client heterogeneity, and limited expert annotations hinder the echocardiogram segmentation based on federated learning. To address these challenges, we propose FedHNR, a federated hierarchical noise-resilient method that identifies and leverages annotation noise across global and local hierarchies. At the global-hierarchy, expert samples fine-tune the global model through a novel weight noise decoupling approach, reducing overfitting while preserving aggregated client knowledge. At the local-hierarchy, FedHNR employs region-level noise assessment and sample-level noise calibration to refine annotations using pseudo-clean labels derived from the global model. These hierarchies together mitigate the negativeness of noise and enhance the model robustness to noise. Extensive experiments on 95,469 echocardiogram frames across public and private datasets demonstrate that FedHNR outperforms ten state-of-the-art methods, showcasing its robustness in both traditional federated learning and real-world scenarios.