Data-efficient federated semi-supervised learning framework via pseudo supervision refinement strategy for lung tumor segmentation

Accurate lung tumor segmentation is crucial for early clinical diagnosis and subsequent treatment planning. While federated learning addresses privacy concerns in centralized model training, leveraging unlabeled data effectively remains a challenge. In this work, we introduce a novel federated semi-supervised learning framework designed for realistic clinical scenarios, efficiently utilizing limited labeled data and heterogeneous unlabeled data from multiple sites. Specifically, we initially conduct federated self-supervised pre-training at completely unlabeled sites using masked image modeling strategy. Subsequently, the downstream segmentation models are fine-tuned with a pseudo supervision refinement strategy to reduce noise in pseudo labels and stabilize training. Moreover, we propose a dynamic model aggregation strategy to assist server in dynamically combining local models during each communication round to enhance the robustness and generalizability of the global model. Extensive experiments validate our framework's superiority over recent state-of-the-art semi- supervised methods in scenarios with limited labeled data. This approach highlights the significant potential of multi-source and heterogeneous unlabeled data in facilitating the training of federated deep learning models and promotes clinical applications, assists radiologists in diagnosis and improves patient outcomes. The source code will be released on (https://github.com/zhenweishi/FedPSR).