A Federated Semi-Supervised Learning Approach for COVID-19 CT Image Segmentation.

COVID-19 CT image segmentation is of great significance in medical image analysis, but it faces numerous challenges, especially in the context of inter-institution collaboration and data privacy protection. In particular, with the scarcity of labeled data, the effective use of limited labeled data and abundant unlabeled data for training remains a pressing issue. To address this challenge, we propose an innovative Federated Semi-Supervised Learning method (FSSLCIS) for COVID-19 CT image segmentation. We apply the MixMatch method for local semi-supervised learning, leveraging data augmentation to generate low-entropy pseudo-labels. Additionally, we use the MixUp strategy to combine labeled and unlabeled data, enhancing the model's overall learning capacity. To address the imbalance in data distribution between clients, we employ a weighted aggregation strategy on the server side to ensure the stability of the global model. Experimental results demonstrate that the proposed method performs exceptionally well on three real-world datasets, significantly improving COVID-19 CT image segmentation accuracy while ensuring data privacy.