Advancing federated semi-supervised medical image segmentation: A duo of interactive denoising pseudo-labels and convolutional contrastive learning.

Many existing studies on federated learning (FL) for segmentation primarily assume that all client data are labeled. However, in reality, due to the high cost of hospital construction and the scarcity of expert annotators, many medical sites can only provide unlabeled data. Therefore, in our work, we focus on a more practical and challenging problem, namely federated semi-supervised segmentation (FSSS), where only a subset of clients possesses labeled data while the remaining clients contribute unlabeled data. To tackle this problem, we propose an effective and generalizable FSSS framework. Specifically, labeled clients are first aggregated to construct a label-based aggregation model, which serves to guide the pseudo-label generation for unlabeled clients. Since the generated initial pseudo-labels often suffer from feature offset, we develop a pixel-level denoising method based on uncertainty feature map estimation, which enhances the quality of pseudo-labels by leveraging local data. Second, we design a model-convolutional contrastive learning to endow unlabeled clients with enhanced feature discrimination capabilities, thereby correcting their inaccurate representations. Finally, an effective dynamic model aggregation method is devised to adjust the aggregation weight of each client by considering the contribution quantified via a one-hot scheme. We comprehensively evaluate our method from multiple perspectives on three non-independent and identically distributed (Non-IID) segmentation tasks, and the experimental results confirm the effectiveness of our method. The codes of this work has been released at the following link: https://github.com/ZhenghuaXu/FedDPCon.