Boundary-precise semi-supervised medical image segmentation via prototypical mutual learning and cyclic task consistency

Consistency regularization and pseudo labeling methods within semi-supervised learning present promising solutions to address the scarcity of labeled data in medical image segmentation. Despite recent advancements achieved by integrating these techniques, existing methods overlook several critical issues, potentially limiting model performance. First, prototype bias arises when class prototypes used for pseudo label reliability assessment fail to capture the overall data distribution, thereby compromising the supervised signal. Second, current approaches often neglect the potential of leveraging complementary information through the learnable transformation between the predictions of different tasks, limiting the effectiveness of multi-task learning. To address these limitations, we propose a novel framework named PMCM, which integrates prototypical mutual learning (PML) and cyclical multi-task learning (CML). PML addresses prototype bias by introducing a unified class prototype that captures the overall distribution of both labeled and unlabeled data, enabling reliable mutual correction between networks. Complementing PML, CML facilitates effective cross-task information exchange by learning the transformation between segmentation and signed distance field through bidirectional supervision. Extensive experiments on two public datasets with CT and MRI modalities demonstrate that our method achieves state-of-the-art performance and substantially improves boundary segmentation precision, exhibiting robust effectiveness even under limited annotation scenarios. The code is available at https://github.com/xup6YJ/PMCM .