Self-distillation double student network for semi-supervised medical image segmentation

Semi-supervised learning (SSL) has garnered considerable attention in medical image segmentation due to its ability to leverage abundant unlabeled data, thereby significantly alleviating the dependency on expensive pixel-wise annotations. However, existing Mean-Teacher-based frameworks typically guide the student network using a teacher model in a unidirectional or loosely interactive manner, which fails to fully exploit the complementary relationships between different networks and often struggles with low-confidence predictions that degrade boundary accuracy. To address these limitations, we propose SDBS-Net, a novel semi-supervised dual-student self-distillation network for 3D medical image segmentation. The architecture comprises a shared encoder and two parallel decoders, forming two networks (Model I and Model II) that process both labeled and unlabeled volumes simultaneously. To enable effective bidirectional knowledge transfer, we introduce two synergistic training schemas: a prior knowledge learning (PKL) schema with hierarchical difference (HD) modules and a Self-Distillation Learning (SDL) schema. The PKL schema extracts and inject multi-scale discrepancy features from Model I into Model II, enhancing discriminative representation and boundary delineation, while the SDL schema employs multi-level soft-label and feature distillation losses to correct confirmation bias and refine overlooked regions in a mutually reinforcing manner. Extensive experiments on two challenging benchmarks such as the Left Atrium (LA) dataset and the Pancreas-CT dataset demonstrate the superiority of SDBS-Net. With only 10% and 20% labeled data, our method achieves Dice scores of 88.19% and 90.51% on the LA dataset, and 71.39% and 79.37% on the Pancreas-CT dataset, respectively, consistently outperforming state-of-the-art semi-supervised approaches while approaching fully supervised performance.