DTSC-Net: Dual-Teacher Semi-Supervised Segmentation with Structural Consistency and Prompted SAM*

Automated medical image segmentation plays a pivotal role in clinical diagnosis and treatment planning. Despite the success of deep neural networks, their reliance on large-scale pixel-wise annotations limits applicability in real-world scenarios where expert labeling is expensive and time-consuming. Semi-supervised learning provides a promising alternative, yet current methods often suffer from inaccurate pseudo-labels and poor generalization in structurally complex regions. To overcome these limitations, we propose DTSC-Net, a dual-teacher semi-supervised segmentation framework that integrates structural consistency constraints and prompt-driven prior knowledge. Specifically, we introduce a reliability assessment mechanism that jointly evaluates mask overlap and boundary deviation to filter high-confidence pseudo-labels. Additionally, a pretrained SAM-Med2D model, guided by structure-derived point and box prompts, provides complementary anatomical cues to refine pseudo supervision. Extensive experiments on the LA dataset demonstrate that our approach outperforms state-of-the-art methods, particularly under low-label regimes, highlighting the effectiveness of integrating structural priors and confidence-aware training in semi-supervised medical image segmentation.