Answers are grounded in this paper and cite their sources.
Precise tissue segmentation in histopathology images is essential for reliable cancer diagnosis; however, existing approaches face two major challenges. First, obtaining pixel-wise annotations is costly and labor-intensive, limiting the availability of labeled training data. Second, conventional convolution-based segmentation network often struggle to accurately delineate complex tissue boundaries because they primarily focus on object interiors rather than boundary structures. We propose a semi-supervised histopathology segmentation framework that integrates pseudo-label-based self-training with Signed Distance Map (SDM)-guided boundary supervision. The proposed architecture employs a dual-head design that simultaneously predicts segmentation masks and SDMs, enabling the network to learn both semantic and geometric representations. By incorporating SDM prediction as an auxiliary objective, the framework provides complementary structural supervision that enhances boundary-aware feature learning under limited annotation settings Experiments on both a public Breast Cancer Semantic Segmentation (BCSS) dataset and a private Invasive Breast Cancer dataset demonstrate the effectiveness of the proposed framework. On the private dataset, the proposed method achieved the best performance, with a DSC of 0.922 and mIoU of 0.855. On both the BCSS and private dataset, the proposed method also improved boundary-sensitive metrics, including HD95 and ASSD, demonstrating that the SDM-guided self-training framework enables accurate boundary delineation for histopathology image segmentation. Qualitative results further demonstrate more accurate and consistent boundary delineation compared with existing segmentation architectures. The proposed SDM-guided self-training framework effectively addresses both label scarcity and boundary delineation challenges in histopathology image segmentation. By explicitly incorporating distance-based structural information, the framework improves boundary-aware representation learning and achieves robust segmentation performance across diverse breast cancer datasets.
No takes yet. Share an insight, caveat, or question.
Jang et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: