Signed Distance Maps Sharpen Tissue Boundary Segmentation in Breast Cancer Slides

Precise segmentation of tissue in histopathology images underpins reliable cancer diagnosis, yet teaching a network to draw those boundaries usually demands large volumes of pixel-by-pixel annotation from expert pathologists. A new study published in BMC Medical Imaging presents S3DM, a semi-supervised framework that combines pseudo-label-based self-training with signed distance map guidance to ease both problems at once. In experiments on two breast cancer histopathology datasets, the framework achieved the study's best reported performance on a private Invasive Breast Cancer dataset, with a Dice similarity coefficient of 0.922 and a mean intersection-over-union of 0.855, while also improving boundary-sensitive metrics on both datasets tested.

The method attacks two known weaknesses of conventional convolution-based segmentation networks. First, labeled training data are scarce because pixel-wise annotation is costly and labor-intensive. Second, such networks tend to focus on the interiors of tissue regions rather than on the boundary structures themselves, producing segmentations that waver exactly where diagnostic accuracy matters most.

S3DM's architecture uses a dual-head design: one head predicts the segmentation mask, while a second simultaneously predicts a signed distance map, an auxiliary representation encoding geometric distance information relative to tissue boundaries. By supervising both outputs, the framework forces the network to learn semantic and geometric representations together, giving boundary-aware features a structural signal to anchor to. The semi-supervised component relies on pseudo-label-based self-training, in which the model generates approximate labels for unlabeled images, and the signed distance map objective provides complementary structural supervision that sharpens boundary learning under limited annotation settings.

The authors describe this as a way to address label scarcity and boundary delineation challenges within a single training framework. Evaluation covered the public Breast Cancer Semantic Segmentation (BCSS) dataset and the private Invasive Breast Cancer dataset. Beyond standard overlap metrics, the framework improved boundary-sensitive measures on both datasets, including the 95th percentile Hausdorff distance (HD95) and the average symmetric surface distance (ASSD).

Qualitative results echoed the numbers: the authors report more accurate and consistent boundary delineation compared with existing segmentation architectures. The findings suggest that distance-based structural supervision can serve as an efficient complement to scarce human annotations in digital pathology. Because the experiments were confined to breast cancer histopathology data, broader applicability to other tissues and staining protocols remains to be demonstrated in future work.

For laboratories facing annotation bottlenecks, the study's central contribution is architectural rather than clinical: signed distance map prediction, used as an auxiliary self-training target, appears to strengthen boundary-aware representation learning across both a public benchmark and a private clinical dataset. The work is presented as an author accepted manuscript posted early for faster access, meaning it is citable but subject to further edits before the final version of record.

Subject of Research: Semi-supervised signed distance map guidance for boundary-aware tissue segmentation in breast cancer histopathology images

Article Title: S3DM: semi-supervised signed distance maps for boundary-aware tissue Segmentation

Article References: Jang, Y., Lee, E., Go, H., Park, J.-Y. N., & Lee, S. (2026). S3DM: semi-supervised signed distance maps for boundary-aware tissue Segmentation. BMC Medical Imaging. https://doi.org/10.1186/s12880-026-02967-z

Image Credits: AI Generated

DOI: 10.1186/s12880-026-02967-z

Keywords: histopathology, image segmentation, semi-supervised learning, signed distance maps, breast cancer, pseudo-labels, boundary detection, deep learning, medical imaging, S3DM, semi-supervised, signed