LNOC-MT: Local Nuclear Occupancy Consistency Mean Teacher for Semi-supervised Nuclei Segmentation

Abstract

Accurate nuclei segmentation in histopathology images is essential for quantitative pathological analysis, but dense pixel-level annotations are costly and time-consuming. Semi-supervised learning offers a practical way to reduce annotation burden by using both labeled and unlabeled images. In this paper, we propose Local Nuclear Occupancy Consistency Mean Teacher (LNOC-MT), a local occupancy-aware Mean Teacher framework for semi-supervised binary nuclei segmentation. Unlike conventional consistency regularization that treats all pixels uniformly, LNOC-MT estimates local nuclear occupancy from Teacher predictions and uses it to guide consistency learning on unlabeled images. Specifically, an occupancy-weighted region consistency loss emphasizes nucleus-rich areas, while a local occupancy alignment loss enforces consistent regional foreground distributions between the Student and Teacher. Additionally, an auxiliary boundary branch provides contour-aware supervision. Experiments on the PanNuke dataset show that LNOC-MT outperforms several semi-supervised baselines, achieving a Dice of 84.66% and a Recall of 88.96%, demonstrating its potential for annotation-efficient nuclei segmentation.