Single-positive multi-label-based model for tissue segmentation in pathological images
JIA Yumian
LIN Jiatai
LE Juncong
LIU Fangfang
LIU Zaiyi
HAN Chu
GUO Xiaojing
Abstract：Objective Weakly Supervised Semantic Segmentation(WSSS)leverages image-level labels instead of pixel-level annotations,significantly improving annotation efficiency.To further reduce the burden on pathologists'annotation,this study proposes a novel Single Positive Multi-Label Weakly Supervised Semantic Segmentation(SP-WSSS)framework,which requires only one positive label per image indicating the presence of any tissue type.Methods To mitigate the impact of extensive noise in pseudo-labels,we introduce an early-learning strategy to prevent the model from overfitting to noisy signals.The tissue segmentation performance is evaluated on two public datasets,LUAD-HistoSeg and BCSS-WSSS.Additionally,three pathologists are invited to assess the annotation efficiency.Results Experiments demonstrate that SP-WSSS achieves outstanding tissue segmentation performance on two public datasets,LUAD-HistoSeg and BCSS-WSSS.Three board-certified pathologists were invited to evaluate annotation efficiency:for a task involving 100 images,pixel-level annotation took approximately 200 minutes on average,while single-positive multi-label annotation required less than 2 minutes.Furthermore,compared to conventional annotation methods,single-positive multi-label annotation achieved higher consistency(consistency scores:pixel-wise annotation 85.64%;image-level multi-label 92.25%;single-positive multi-label 98%).Conclusion The proposed novel annotation paradigm for segmentation SP-WSSS shows strong potential to replace traditional annotation approaches,substantially alleviating the annotation burden on pathologists.
Keywords：Single-positivemulti-labelWeakly-SupervisedSemantic Segmentation(WSSS)
Publication Date：2025-12-28
Online Publishing Date：2026-01-28(First online date of this platform, not the publication date of the document)
Pages：7( 1577-1583 )