伪标签置信度调控结直肠癌病理图像半监督语义分割  

Pseudo-label confidence regulates semi-supervised semantic segmentation of pathological images of colorectal cancer

作　　者：徐晗晗 张印辉[1] 何自芬[1] 刘珈岑 李振辉[2] 吴琳 史本杰 XU Hanhan;ZHANG Yinhui;HE Zifen;LIU Jiacen;LI Zhenhui;WU Lin;SHI Benjie(Faculty of Mechanical and Electrical Engineering,Kunming University of Science and Technology,Kunming 650500,China;Department of Radiology,Yunnan Cancer Hospital,Kunming 650106,China;Department of Pathology,Yunnan Cancer Hospital,Kunming 650106,China)

机构地区：[1]昆明理工大学机电工程学院,云南昆明650500 [2]云南省肿瘤医院放射科,云南昆明650106 [3]云南省肿瘤医院病理科,云南昆明650106

出　　处：《光学精密工程》2025年第4期591-609,共19页Optics and Precision Engineering

基　　金：国家自然科学基金(No.62061022,No.62171206);装备智能运用教育部重点实验室开放基金项目(No.AAIE-2023-0203)。

摘　　要：为了改善结直肠癌病理图像半监督语义分割任务中存在的低置信度伪标签利用不充分、高置信度伪标签准确性亟需优化和伪标签类别不平衡等问题,本文提出了一种伪标签置信度调控方法,旨在实现结直肠癌病理图像的高质量多类别半监督语义分割。首先,基于教师-学生模型的半监督语义分割框架,提出在一致性正则化中嵌入类别置信度调控,通过对未训练教师模型生成的低置信度伪标签中的混淆类别进行移除以增强确定性,从而提升低置信度伪标签的贡献率。其次,提出对训练后教师模型生成的伪标签进行先筛选后细化的操作范式,通过对筛选后的高置信度伪标签进行基于条件随机场的细化操作,以改善高置信度伪标签中边界模糊和缺乏语义信息的问题。最后,为缓解伪标签数据中的类别不平衡,设计了一种基于伪标签类别数判定的自适应随机级联强数据增强的方法。通过自建结直肠癌病理图像数据集以及公开的多类别病理图像数据集进行实验验证,本文方法实现了74.09%的结直肠癌病理图像四个类的平均分割精度,相比于基准网络提高6.43%,为结直肠癌病理图像半监督语义分割提供有力的算法支持。In order to improve the under-utilization of low-confidence pseudo-labels,the need to optimize the accuracy of high-confidence pseudo-labels and the imbalance of pseudo-label categories in the semi-supervised semantic segmentation task of colorectal cancer pathological images,this paper proposed a pseudo-label confidence regulation method to achieve high-quality multi-class semi-supervised semantic segmentation of colorectal cancer pathological images.First,based on the semi-supervised semantic segmentation framework of the teacher-student model,we propose to embed class confidence regulation in the consistency regularization,and to enhance the certainty by removing the confusing classes in the low confidence pseudo-labels generated by the untrained teacher model,so as to increase the contribution rate of the low confidence pseudo-labels.Secondly,an operation paradigm of first screening and then refining the pseudo-tags generated by the teacher model after training is proposed.By refining the filtered high-confidence pseudo-tags based on conditional random fields,the problems of boundary ambiguity and lack of semantic information in high-confidence pseudo-tags are improved.Finally,in order to alleviate the category imbalance in pseudo-label data,an adaptive random cascade strong data enhancement method based on the classification number of pseudo-label is designed.Through the experimental verification of the self-built colorectal cancer pathological image dataset and the published multi-class pathological image dataset,the proposed method achieves 74.09%average segmentation accuracy of four categories of colorectal cancer pathological images,which is 6.43%higher than that of the benchmark network,and provides powerful algorithm support for semi-supervised semantic segmentation of colorectal cancer pathological images.

关 键 词：结直肠癌病理图像 半监督语义分割 教师-学生模型 一致性正则化 条件随机场 数据增强 

分 类 号：TP394.1[自动化与计算机技术—计算机应用技术] TH691.9[自动化与计算机技术—计算机科学与技术]