不同分类器对苏木精-伊红染色在早期胃癌组织病理图像分级研究  

Study on the different classifier on pathological image grade of early gastric cancer tissue with HE staining

作　　者：胡嘉影 陈雨萍 罗志飞[1] 蔡仁桑[1] 马梦晨 黄幼生[1] 陈梦柚 HU Jia-ying;CHEN Yu-ping;LUO Zhi-fei(Department of Pathology,First Affiliated Hospital of Hainan Medical University,Hainan 570102,China)

机构地区：[1]海南医学院第一附属医院病理科,海南海口570102 [2]海口市第三人民医院病理科,海南海口570000

出　　处：《中国医学装备》2020年第12期63-68,共6页China Medical Equipment

基　　金：海南省科学技术厅重点研发计划(2019JM137)“早期胃癌组织病理图像分类及精细化算法研究”。

摘　　要：目的:研究不同分类器对苏木精-伊红(HE)染色的早期胃癌组织病理图像分类效果及病理分级的价值。方法:收集医院病理科原始HE染色的420幅胃癌组织病理图像,采用逻辑回归(LR)、支持向量机(SVM)和朴素贝叶斯(NB)3个机器学习算法分类器,对图像进行基于像素分类的癌巢分割。对比最佳的分类器对黏膜内癌[cT1a(M)]和黏膜下癌[cT1b(SM)]患者在像素级别特征(PLF)、对象级别特征(OLF)及两者组合(PLF+OLF)上的分类准确率。结果:SVM分类器在癌巢分割的查准率和查全率高于LR和NB分类器;SVM分类器在(PLF+OLF)上的准确率(76.7%)高于单独PLF(71.9%)与OLF(70.8%)。cT1a(M)被正确分类的准确率为69.7%,cT1b(SM)被正确分类的准确率为82.1%。结论:在HE染色早期胃癌组织病理图像中癌巢分割中SVM分类器的查准率和查全率较高,同时其对组织学分级效果较好。HE染色的早期胃癌组织病理图像的组织学分级可以在一定程度上代表患者的组织学分级。Objective:To study the value of different classifier on classification effect and pathological grade of early gastric cancer tissue with hematoxylin-eosin(HE)staining.Methods:420 pathology images of gastric cancer tissue with original HE staining in hospital were collected.And three classifiers of machine learning algorithm included logistic regression(LR),support vector machine(SVM)and naive Bayes(NB)were adopted.The images were segmented as cancer nest based on pixel classification.The classification accuracy of the optimal classifier on pixel level features(PLF),object level features(OLF)and the combination of PLF and OLF(PLF+OLF)of patients with intra-mucosal carcinoma[cT1a(M)]and who with submucosal carcinoma[cT1b(SM)]were further compared.Results:The precision rate of inquire and entirety rate of inquire of SVM classifier in the segmentation of cancer nest were higher than those of LR classifier and NB classifier,respectively.And the accuracy rate of SVM classifier(76.7%)was higher than that of single PLF(71.9%)and OLF(70.8%),respectively.And the accuracy rate that cT1a(M)was correctly classified was 69.7%,and that of cT1b(SM)was 82.1%.Conclusion:The precision rate of inquire and entirety rate of inquire of SVM classifier are higher in the segmentation of cancer nest of pathological images of early gastric cancer tissue with HE staining,and it has better effect on histological grade.And the histological grade of pathological images of early gastric cancer tissue with HE staining can represent histological grade of patients to a certain degree.

关 键 词：早期胃癌 癌巢分割 细胞核分割 组织学分级 机器学习算法 分类器 

分 类 号：R735.2[医药卫生—肿瘤]