摘要
目的构建基于病理数字切片的肿瘤人工智能(AI)诊断模型,并验证其诊断效能。方法选取2022年1月—2024年1月复旦大学附属华东医院和复旦大学附属上海市第五人民医院肿瘤(肺癌、胃癌、肝癌、大肠癌、乳腺癌、前列腺癌)患者100例,收集其病理组织切片共4000张,按8∶2的比例将患者分为训练组80例和验证组20例。数据集图像输入DeepLabV3+语义分割模型中,提取出数字图像的病理特征后输入ResNet50分类模型中训练,训练组分别输入两种数据集以构建肿瘤AI诊断模型,模型1输入仅标注病理组织(阳性标注)的数据集,模型2输入标注病理组织和正常组织(阳性及阴性标注)的数据集;验证组用于验证上述两种模型的诊断效能。对比两种模型诊断肿瘤的HE染色、ROC曲线分析、诊断效能分析结果。结果HE染色显示,相比模型1,模型2能更清晰地显示肿瘤的病理组织范围与边界,更好地区分病理组织与正常组织;ROC曲线结果表明,模型1、模型2均可有效诊断肿瘤(P<0.05),且模型2诊断肿瘤的AUC、约登指数比模型1大,差异明显(P<0.05);模型2诊断肿瘤的敏感性、特异度、准确值及阳、阴性预测值比模型1大,差异明显(P<0.05)。结论基于病理数字切片构建的肿瘤AI诊断模型对肿瘤的诊断效果优良,其中阳性及阴性标注模型对肿瘤有着更高的诊断效能,值得临床尝试实践。
Objective To construct an artificial intelligence(AI)diagnostic model for tumors based on digital pathology sections and validating its diagnostic efficacy.Methods Patients with tumors(lung cancer,stomach cancer,liver cancer,colorectal cancer,breast cancer,prostate cancer)at Huadong Hospital Affiliated to Fudan University and Shanghai Fifth People's Hospital Affiliated to Fudan University were selected from January 2022 to January 2024,with a total of 100 cases,and a total of 4,000 pathological tissue sections were collected,and the patients were divided into a training group and a validation group according to the ratio of 8∶2,with 80 cases and 20 cases,respectively.The dataset images were input into DeepLabV3+semantic segmentation model,and the pathological features of the digital images were extracted and input into ResNet50 classification model for training,and two kinds of datasets were input into the training group to construct the tumor AI diagnostic model respectively,model 1 inputs the dataset with only pathological tissues(positive annotation),and model 2 inputs the dataset with pathological tissues and normal tissues(positive and negative annotations).The validation group was used to verify the diagnostic efficacy of the above two models.Comparison of HE staining,ROC curve analysis and diagnostic efficacy analysis of the two models for tumor diagnosis.Results HE staining showed that compared with model 1,model 2 could display the pathological tissue extent and boundary of the tumor more clearly,and better distinguish pathological tissue from normal tissue;the results of ROC curve showed that both model 1 and model 2 could diagnose the tumor efficiently(P<0.05),and the AUC and Jordon's index of diagnosing the tumor in model 2 were larger than that of model 1,with obvious differences(P<0.05);the Sensitivity,specificity,accuracy value and positive and negative predictive values were greater than model 1,with significant differences(P<0.05).Conclusion Tumor AI diagnostic models constructed based on pathology digital sections have excellent diagnostic effects on tumors,in which positive and negative labeling models have higher diagnostic efficacy for tumors,which is worth clinical attempts and practice.
作者
肖立
李小静
张美花
杨菊
熊海云
张晓军
刘秀萍
XIAO Li;LI Xiaojing;ZHANG Meihua;YANG Ju;XIONG Haiyun;ZHANG Xiaojun;LIU Xiuping(Department of Pathology,Huadong Hospital Affiliated to Fudan University,Shanghai 200040,China;Department of Pathology,Shanghai Fifth People's Hospital Affiliated to Fudan University,Shanghai 200240,China;Department of Pathology,Shanghai Lanwei Medical Laboratory Co.,Ltd.,Shanghai 200335,China;R&D Department,Changsha Langjia Software Co.,Ltd.,Changsha 410600,China;Thorough Laboratory,Beijing Thorough Future Technology Co.,Ltd.,Beijing 100036,China)
出处
《西部医学》
2025年第3期366-371,共6页
Medical Journal of West China
基金
2021—2022年度上海市促进产业高质量发展专项(人工智能专题)项目(2021-GZL-RGZN-01031)。
关键词
病理数字切片
肿瘤
AI
诊断模型
诊断效能
Pathology digital slices
Tumor
AI
Diagnostic model
Diagnostic efficacy
作者简介
通信作者:刘秀萍,教授,E-mail:Feiying459@163.com。
