摘要
目的:分析胃癌的临床病理特征,评估基于机器学习的单病种病理切片标注技术在胃癌病理图像初筛系统中的临床应用价值。方法:收集2023年3月至2024年3月复旦大学附属中山医院病理科的600例胃肿瘤苏木精-伊红(HE)染色法染色切片并扫描成数字切片,按照2:1的比例分为400例训练集和200例内部验证集。训练集由高级职称病理医师进行像素级标注,并进行病理分型,基于R-CNN建立胃癌区域检测与分类模型,结合ResNet-50-DeepLabv3+与ce_jaccard_loss实现病灶区域精确分割。将胃肿瘤组织切片图像癌区分类模型的参数作为初始值,并经过深度迁移学习算法,以Adam作为优化器对胃癌区分类模型进行二次训练优化。再利用内部测试集和复旦大学附属闵行医院病理科的300例胃肿瘤HE切片图像作为外部测试集,对已建立的辅助诊断模型进行性能评估。结果:内部验证集中,单标注从零学习模型的准确率及AUC分别为0.836、0.917,双标注分别为0.867、0.932;迁移学习模型中,单标注准确率及AUC分别为0.922、0.942,双标注为0.926、0.953。外部验证集中,单标注从零学习模型准确率及AUC分别为0.865、0.916,双标注准确率及AUC分别为0.873、0.924;迁移学习模型中,单标注准确率及AUC分别为0.921、0.925,双标注准确率及AUC分别为0.933、0.958。结论:双标注模型能够精准识别并标注出病理图像中的胃癌区域,灵敏度和特异度较好,有助于在实际应用中辅助诊断。
Objective To analyze the clinicopathological features of gastric cancer and evaluate the clinical application value of machine learning-based single-disease pathological slice annotation technology in the primary screening system for gastric cancer pathological images.Methods A total of 600 hematoxylin-eosin(HE)stained slices of gastric tumors from the Department of Pathology,Zhongshan Hospital Affiliated to Fudan University,collected from March 2023 to March 2024,were scanned into digital slices.They were divided into a training set(400 cases)and an internal validation set(200 cases)at a ratio of 2:1.The training set was subjected to pixel-level annotation by pathologists with senior professional titles,and pathological typing was performed.A model for detection and classification of gastric cancer regions was established based on R-CNN,and precise segmentation of lesion regions was achieved by combining ResNet-50-DeepLabv3+with ce_jaccard_loss.The parameters of the cancer region classification model for gastric tumor tissue slice images were used as initial values,and through the deep transfer learning algorithm,the Adam optimizer was used for secondary training and optimization of the gastric cancer region classification model.Then,the internal test set and 300 HE-stained slice images of gastric tumors from the Department of Pathology,Minhang Hospital Affiliated to Fudan University were used as the external test set to evaluate the performance of the established auxiliary diagnosis model.Results In the internal validation set,the accuracy and AUC of the singleannotation zero-shot learning model were 0.836 and 0.917,respectively,while those of the double-annotation model were 0.867 and 0.932,respectively;in the transfer learning model,the accuracy and AUC of the single-annotation model were 0.922 and 0.942,respectively,and those of the double-annotation model were 0.926 and 0.953,respectively.In the external validation set,the accuracy and AUC of the single-annotation zero-shot learning model were 0.865 and 0.916,respectively,and those of the double-annotation model were 0.873 and 0.924,respectively;in the transfer learning model,the accuracy and AUC of the single-annotation model were 0.921 and 0.925,respectively,and those of the double-annotation model were 0.933 and 0.958,respectively.Conclusion The double-annotation model can accurately identify and annotate gastric cancer regions in pathological images with good sensitivity and specificity,which is helpful for auxiliary diagnosis in practical applications.
作者
盛霞
徐晨
朱剑
袁伟
蒋雪兵
熊海云
王书浩
朱虹光
SHENG Xia;XU Chen;ZHU Jian;YUAN Wei;JIANG Xuebing;XIONG Haiyun;WANG Shuhao;ZHU Hongguang(Minhang Hospital Affiliated to Fudan University,Shanghai 201199,China;Zhongshan Hospital Affiliated to Fudan University;Shanghai Lanwei Medical Laboratory Co.,Ltd.;Changsha LangJia Software Co.,Ltd.;Beijing Touche Future Technology Co.,Ltd.)
出处
《中国数字医学》
2025年第8期76-84,共9页
China Digital Medicine
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)
上海市闵行区科技学术委员会项目(2024MH2065)。
关键词
胃癌
人工智能
切片标注
病理诊断
Gastric cancer
Artificial intelligence
Slice annotation
Pathological diagnosis
