摘要
目的 胃上皮异型增生(GED)是大部分胃癌的癌前病变,受到临床的高度重视。但胃上皮异型增生的病理诊断标准、分类和分级极不统一,病理医师之间存在高度不一致性。本研究实施机器学习,以期提高其病理诊断准确率。方法 收集武汉大学人民医院病理科1560张胃黏膜上皮病变组织HE切片(包括胃炎、不确定上皮内瘤变、低级别上皮内瘤变、高级别上皮内瘤变、肠型腺癌和弥漫型腺癌),扫描成数字切片后,自动提取有效病理区域后在×20下分割成567×567的图像块。其中1255张WSI图像用于卷积神经网络(CNN)的训练。另外305张WSI图像用于CNN的测试,采用Accuracy、Precision、Recall、F1-socre值、宏平均ROC、微平均ROC和AUC等指标对胃黏膜上皮病变多分类模型进行评价。结果 基于CNN的分类模型在测试集中的总体准确率达83.6%。测试集中的宏平均ROC曲线下面积AUC=0.97,微平均ROC曲线下面积AUC=0.96,表明本研究提出的CNN模型在胃黏膜病变分类中具有较高分类价值,具备较高的稳定性。结论 本研究建立的基于CNN多分类模型在胃黏膜上皮病变病理辅助诊断中具有较高的准确率和较好的稳定性。随着人工智能技术的在医学图像研究领域的深入发展,基于CNN病理辅助诊断模型将成为辅助病理医师进行精准诊断的重要工具。
Objective Gastric epithelial dysplasia(GED) is a precancerous lesion of most gastric cancers. However, the current pathological diagnostic criteria, classification and grading of gastric dysplasia are extremely inconsistent. The aim of this study was to improve the accuracy of pathological diagnosis of gastric epithelial lesions, by implementing machine learning to improve the accuracy of pathological diagnosis. Methods A total of 1560 H&E sections of gastric epithelial lesions(including gastritis, indeterminate intraepithelial neoplasia, low-grade intraepithelial neoplasia, high-grade intraepithelial neoplasia, diffuse adenocarcinoma, and intestinal adenocarcinoma) were collected from the pathology department of our hospital. After scanning into digital slices, the pathological area was automatically extracted and segmented into 567×567 image blocks under 20× magnifying. Among them, 1255 WSI were used for the training of convolutional neural network(CNN). An additional 305 WSI was used to evaluate the classification results of the CNN compared with the original classification labels. The multi-classification models of gastric epithelial lesions were evaluated by indicators such as accuracy, precision, recall, F1-socre value, macro-average ROC, micro-average ROC, and AUC. Results The overall accuracy of CNN in the test set reached 83.6%. The area under the macro-average ROC curve in the test set was AUC = 0.97, and the area under the micro-average ROC curve was 0.96, indicating that the CNN-based ResNet-50 algorithm model proposed in this study could be used in the classification of gastric lesions. Conclusion The model based on deep learning CNN established in this study has high accuracy and stability in the pathological auxiliary diagnosis of GIN. With in-depth development of artificial intelligence technology in the field of medical image research, the CNN-based pathological auxiliary diagnosis model based on deep learning will become an important tool to assist pathologists in accurate diagnosis.
作者
严丹丹
尹修恒
阎红琳
饶洁
罗斌
袁静萍
YAN Dan-dan;YIN Xiu-heng;YAN Hong-lin;RAO Jie;LUO Bin;YUAN Jing-ping(Department of Pathology,Renmin Hospital of Wuhan University,Wuhan 430060,China;School of Mathematics and Statistics,Wuhan University,Wuhan 430072,China)
出处
《诊断病理学杂志》
2022年第12期1097-1100,1110,共5页
Chinese Journal of Diagnostic Pathology
基金
湖北省卫生健康委员会科研项目立项(WJ2021M151)
国家自然科学基金(82003490)。
关键词
胃黏膜病变
异型增生
人工智能
深度学习
病理诊断
Gastric lesions
Dysplasia
Artificial intelligence
Deep learning
Pathological diagnosis
