深度学习神经网络在非炎性主动脉中膜变性病理图像分类中的应用  

Application of deep learning neural network in pathological image classification of non-inflammatory aortic membrane degeneration

作　　者：汪昊 陈东[1] 万涛 赵艳丽 孙中杰 方微[1] 董方 连国亮[1] 韩丽媛[1] Wang Hao;Chen Dong;Wan Tao;Zhao Yanli;Sun Zhongjie;Fang Wei;Dong Fang;Lian Guoliang;Han Liyuan(Department of Pathology,Beijing Anzhen Hospital,Capital Medical University,Beijing 100029,China;School of Biomedical Science and Medical Engineering,Beijing Advanced Innovation Center for Biomedical Engineering,Beihang University,Beijing 100191,China)

机构地区：[1]首都医科大学附属北京安贞医院病理科,100029 [2]北京航空航天大学生物与医学工程学院,生物医学工程高精尖创新中心,100191

出　　处：《中华病理学杂志》2021年第6期620-625,共6页Chinese Journal of Pathology

基　　金：国家自然科学基金(61876197);北京市自然科学基金(7192105);北京市优秀人才青年骨干个人项目(2018000021469G240);北京市医院管理局临床技术创新项目(XMLX201814)。

摘　　要：目的探讨基于深度学习的人工智能在非炎性主动脉中膜变性中的辅助诊断及其应用价值。方法选取2018年1—6月首都医科大学附属北京安贞医院保存的89例非炎性主动脉中膜变性标本组织HE切片,扫描成数字切片后进行人工标注,在标注区域总提取1627幅中膜病变HE图像。结合一种改进的基于ResNet18的卷积神经网络模型,进行非炎性主动脉病理图像的4分类研究,并对模型应用进行检测。结果4分类模型对中膜变性病理改变中最常见的平滑肌细胞核缺失病变的识别准确率、灵敏度及精确率分别为99.39%、98.36%、98.36%。弹力纤维断裂和/或缺失病变识别精确率为98.08%;层内型黏液样细胞外基质聚集病变识别准确率为96.93%。模型整体准确率为96.32%,受试者工作特征曲线下面积值可达0.982。结论初步验证了深度学习神经网络模型在非炎性主动脉病变图像分类方面的准确性,该方法可以有效提升病理医师诊断效率。Objective To investigate the value of deep learning in classifying non-inflammatory aortic membrane degeneration.Methods Eighty-nine cases of non-inflammatory aortic media degeneration diagnosed from January to June 2018 were collected at Beijing Anzhen Hospital,Capital Medical University,China and scanned into digital sections.1627 hematoxylin and eosin stained photomicrographs were extracted.Combined with the ResNet18-based deep convolution neural network model,4-category classification of pathological images were performed to diagnose the non-inflammatory aortic lesion.Results The prediction model of artificial intelligence assisted diagnosis had the best accuracy,sensitivity and precision in identifying lesions with smooth muscle cell nuclei loss,which were 99.39%,98.36%and 98.36%,respectively.The classification accuracy of elastic fiber fragmentation and/or loss lesions was 98.08%,while that of intralamellar mucoid extracellular matrix accumulation lesions was 96.93%.The overall accuracy of the classification model was 96.32%,and the area under the curve was 0.982.Conclusions The accuracy of deep learning neural network model in the 4-category classification of non-inflammatory aortic lesionsis confirmed based on digital photomicrographs.This method can effectively improve the diagnostic efficiency of pathologists.

关 键 词：主动脉疾病 血管中膜 人工智能 神经网络(计算机) 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术] TP18[自动化与计算机技术—计算机科学与技术] R543.1[医药卫生—心血管疾病]