深度学习技术辅助诊断结直肠息肉的临床分析  

Clinical Analysis of Deep Learning Technology in Assisting Diagnosis of Colorectal Polyps

作　　者：姜良慧 孙昕[2] 张荣秋 孟欣颖[1] 李雪桐 周长宏[1] JIANG Lianghui;SUN Xin;ZHANG Rongqiu;MENG Xinying;LI Xuetong;ZHOU Changhong(Department of Health Care,Qingdao Municipal HospitalEastAffiliated to Qingdao University,Qingdao,Shandong Province,266071;Department of Endoscopy Center,Qingdao Municipal HospitalEastAffiliated to Qingdao University,Qingdao,Shandong Province,266071;Sino-German Joint Software Institute,Beihang University,Beijing)

机构地区：[1]青岛大学附属青岛市市立医院东院保健科,266071 [2]青岛大学附属青岛市市立医院东院内镜中心 [3]北京航空航天大学中德软件技术联合研究所

出　　处：《胃肠病学》2020年第7期389-394,共6页Chinese Journal of Gastroenterology

摘　　要：背景:基于深度学习技术的计算机辅助诊断已成为胃肠病学领域的研究热点,计算机辅助诊断结直肠息肉已引起越来越多的关注。目的:验证一个自动识别结直肠息肉的深度学习模型,分析该模型对新手内镜医师的辅助学习功能。方法:回顾性收集2019年1月—2020年1月青岛市市立医院东院内镜中心数据库中的结肠镜图像共1200张,其中结直肠息肉图像600张,正常结肠图像600张。以深度学习技术模型对1200张内镜图像进行验证,并比较该模型与5名新手内镜医师诊断结直肠息肉的敏感性、特异性、准确率、时间。结果:深度学习模型诊断结直肠息肉的敏感性为93.2%,特异性为98.7%,准确率为95.9%,每张图像的诊断时间为(0.20±0.03)s,模型的敏感性、准确率、诊断时间优于5名新手内镜医师,特异性优于部分新手内镜医师。当息肉≤5 mm或6~9 mm时,模型的准确率分别为88.1%、96.8%,优于5名新手内镜医师;当息肉≥10 mm时,模型的准确率为100%,与5名新手内镜医师无明显差异。模型识别隆起型息肉的准确率为94.8%,优于部分新手内镜医师;模型识别扁平型息肉的准确率为91.7%,优于5名新手内镜医师。扁平型息肉未能识别(38.8%)、黏膜皱襞处息肉(32.7%)、误认黏膜皱襞为息肉(12.2%)为模型假阴性或假阳性的主要原因。结论:深度学习模型对结直肠息肉的辅助诊断有较高的准确率、敏感性、特异性,且诊断时间较短,可辅助新手内镜医师识别小息肉和扁平型息肉。Background:Computer-aided diagnosis based on deep learning technology is a research hotspot in the field of gastroenterology,and computer-aided diagnosis of colorectal polyps has received more and more attention.Aims:To validate a model based on deep learning for the automatic identification of colorectal polyps,and to analyze its auxiliary learning function for helping novice endoscopists.Methods:A total of 1200 colonoscopy images(600 colorectal polyp images and 600 normal images)in the endoscopy center database of Qingdao Municipal Hospital(East)from January 2019 to January 2020 were retrospectively collected.Deep learning model was used to identify the 1200 images.The sensitivity,specificity,accuracy and diagnosis time of deep learning model and 5 novice endoscopists for diagnosis of colorectal polyps were compared.Results:The deep learning model showed a sensitivity of 93.2%,specificity of 98.7%,accuracy of 95.9%for detecting colorectal polyps,and the diagnosis time of each image was(0.20±0.03)second.The sensitivity,accuracy,and diagnosis time of the model were superior to 5 novice endoscopists,and the specificity was superior to some novice endoscopists.The accuracies of model for polyps with size≤5 mm and 6~9 mm were 88.1%and 96.8%,respectively,and were superior to 5 novice endoscopists;the accuracy of model for polyps with size≥10 mm was 100%,and was similar to 5 novice endoscopists.The accuracy of model for polyps with protrude type was 94.8%,and was superior to some novice endoscopists;the accuracy of model for polyps with flat type was 91.7%,and was superior to 5 novice endoscopists.Missing the polyps with flat type(38.8%),polyps at mucosal folds(32.7%),and mistaking the mucosal folds as polyps(12.2%)were the main causes of false negative or false positive results of the model.Conclusions:The deep learning model has a high accuracy,sensitivity,specificity and shorter diagnosis time for diagnosis of colorectal polyps,and can be used to assist novice endoscopists in diagnosing small polyps and flat p

关 键 词：深度学习 人工智能 肠息肉 诊断 

分 类 号：R73[医药卫生—肿瘤]