人工智能实时辅助消化内镜检出消化道隆起型病变的设备研发和临床评价  

Development and clinical evaluation of an equipment with artificial intelligence real-time assistance in detection of gastrointestinal protruding lesions under endoscopy

作　　者：黄志寅[1] 蒋竞荪 张琼英[1] 谭庆华[1] 龚慧[1] 郭林杰[1] 李传慧 杜江[1] 童欢[1] 胡兵[1] 宋捷 唐承薇[1] 李静[1] 刘苓[1] Huang Zhiyin;Jiang Jingsun;Zhang Qiongying;Tan Qinghua;Gong Hui;Guo Linjie;Li Chuanhui;Du Jiang;Tong Huan;Hu Bing;Song Jie;Tang Chengwei;Li Jing;Liu Ling(Department of Gastroenterology,Research Office of Digestive Diseases,West China Hospital,Sichuan University,Chengdu 610041,China;Department of Clinical Research Management,West China Hospital,Sichuan University,Chengdu 610041,China;Seedsmed Corporation Ltd.,Chengdu 611130,China)

机构地区：[1]四川大学华西医院消化内科,消化疾病研究室,成都610041 [2]四川大学华西医院临床研究管理部,成都610041 [3]四川希氏异构医疗科技有限公司,成都611130

出　　处：《中华消化杂志》2020年第11期745-750,共6页Chinese Journal of Digestion

基　　金：成都市科学技术局产研院技术创新项目(2017-CY02-00023-GX);四川省科学技术厅重点研发项目(2018GZ0088)。

摘　　要：目的研发消化道隆起型病变人工智能实时辅助消化内镜影像诊断设备(简称内镜人工智能设备),并评价其性能和安全性。方法收集2017年1至12月于四川大学华西医院内镜中心常规行胃镜和肠镜检查患者的内镜图像,基于深度卷积神经网络建立模型,研发内镜人工智能设备。2019年6至12月采用前瞻性、单中心、盲法、平行对照研究设计,比较内镜医师和内镜人工智能设备同时评估同一例患者胃镜和肠镜下隆起型病变的差异性,评估内镜下病变大小(病变长径<5 mm和≥5 mm)对内镜人工智能设备检出的影响。主要评价指标为内镜医师和内镜人工智能设备报告隆起型病变的时间差值,次要评价指标为内镜人工智能设备报告隆起型病变的准确度。采用Wilcoxon秩和检验和卡方检验进行统计学分析。结果共71582张白光内镜图像用于内镜人工智能设备训练,其中隆起型病变图像41376张,内镜人工智能设备研制成功,已获中华人民共和国医疗器械注册证(川械注准20202060049)。内镜人工智能设备判断隆起型病变的准确度为96.4%,灵敏度为95.1%,特异度为92.8%。内镜人工智能设备每个胃镜下隆起型病变的检出时间比内镜医师快1.524 s,但每个肠镜下隆起型病变的检出时间比内镜医师慢0.070 s,差异均有统计学意义(Z=-5.505、-4.394,P均<0.01),内镜人工智能设备每个胃镜下和肠镜下隆起型病变的检出时间均非劣效于内镜医师。内镜人工智能设备对胃镜下隆起型病变的检出率为89.9%(249/277),灵敏度为89.9%;对肠镜下隆起型病变的检出率为87.0%(450/517),灵敏度为86.9%。内镜人工智能设备对于胃镜下病变长径<5 mm和≥5 mm的隆起型病变的检出时间差值、灵敏度和漏诊率比较差异均无统计学意义(P均>0.05);内镜人工智能设备对于肠镜下病变长径≥5 mm的隆起型病变的灵敏度高于肠镜下病变长径<5 mm的隆起型病变(96.8Objective To develop an diagnostic equipment with artificial intelligence(AI)real-time assistance under endoscopy(endoscopic AI equipment)for the detection of gastrointestinal protruding lesions,and to evaluate its performance and safety.Methods From January to December 2017,at Endoscopy Center of West China Hospital,Sichuan University,the endoscopic images of individuals who underwent routine gastroscopy and colonoscopy were collected.The model was established based on convolutional neural network and the endoscopic AI equipment was developed.From June to December 2019,a prospective,single center,blinded and parallel controlled study was conducted to compare the differences in evaluation of protruding lesions of the same patient under gastroscopy or colonoscopy between endoscopist and the endoscopic AI equipment and to evaluated the impact of lesion size(lesions<5 mm and≥5 mm)on the detection of endoscopic AI equipment.The main outcome measure was the detection time difference in reporting the protruding lesion between endoscopic AI equipment and endoscopist;and the secondary indicator was the accuracy of endoscopic AI equipment in detecting the protruding lesion.Wilcoxon rank sum test and chi-square test were used for statistical analysis.Results A total of 71582 white light endoscopy images were used for endoscopic AI equipment training,which included 41376 images of protruding lesions.The endoscopic AI equipment was successfully developed and obtained the registration certificate of medical devices of the People′s Republic of China(Sichuan Instrument Standard,20202060049).The accuracy,sensitivity,and specificity of endoscopic AI equipment in detecting protruding lesions were 96.4%,95.1%and 92.8%,respectively.The detection time of each protruding lesions under gastroscopy of endoscopic AI equipment was 1.524 seconds faster than that of endoscopist;but the detection time of each protruding lesions under colonoscopy was 0.070 seconds slower than that of endoscopist,and the differences were statistically sign

关 键 词：胃肠内窥镜 人工智能 隆起型病变 卷积神经网络 检出率 

分 类 号：R656[医药卫生—外科学] TP18[医药卫生—临床医学]