基于窄带成像和放大内镜的诊断模型在分类结直肠腺瘤中的价值  

The value of a diagnostic model based on narrow band imaging and magnifying endoscopy in classifying colorectal adenoma grades

作　　者：王瑞刚[1] 蒋绚 任渝棠 魏来 姜泊 WANG Rui-gang;JIANG Xuan;REN Yu-tang;WEI Lai;JIANG Bo(Department of Gastroenterology,Beijing Tsinghua Changgung Hospital,School of Clinical Medicine,Tsinghua University;Hepatobiliary and Pancreatic Center,Beijing Tsinghua Changgung Hospital,School of Clinical Medicine,Tsinghua University)

机构地区：[1]清华大学附属北京清华长庚医院消化内科,清华大学临床医学院 [2]清华大学附属北京清华长庚医院肝胆胰中心,清华大学临床医学院

出　　处：《现代消化及介入诊疗》2023年第5期554-558,563,共6页Modern Interventional Diagnosis and Treatment in Gastroenterology

基　　金：北京清华长庚医院青年启动基金资助项目(12021C1011)。

摘　　要：目的结直肠腺瘤是一种常见的胃肠道疾病,结肠镜检查被认为是检出结直肠腺瘤的主要工具。本研究拟建立一套人工智能(AI)辅助的图像诊断系统,该系统具有预测腺瘤上皮内瘤变等级的功能。借助窄带成像(NBI)或放大内镜结合NBI(ME-NBI)的帮助,该系统最终将有利于提高结肠镜检查中对腺瘤诊断的准确性,提升工作效率。方法收集2015.10-2021.10期间,北京清华长庚医院消化内镜中心的结直肠腺瘤病例,筛选出具有代表性的763名患者,组成数据集,其中包含1049张标记的图像用于训练模型。改进和训练深度卷积神经网络,对结肠镜检查中的腺瘤进行分类。在临床验证阶段,使用一个独立的数据集,包括100张图像来测试AI诊断系统的能力。将研究对象分为五组:AI诊断系统、非内镜专家、内镜专家、非内镜专家+AI、内镜专家+AI。这五组需要逐一识别所有的NBI和ME-NBI图像,并将结论与组织病理学诊断进行比较。所有数据输入统计软件SPSS 20中进行分析,P<0.05为有统计学差异。结果在标记图像上测试时,AI诊断系统对腺瘤的分类能力已经得到证实,曲线下面积(ROC)为0.895,平均精度(mAP)为95.90%。在临床测试实验中:AI的准确性为94.00%(94/100),敏感性和特异性为94.74%(36/38)和93.55%(58/62),阳性和阴性预测值(PPV和NPV)为90.00%(36/40)和96.67%(58/60)。AI诊断系统和内镜专家组的能力是相当的,而且速度更快。非内镜专家医生组使用AI诊断系统的敏感性和特异性为(93.69±5.20)%和(93.92±3.97)%,PPV和NPV为(90.31±6.74)%和(93.54±9.74%)%。与不使用AI诊断系统相比,其诊断准确性明显提高。结论AI诊断系统在使用NBI和ME-NBI诊断结直肠上皮内瘤的级等级方面有出色的能力。更值得注意的是,AI诊断系统可以指导结肠镜初学者判断结直肠腺瘤的性质,提高其诊断准确性。Objective Colorectal adenoma is a common gastrointestinal disease and its primary tool for detection is colonoscopy.The aim of this study is to establish an AI-assisted image diagnostic system which has the function of predicting adenomas’high or low grade intraepithelial neoplasia(HGIN or LGIN).This function is realized with the help of narrow band imaging(NBI)or magnifying endoscopy combined with NBI(ME-NBI),and will ultimately benefit the improvement of detection accuracy and productivity during colonoscopy examination.Methods Using a diverse and representative dataset,comes from 763 patients earlier,which contains 1049 hand labeled images,we design and train deep convolutional neural networks,AI diagnostic system,to classify adenomas after colonoscopy.In the clinical validation phase,we use an independent dataset that including a total of 100 images to test the capability of the AI system.We divide the research objects into five groups:AI,non-expert endoscopists,export endoscopists,non-expert endoscopists with AI and expert endoscopists with AI.These five groups need to identify all the NBI and ME-NBI images one by one,and compare the conclusions with histopathological diagnosis.All data are entered into SPSS 20 for analysis,and P<0.05 was defined as statistically significant.Results When test on labeled images,the capability of our AI system to classify adenomas has been demonstrated,with an area under the receiver operating characteristic curve(ROC)of 0.895 and a mean average precision(mAP)of 95.90%.In clinical test experiments:the sensitivity and specificity of the AI is 94.74%(36/38)and 93.55%(58/62),the positive and negative predictive value(PPV and NPV)is 90.00%(36/40)and 96.67%(58/60).The capability of the AI diagnostic system and the expert endoscopists group are comparable and faster.The sensitivity and specificity of non-expert endoscopists with AI is(93.69±5.20)%and(93.92±3.97)%,the PPV and NPV is(90.31±6.74)%and(93.54±9.74)%.The diagnostic efficacy of non-expert endoscopists with AI group is

关 键 词：结直肠腺瘤 上皮内瘤变 AI辅助诊断 

分 类 号：R735.3[医药卫生—肿瘤] R57[医药卫生—临床医学]