基于多模态内镜图像的人工智能模型诊断胃肿瘤性病变研究(含视频)  

An artificial intelligence system based on multi-modal endoscopic images for the diagnosis of gastric neoplasms(with video)

作　　者：陶逍 吴练练 杜泓柳 董泽华 于红刚[1] Tao Xiao;Wu Lianlian;Du Hongliu;Dong Zehua;Yu Honggang(Department of Gastroenterology,Renmin Hospital of Wuhan University,Hubei Key Laboratory of Digestive Diseases,Wuhan 430060,China)

机构地区：[1]武汉大学人民医院消化内科、消化系统疾病湖北省重点实验室,武汉430060

出　　处：《中华消化内镜杂志》2024年第9期690-696,共7页Chinese Journal of Digestive Endoscopy

基　　金：国家自然科学基金(82202257);武汉市人工智能示范应用场景项目(2022YYCJ01);武汉市知识创新专项-曙光计划项目(2022020801020482);武汉大学深化校企改革项目。

摘　　要：目的构建一种基于多模态内镜图像数据的人工智能胃肿瘤性病变诊断模型,并将其诊断效能与基于单模态图像数据的模型及内镜医师诊断水平进行比较。方法收集武汉大学人民医院内镜中心2018年3月至2019年12月期间463例患者的3267张白光成像和弱放大成像的胃肿瘤性病变及非肿瘤性病变图像,用于构建基于白光图像和弱放大图像的单模态模型(白光模型和弱放大模型)。将同一病变的白光和弱放大图像组合成图像对,用于构建多模态特征融合模型(多模态模型)。使用2020年3月至2021年3月间97例患者(102个病变)的696张图像作为图像测试集,比较单模态模型与多模态模型在图像水平和病变水平识别胃肿瘤性病变的诊断效能。收集2022年1月至2022年6月期间80例患者(80个病变)的视频片段,用于比较弱放大模型、多模态模型和7名内镜医师在在病变水平对胃肿瘤性病变的诊断效能。结果在图像测试集中,多模态模型在图像水平诊断胃肿瘤性病变的灵敏度、准确率分别为84.96%(576/678)、86.89%(1220/1289),优于弱放大模型的灵敏度[63.13%(113/179),χ^(2)=42.81,P<0.001]和准确率[80.59%(353/438),χ^(2)=10.33,P=0.001],同样优于白光模型的灵敏度[70.47%(74/105),χ^(2)=13.52,P<0.001]和准确率[67.82%(175/258),χ^(2)=57.27,P<0.001]。多模态模型在病变水平诊断的灵敏度、特异度和准确率分别为87.50%(28/32)、88.57%(62/70)、88.24%(90/102),特异度(χ^(2)=22.99,P<0.001)和准确率(χ^(2)=19.06,P<0.001)高于白光模型;但与弱放大模型间差异均无统计学意义(P>0.05)。在视频测试中,多模态模型在病变水平诊断胃肿瘤性病变的灵敏度、特异度和准确率分别为95.00%(19/20)、93.33%(56/60)和93.75%(75/80),分别优于内镜医师总体的77.14%(108/140)、79.29%(333/420)和78.75%(441/560),差异具有统计学意义(χ^(2)=18.62,P<0.001;χ^(2)=35.07,P<0.001;χ^(2)=53.12,P<0.001),且高于资深内镜医师�Objective To develop an artificial intelligence model based on multi-modal endoscopic images for identifying gastric neoplasms and to compare its diagnostic efficacy with traditional models and endoscopists.Methods A total of 3267 images of gastric neoplasms and non-neoplastic lesions under white light(WL)endoscopy and weak magnification(WM)endoscopy from 463 patients at the Digestive Endoscopy Center of Renmin Hospital of Wuhan University from March 2018 to December 2019 were utilized.Two single-modal models(WL model and WM model)were constructed based on WL and WM images separately.WL and WM images of corresponding lesions were combined into image pairs for creating a multi-modal(MM)characteristics integration model.A test set consisting of 696 images of 102 lesions from 97 patients from March 2020 to March 2021 was used to compare the diagnostic efficacy of the single-modal models and a multi-modal model for gastric neoplastic lesions at both the image and the lesion levels.Additionally,video clips of 80 lesions from 80 patients from January 2022 to June 2022 were employed to compare diagnostic efficacy of the WM model,the MM model and 7 endoscopists at the lesion level for gastric neoplasms.Results In the image test set,the sensitivity and accuracy of MM model were 84.96%(576/678),and 86.89%(1220/1289),respectively,for diagnosing gastric neoplasms at the image level,which were superior to 63.13%(113/179)and 80.59%(353/438)of WM model(χ^(2)=42.81,P<0.001;χ^(2)=10.33,P=0.001),and also better than those of WL model[70.47%(74/105),χ^(2)=13.52,P<0.001;67.82%(175/258),χ^(2)=57.27,P<0.001].The MM model showed a sensitivity of 87.50%(28/32),a specificity of 88.57%(62/70),and an accuracy of 88.24%(90/102)at the lesion level.The specificity(χ^(2)=22.99,P<0.001)and accuracy(χ^(2)=19.06,P<0.001)were significantly higher than those of WL model;however,there was no significant difference compared with those of the WM model(P>0.05).In the video test,the sensitivity,specificity and accuracy of the MM model at the lesio

关 键 词：人工智能 胃肿瘤性病变 弱放大成像 多模态 

分 类 号：R735.2[医药卫生—肿瘤]