新型人工智能辅助结肠镜诊断复合模型对结肠常见病变诊断价值的临床研究  

Clinical study on the diagnostic value of new artificial intelligence assisted diagnosis composite model of colonoscopy for common colonic lesions

作　　者：王晓旭 温静[2] 冯佳[3] 卢娜利 刘翠华[2] 智佳 王子阳 黄锦 WANG Xiaoxu;WEN Jing;FENG Jia;LU Nali;LIU Cuihua;ZHI Jia;WANG Ziyang;HUANG Jin(Graduate School,Xinxiang Medical University,Xinxiang 453003;Department of Gastroenterology,the 984th Hospital of PLA Joint Logistics Support Force;Department of Gastroenterology,the 980th Hospital of PLA Joint Logistics Support Force;Department of Gastroenterology,the 988th Hospital of PLA Joint Logistics Support Force,China)

机构地区：[1]新乡医学院研究生院,河南新乡453003 [2]中国人民解放军联勤保障部队第九八四医院消化内科 [3]中国人民解放军联勤保障部队第九八零医院消化内科 [4]中国人民解放军联勤保障部队第九八八医院消化内科

出　　处：《胃肠病学和肝病学杂志》2023年第11期1224-1228,共5页Chinese Journal of Gastroenterology and Hepatology

摘　　要：目的研究新型人工智能辅助结肠镜诊断复合模型对常见结肠病变的诊断价值。方法前瞻性地收集大量高质量结肠镜内镜下检测的完整视频,作为分析验证数据,以模拟临床实时使用人工智能辅助结肠镜诊断复合模型行结肠镜检查的环境,评价该复合模型在结肠常见病变中的诊断效能。结果将模型测试结果与病理结果对比,该模型对病变整体诊断的准确度、特异度、灵敏度分别为95.5%、90.6%和96.8%。在不同病变类型中,对腺瘤性息肉诊断的准确度、敏感度、特异度分别为95.9%、90.4%和98.4%;非腺瘤性息肉诊断的准确度、敏感度和特异度分别为97.9%、90.5%、99.4%;对结直肠癌的准确度、敏感度、特异度分别为94.0%、94.4%、92.6%;对结肠憩室诊断的准确度、敏感度、特异度分别为97.2%、91.8%、98.7%;对溃疡性结肠炎的准确度、敏感度和特异度分别为93.7%、87.7%和94.5%;对结直肠黏膜下肿物诊断的准确性相对较低为91.8%,敏感度和特异度分别为84.0%、93.4%。结论本团队研发的新型人工智能辅助结肠镜诊断复合模型可准确快速识别包括腺瘤性息肉、非腺瘤性息肉、结直肠癌、结肠憩室、黏膜下肿物和溃疡性结肠炎在内的结肠常见病变,减少医师间诊断差异,特别是能指导初学者进行结肠镜检查,提高结肠镜检查质量。Objective To investigate the diagnostic value of new artificial intelligence diagnosis composite model for common colonic lesions.Methods A large number of high-quality complete videos of colonoscopy were prospectively collected as validation datasets to simulate the environment of real-time colonoscopy using artificial intelligence assisted colonoscopy diagnostic composite model,and to evaluate the diagnostic efficacy of this model in common colon lesions.Results By comparing the model test results with pathological results,the accuracy,specificity and sensitivity of the model for diagnosis of the overall lesions were 95.5%,90.6%and 96.8%,respectively.Among different lesion types,the accuracy,sensitivity and specificity of diagnosis of adenomatous polyps were 95.9%,90.4%and 98.4%,respectively;while those of non-adenomatous polyps were 97.9%,90.5%and 99.4%,respectively.The accuracy,sensitivity and specificity for colorectal cancer were 94.0%,94.4%and 92.6%,respectively.The accuracy,sensitivity and specificity for the diagnosis of colonic diverticulum were 97.2%,91.8%and 98.7%,respectively,and the accuracy,sensitivity and specificity for ulcerative colitis were 93.7%,87.7%and 94.5%,respectively.The diagnostic accuracy of colorectal submucosal mass was relatively low(91.8%),and the sensitivity and specificity were 84.0%and 93.4%,respectively.Conclusion The new artificial intelligence assisted diagnosis composite model of colonoscopy developed by our team can accurately and rapidly identify common colon lesions including adenomatous polyps,non-adenomatous polyps,colorectal cancer,colonic diverticulum,submucosal mass and ulcerative colitis.It can reduce diagnostic differences among endoscopist,especially guide beginners to carry out colonoscopy and improve the quality of colonoscopy.

关 键 词：结肠镜检查 结肠病变 人工智能 辅助诊断 

分 类 号：R574[医药卫生—消化系统]