目的:旨在系统评估基于深度学习的智能辅助内镜诊断模型(Intelligence-assisted endoscopic diagnosis model based on deep learning,DL-IEDM)对萎缩性胃炎和肠上皮化生的诊断效果.方法:系统检索PubMed、Embase、Web of Science、Cochrane Library、CNKI、维普及万方等中英文数据库中有关DL-IEDM诊断萎缩性胃炎和肠上皮化生的研究.纳入的研究根据诊断准确性试验质量评价工具-2进行质量评价,通过Rev Man 5.4、Meta-Disc1.4和Stata17.0软件计算合并后的敏感性、特异性、阳性似然比、阴性似然比、诊断优势比等诊断效能评价指标,并和内镜医师的诊断性能进行比较.结果:纳入的 13 项研究中,总共图片14 447张,其中萎缩性胃炎的图片为6 985张,肠上皮化生的图片为1 073张.Meta分析后DL-IEDM诊断癌前疾病萎缩性胃炎和肠上皮化生的灵敏度、特异度、阳性似然比、阴性似然比、诊断优势比分别为 0.93(95%CI,0.91～0.94)、0.90(95%CI,0.86～0.93)、9.2(95%CI,6.5～13.2)、0.08(95%CI,0.07～0.10)、111(95%CI,71～174),综合受试者工作特征曲线的曲线下面积(Area under the curve,AUC)为 0.96(95%CI,0.94～0.97).经过亚组和回归分析发现(1)DL-IEDM在内镜中识别萎缩性胃炎和肠上皮化生的AUC值分别为0.96(95%CI,0.94～0.97)、0.95(95%CI,0.93～0.97),两者无显著差异性;(2)使用白光内镜图像构建的DL-IEDM在识别萎缩性胃炎和肠上皮化生的性能优于使用图像增强内镜构建的DL-IEDM,差异具有统计学意义(χ2=32.53,P<0.05);3)内镜视频参与模型的训练或验证可提高DL-IEDM的诊断性能(χ2=7.47,P<0.05).最后,将DL-IEDM的AUC值与内镜专家和内镜初学者的AUC值进行相互比较发现,DL-IEDM的诊断效能显著高于内镜专家和内镜初学者(AUC=0.96,AUC=0.91,AUC=0.78),差异有统计学意义(Z=3.361,P<0.001和Z=9.265,P<0.000 1).结论:DL-IEDM对癌前疾病萎缩性胃炎和肠上皮化生具有较高的诊断准确性.除此之外,DL-IEDM可以显著提高内镜医师的诊疗水平,尤其是内镜初学者.然而,为了深度学习技术能够在临床得到更为全面的推广应用,尚需更多大样本、多中心、前瞻性的研究证实.
The performance of diagnosis deep learning for atrophic gastritis and intestinal metaplasia based on endoscope image:A systematic review and meta-analysis
Objective:To systematically evaluate the effectiveness of an intelligence-assisted endoscopic diagnosis model based on deep learning(DL-IEDM)in identifying atrophic gastritis and intestinal metaplasia.Methods:Studies that investigated the diagnosis of atrophic gastritis and intestinal metaplasia using DL-IEDM were collected from English databases(PubMed,Em-base,Web of Science,and Cochrane Library)and Chinese databases(CNKI、VIP and Wanfang).The quality of these studies was evaluated using the Quality Assessment of Diagnostic Accuracy Studies-2(QUADAS-2)tool.Sensitivity,specificity,posi-tive likelihood ratio,negative likelihood ratio,diagnostic odds ratio,and other evaluation metrics were calculated using Rev Man 5.4、Meta-Disc1.4 and Stata17.0 software.The results were also compared to the performance of endoscopists.Results:A total of 13 studies were selected,comprising a total of 14 447 images,including 6 985 images of atrophic gastritis and 1 073 images of in-testinal metaplasia.After conducting a meta-analysis,it was found that DL-IEDM has a sensitivity of 0.93(95%CI,0.91-0.94),specificity of 0.90(95%CI,0.86-0.93),positive likelihood ratio of 9.2(95%CI,6.5-13.2),negative likelihood ratio of 0.08(95%CI,0.07-0.10),and diagnostic odds ratio of 111(95%CI,71-174)for diagnosing precancerous diseases GA and IM.Addi-tionally,a Summary Receiver Operating Characteristic Curve was plotted,and the Area Under the Curve(AUC)was calculated to be 0.96(95%CI,0.94-0.97).Subgroup and regression analysis showed that:(1)there was no distinct difference in the AUC value between DL-IEDM diagnose atrophic gastritis and intestinal metaplasia,with values of 0.96(95%CI,0.94-0.97)and 0.95(95%CI,0.93-0.97),respectively;(2)the performance of DL-IEDM based on white light endoscope images was superior to that relying on image-enhance endoscope in identifying atrophic gastritis and intestinal metaplasia,and the difference was statistical-ly significant(χ2=32.53,P<0.05);(3)Video participation in model training or validation can improve the diagnostic perfor-mance of DL-IEDM(χ2=7.47,P<0.05).Finally,compared with expert and novice endoscopists,the AUC value of DL-IEDM was outstanding(AUC=0.96,AUC=0.91,AUC=0.78,respectively),and the difference was statistically significant(Z=3.361,P<0.001 and Z=9.265,P<0.000 1).Conlusion:DL-IEDM shows high diagnostic accuracy for precancerous diseases such as atrophic gastritis and intestinal metaplasia.Furthermore,it improves the diagnostic and treatment levels of endoscopists,especially those who are inexperienced.However,to promote the application of deep learning technology in clinical practice,more large sample,multicent and prospective studies are needed.
Deep learningAtrophic gastritisIntestinal metaplasiaEndoscopeMeta-analysis
万方数据
维普
