基于卷积神经网络的辅助识别模型在白光内镜诊断慢性胃炎中的应用价值

Application of Convolutional Neural Network-based Auxiliary Recognition Model in Diagnosis of Chronic Gastritis by White Light Endoscopy

背景:慢性萎缩性胃炎的准确诊断对于胃癌预防具有重要意义。卷积神经网络在消化内镜领域的潜力已得到证实。目的:基于卷积神经网络建立白光内镜诊断慢性胃炎的辅助识别模型,并评估其应用价值。方法:回顾性收集武汉大学人民医院和北京大学第三医院的慢性胃炎白光内镜图片作为训练集和测试集。辅助识别模型由Unet++和Resnet-50卷积神经网络构建,包括两层结构。采用ROC曲线评估Resnet-50卷积神经网络的诊断效能;采用Kappa一致性检验评估辅助识别模型和3名内镜医师的诊断结果与标准答案的一致性。结果:训练集含5200张慢性胃炎白光内镜图片,测试集含668张慢性胃炎白光内镜图片。Unet++对慢性胃炎病变的总体命中率为97.2%。Resnet-50卷积神经网络诊断慢性萎缩性胃炎的ROC曲线下面积为0.97,准确性、敏感性和特异性分别为90.0%、90.3%和89.6%。其对慢性萎缩性胃炎的诊断结果与标准答案具有高度一致性(κ=0.763),与内镜专家(κ=0.712,κ=0.698)水平相近,明显优于普通内镜医师(κ=0.585)。结论:本研究建立的基于卷积神经网络的辅助识别模型能为白光内镜诊断慢性胃炎提供有效帮助,具有良好的临床应用前景。

Background:Accurate diagnosis of chronic atrophic gastritis is of great significance for the prevention of gastric cancer.The potential of convolutional neural network in digestive system endoscopy has been confirmed.Aims:To establish an auxiliary recognition model based on convolutional neural network in diagnosis of chronic gastritis by white light endoscopy and to evaluate its application value.Methods:White light endoscopy images of chronic gastritis patients were collected retrospectively from Renmin Hospital of Wuhan University and Peking University Third Hospital as the training set and test set.The auxiliary recognition model was constructed by Unet++and Resnet-50 convolutional neural network and consisted of two layers.The diagnostic performance of Resnet-50 convolutional neural network was evaluated by ROC curve;the consistency between the results of auxiliary recognition model,three endoscopists,and standard answers were assessed by Kappa coefficients.Results:A total of 5200 white light endoscopic images of chronic gastritis were collected as the training set,and 668 images as the test set.Unet++had a 97.2%hit rate for chronic gastritis.The area under the ROC curve of Resnet-50 convolutional neural network in diagnosis of chronic atrophic gastritis was 0.97,with the accuracy,sensitivity and specificity of 90.0%,90.3%and 89.6%,respectively.Resnet-50 convolutional neural network showed high consistency with standard answers for diagnosis of chronic atrophic gastritis(κ=0.763),which was similar to those of endoscopy experts(κ=0.712,κ=0.698),and superior to that of general endoscopist(κ=0.585).Conclusions:The convolutional neural network-based auxiliary recognition model established in this study is helpful for the diagnosis of chronic gastritis by white light endoscopy,which has a good prospect of application in clinical practice.