基于深度卷积神经网络的上消化道内镜解剖分类模型构建  

Development of anatomical classification models for upper gastrointestinal endoscopy based on deep convolutional neural networks

作　　者：许郭婷 吴爱荣[1,2] 林嘉希 高欣 周鑫[3] 顾慧媛 许春芳 朱锦舟 XU Guoting;WU Airong;LIN Jiaxi;GAO Xin;ZHOU Xin;GU Huiyuan;XU Chunfang;ZHU Jinzhou(Department of Gastroenterology,the First Affiliated Hospital of Soochow University,Suzhou 215000,China;Suzhou Clinical Center of Digestive Disease,Suzhou 215000,China;Department of Gastroenterology,Jintan Hospital Affiliated to Jiangsu University,Changzhou 213200,China)

机构地区：[1]苏州大学附属第一医院消化内科,江苏苏州215000 [2]苏州市消化病临床医学中心,江苏苏州215000 [3]江苏大学附属金坛医院消化内科,江苏常州213200

出　　处：《中国医学物理学杂志》2023年第8期1051-1056,共6页Chinese Journal of Medical Physics

基　　金：国家自然科学基金(82000540);苏州市科技计划项目(SKY2021038);苏州市科教兴卫项目(KJXW2019001)。

摘　　要：目的:利用深度卷积神经网络构建上消化道内镜解剖分类模型。方法:收集苏州大学附属第一医院消化内镜中心4183张胃镜图片,按照8:2的比例随机分为训练集和验证集;同时收集江苏大学附属金坛医院270张胃镜图片作为测试集。以上图片标注上消化道解剖位置(包括食管、贲门、胃底、胃体、胃角、胃窦、幽门、十二指肠球部及降部)。选择ImageNet数据库中预训练的Xception、NASNet Large(NASNet)和ResNet50V2(ResNet)3个深度卷积神经网络,在训练集及验证集中训练,构建上消化道图片解剖部位分类模型。使用梯度加权分类激活映射对模型的分类结果进行可视化解释。在验证集和测试集中评价模型分类能力。结果:成功构建了基于深度卷积神经网络的上消化道内镜解剖分类的3个模型,各模型均具备较高的分类能力。在验证集中,平均分类准确性为0.980,平均分类召回率为0.894,平均分类精确度为0.920;其中,ResNet模型表现最优,其分类准确性(0.982)、分类召回率(0.905)和分类精确度(0.933)最高。在测试集中,平均分类准确性为0.988,平均分类召回率为0.942,平均分类精确度为0.950;其中,NASNet模型表现最优,其分类准确性(0.992)、分类召回率(0.959)和分类精确度(0.970)最高。梯度加权分类激活映射以热力图形式对模型分类结果提供可视化解释。结论:利用深度卷积神经网络,构建的上消化道内镜解剖分类模型具有较好的分类能力。Objective To develop anatomical classification models for upper gastrointestinal endoscopy using deep convolutional neural networks.Methods:A total of 4183 gastroscopic images collected from the Gastrointestinal Endoscopy Center of the First Affiliated Hospital of Soochow University were randomly divided into training set and validation set at a ratio of 8:2,while 270 gastroscopic images from Jintan Hospital Affiliated to Jiangsu University were collected as the test set.The anatomical structures(esophagus,cardia,gastric fundus,gastric body,gastric angle,gastric antrum,pylorus,duodenal bulb and descending)were labeled in the gastroscopic images.Three deep convolutional neural networks,namely Xception,NASNet Large(NASNet)and ResNet50V2(ResNet),which had been pre-trained in ImageNet database,were trained in training set and validation set for constructing the anatomical classification models for upper gastrointestinal endoscopy.The gradient-weighted class activation mapping was used to visually interpret the classification results of the models,and the classification abilities of the models were evaluated in validation set and test set.Results Three anatomical classification models for upper gastrointestinal endoscopy based on deep convolutional neural network were successfully constructed.All models had high classification ability.In the validation set,the average classification accuracy,recall and precision were 0.980,0.894 and 0.920,respectively.Among them,ResNet model performed best,with the highest classification accuracy(0.982),classification recall(0.905)and classification precision(0.933).In the test set,the average classification accuracy,recall and precision were 0.988,0.942 and 0.950,respectively.Among them,NASNet model performed best,with the highest classification accuracy(0.992),classification recall(0.959)and classification precision(0.970).The gradient-weighted class activation mapping provides a visual interpretation of the model classification results in the form of thermal map.Conclusion:The anat

关 键 词：上消化道 胃镜 解剖定位 深度卷积神经网络 模型构建 

分 类 号：R318[医药卫生—生物医学工程]