基于深度学习的内镜肠道准备评分模型的建立  

Computer Vision Models for Endoscopic Bowel Preparation Scoring Based on Deep Learning

作　　者：沈文娟[1] 徐昶 林嘉希 许春芳[1] 陆建英[1] 朱锦舟 SHEN Wenjuan;XU Chang;LIN Jiaxi;XU Chunfang;LU Jianying;ZHU Jinzhou(Department of Gastroenterology,The First Affiliated Hospital of Soochow University,Suzhou Jiangsu 215006,China)

机构地区：[1]苏州大学附属第一医院消化内科,江苏苏州215006

出　　处：《中国医疗设备》2023年第11期11-15,共5页China Medical Devices

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

摘　　要：目的基于深度学习算法构建内镜肠道准备评分的计算机视觉模型。方法收集苏州大学附属第一医院消化内镜中心(600张)及HyperKvasir数据库(1794张)的内镜图片共2394张,根据Boston肠道准备量表完成肠道清洁度评分(0~3分,四分类),按6∶2∶2随机分为训练集(1439张)、验证集(478张)和测试集(477张)。选取3种深度学习网络(DenseNet169、DenseNet121、EfficientNet B3),利用迁移学习方式训练肠道准备分类模型,并采用测试集的混淆矩阵等指标评价模型分类能力,与高、低年资医师的分类能力进行对比。结果成功构建3个基于深度学习的肠道准备分类模型。各模型的分类准确度均较高,平均分类准确度为0.897,近似于低年资内镜医师(0.914),低于高年资内镜医师(0.941)的分类表现。其中,DenseNet169模型表现最好,分类准确度(0.914)及平均精确度(0.892)均为最高。此外,采用梯度加权分类激活映射算法,用热力图形式对模型的分类推理进行可视化呈现。结论运用深度学习算法构建的内镜肠道准备分类模型具有可行性,可通过多中心研究扩大样本来源进一步提高模型的分类及泛化能力。Objective To develop computer vision models for endoscopic bowel preparation scoring based on deep learning.Methods A total of 2394 endoscopic images from the Gastrointestinal Endoscopy Centre of the First Affiliated Hospital of Soochow University(n=600)and the HyperKvasir database(n=1794)were collected,scored by endoscopists according to the Boston bowel preparation scale(BBPS,0-3,four categories).They were randomly divided into training sets(1439 pieces),verification sets(478 pieces)and test sets(477 pieces)according to 6∶2∶2.Three deep learning networks(DenseNet169,DenseNet121,EfficientNet B3)were used to develop the bowel preparation classification models by transfer learning.Metrics such as confusion matrices in the test set were used for model evaluation.Meanwhile,the models were compared with senior and junior endoscopists.Results The three deep learning-based bowel preparation classification models were successfully developed.The classification accuracy of all models was high,and the average classification accuracy was 0.897,which was similar to the clarification performance of junior endoscopist(0.914)and lower than that of senior endoscopist(0.941).Among them,the best performing model was the DenseNet169 model,which had the highest classification accuracy(0.914)and the highest average precision(0.892).In addition,the visual interpretation of the models’classification results was presented in the form of heat maps by using gradient-weighted class activation mapping.Conclusion The endoscopic bowel preparation classification model developed using deep learning is feasible,and the classification and generalization ability of the model can be further improved by expanding the sample source through multicenter studies.

关 键 词：深度学习 计算机视觉 卷积神经网络 梯度加权分类激活映射 

分 类 号：R197.39[医药卫生—卫生事业管理] TP391.4[医药卫生—公共卫生与预防医学]