基于Faster R卷积神经网络构建胸部X线片异物智能检测模型的可行性研究  

Feasibility of constructing the intelligent detection model for foreign bodies on chest X-ray based on Faster R-convolutional neural network

作　　者：孟宇 马之骋 阮敬儒 高阳 杨柏林[2] 何林阳 龚向阳[1] Meng Yu;Ma Zhicheng;Ruan Jingru;Gao Yang;Yang Bailin;He Linyang;Gong Xiangyang(Department of Radiology,Center for Rehabilitation Medicine,Zhejiang Provincial People′s Hospital,Affiliated People′s Hospital,Hangzhou Medical College,Hangzhou 310014,China;School of Computer Science and Information Engineering,Zhejiang Gongshang Univeristy,Hangzhou 310018,China;Hangzhou Jianpei Technology Company Ltd.,Hangzhou 311200,China)

机构地区：[1]浙江省人民医院,杭州医学院附属人民医院康复医学中心放射科,杭州310014 [2]浙江工商大学计算机与信息工程学院,杭州310018 [3]杭州健培科技有限公司,杭州311200

出　　处：《中华放射学杂志》2022年第12期1359-1364,共6页Chinese Journal of Radiology

基　　金：浙江省重点研发计划(2020C01058)。

摘　　要：目的构建基于Faster R卷积神经网络的后前位胸部X线片异物智能检测模型,并评估模型的性能。方法回顾性分析2019年6月至2020年3月浙江省人民医院和淳安县人民医院的成人后前位DR胸片5567张,其中含异物胸片4247张。分为异物训练集(2911张异物胸片)、验证集(1456张,733张含异物、723张无异物)和测试集(1200张,603张含异物、597张无异物)。每张胸片的异物经过2名放射住院医师标注和1名高年资放射技师审核校正后的结果作为参考金标准。采用受试者操作特征(ROC)曲线及曲线下面积分析深度学习模型在测试集中区分胸片有无异物的效能,采用精准率-召回率曲线及平均精确度(mAP)分析模型在不同层级的稳定性。最后分析不同位置、患者性别、患者年龄对于深度学习模型的异物召回率的影响。结果测试集中,深度学习模型诊断胸片是否含有异物的灵敏度为93.2%(562/603),特异度为92.6%(553/597),F1分数为0.94,ROC曲线下面积为0.97,mAP值为0.69。对于不同位置的异物,肺野内和肺野外的异物检测的召回率分别为91.2%(674/739)和89.0%(1411/1585)。对于不同性别的患者,男性和女性的异物检测召回率分别为87.3%(337/386)和90.0%(1745/1938)。对于不同的年龄分段,18~38岁的异物检测召回率为92.5%(1041/1126),39~58岁的异物检测召回率为89.7%(505/563),59~78岁的异物检测召回率为83.5%(335/401),≥79岁的异物检测召回率为85.9%(201/234)。结论构建的基于深度学习的成人后前位胸部X线片异物检测模型具有很高的灵敏度和稳定性,可以快速准确地识别胸片中的异物。Objective To construct an intelligent foreign bodies detection model based on Faster R-convolutional neural network in posterior-anterior chest X-ray and evaluate the performance of the model.Methods Totally 5567 adult posterior-anterior DR chest radiographs from Zhejiang Provincial People′s Hospital and Chun′an County People′s Hospital from June 2019 to March 2020,with 4247 foreign body-containing chest radiographs were analyzed retrospectively.All data were randomly divided into training set(2911 foreign body-containing),validation set(n=1456,733 foreign body-containing,723 free of foreign body)and testing set(n=1200,603 foreign body-containing,597 free of foreign body).The reference gold standard was set as the results of each chest radiography with foreign body annotated by two radiology residents and reviewed and corrected by a senior radiographer.The receiver operating characteristic(ROC)curve and the area under the curve were used to analyze the efficiency of the deep learning model to distinguish the presence or absence of foreign bodies on chest radiography in the testing set.The precision-recall curve and mean precision(mAP)were used to analyze the stability of the model at different levels.Finally,the influence of different locations,patient gender,and patient age on the foreign body recall of the deep learning model were analyzed.Results In the testing set,the sensitivity of the deep learning model in diagnosing whether chest radiograph contained foreign bodies was 93.2%(562/603),the specificity was 92.6%(553/597),and the F1 score was 0.94.The area under the ROC curve was 0.97,and the mAP value was 0.69.For foreign bodies in different locations,the recall rates of foreign bodies in lung field and outside lung field were 91.2%(674/739)and 89.0%(1411/1585),respectively.For different genders,the recall rates for male and female foreign body detection were 87.3%(337/386)and 90.0%(1745/1938),respectively.For different age ranges,the recall rate of foreign body detection was 92.5%(1041/1126)for 18-38

关 键 词：X线 胸 异物 质量控制 深度学习 

分 类 号：R816.4[医药卫生—放射医学]