基于卷积神经网络的胸片肺野分割与肺结核的筛查研究  

Study on Lung Field Segmentation and Tuberculosis Diagnosis Screening on Chest Radiographs Based on Convolutional Neural Network

作　　者：刘梦秋 伍志发 丁虎 刘影[1] LIU Mengqiu;WU Zhifa;DING Hu(Department of Radiology,The First Affiliated Hospital of USTC,Hefei,Anhui Province 230001,P.R.China)

机构地区：[1]中国科学技术大学附属第一医院(安徽省立医院)影像科,合肥230001 [2]中国科学技术大学计算机科学与技术学院,合肥230026

出　　处：《临床放射学杂志》2023年第4期671-677,共7页Journal of Clinical Radiology

基　　金：国家重点研发计划资助项目(编号:2021YFA000900)。

摘　　要：目的探究基于卷积神经网络的级联深度学习模型在胸部X线平片图像上对肺野分割以及肺结核筛查的应用价值。方法搜集2018年10月至2020年2月行胸部X线摄影检查的健康对照组1300名和肺结核患者825例,随机选择140名健康对照组和60例患者组成肺野分割数据集,评价基于U-net++网络的深度学习模型对胸片肺野的分割效果。划分数据集中的80%(1700例)作为训练集,20%(425例)作为测试集,使用四种分类网络(VGG 16、Inception V3、Resnet 101、Densenet 121)对分割结果内是否存在结核病灶进行判断,并使用网络公开的深圳市第三人民医院肺结核数据集(CHX)对模型的检出效能进行评价。结果级联模型中U-net++分割网络对胸片肺野分割的Dice相似指数与交并比(IOU)分别达到99.42%和98.84%;VGG 16、Inception V3、Resnet 101及Densenet 121四种分类网络对肺结核的筛查率最高分别为95.77%、96.00%、94.35%和95.06%;四种分类网络在CHX数据集上的最高检出率分别为84.44%、83.99%、81.42%和86.25%,曲线下面积(AUC)分别达到0.896、0.881、0.919和0.935。结论基于卷积神经网络的级联深度学习模型对胸片上肺野的分割效果良好,对肺结核的筛出具有较高的应用价值,模型具有一定的泛化能力。Objective The application value of a cascaded deep learning model based on convolutional neural network for lung field segmentation and tuberculosis screening on chest X⁃ray images was disscused.Methods Collected from Oc⁃tober 2018 to February 2020 in our hospital for 1,300 healthy people and 825 tuberculosis patients who underwent chest X⁃ray examination.140 healthy people and 60 patients were randomly selected to form the lung field segmentation dataset,and the effect of the segmentation model based on U⁃net++network was evaluated.We divided 80%(1700 cases)of the dataset as the training set and 20%(425 cases)as the test set,four kinds of classification networks(VGG 16,Inception V3,Resnet 101 and Densenet 121)were used to judge whether there were tuberculosis foci in the segmentation results.The tu⁃berculosis dataset of Shenzhen third people's hospital was used to evaluate the diagnostic efficiency of the model.Results The Dice similarity index and IOU of the segmented network in the cascades model reached 99.42%,98.84%.The four classification networks had the highest screening rates for pulmonary tuberculosis,which were 95.77%(VGG 16),96.00%(Inception V3),94.35%(Resnet 101)and 95.06%(Densenet 121),respectively.The highest detection rates on the CHX data set were 84.44%,83.99%,81.42%and 86.25%,and the area under the curve(AUC)reached 0.896,0.881,0.919 and 0.935 respectively.Conclusion The cascade deep learning model based on convolutional neural network has good segmentation effect on lung field on chest radiography,and has high application value to the screening of tubercu⁃losis,and has certain generalization ability.

关 键 词：深度学习 卷积神经网络 级联模型 肺结核 

分 类 号：R521[医药卫生—内科学] R816.4[医药卫生—临床医学]