基于快速区域卷积神经网络对于胸部数字X线摄影图像异物的自动检测  

Automatic Detection of Foreign Objects in Chest Digital Radiography Image Based on Faster Region Convolutional Neural Network

作　　者：张庆雷[1] 陈飞[2] 李茗[1] 戴真煜 姚立正 张鑫[1] 蔡金凤[3] 方宇 郭晓鹏 辛小燕[1] 青钊 张冰[4] ZHANG Qinglei;CHEN Fei;LI Ming(Department of Radiology,The Affiliated Drum Tower Hospital of Nanjing University Medical School,Nanjing,Jiangsu Province 210008,P.R.China)

机构地区：[1]南京大学医学院附属鼓楼医院医学影像科,210008 [2]南京医科大学盐城临床医学院(盐城市第三人民医院)影像科,盐城224005 [3]东软集团股份有限公司临床医疗事业部,沈阳110179 [4]南京大学脑科学研究院,210008

出　　处：《临床放射学杂志》2022年第2期367-371,共5页Journal of Clinical Radiology

基　　金：国家自然科学基金项目(编号:81720108022);中央高校基本科研业务费专项资金资助项目(编号:021414380462);江苏省重点研发计划(社会发展)项目(编号:BE2016605)。

摘　　要：目的探讨快速区域卷积神经网络(Faster RCNN)在胸部数字X线摄影(DR)图像异物位置和类型自动检测中的应用价值。方法对960张胸部DR图像进行Faster RCNN训练、验证及测试,按3∶1∶1的比例随机划分数据集为训练集(576张)、验证集(192张)和测试集(192张)。使用开源分割工具ImageJ对左、右肺野及异物(共13类)进行标注,其中左、右肺野采用区域标记,异物采用包围框标记。肺野分割精确程度采用肺野预测模型度量函数IoU值表示。异物位置标注结果分为三类:图像无异物,图像有肺野内异物,图像有肺野外异物。将检测结果用三分类混淆矩阵表示,并计算三类图像的召回率、精确率及最终分类的准确率。并输出对各类异物检测的平均精确率(AP)和总平均精确率(mAP)。结果训练集和验证集共768张胸部DR中的645张含有异物,共4655个。测试集左肺分割平均IoU值为0.959,右肺分割平均IoU值为0.958。单例测试耗时5 s。图像无异物,图像有肺野内异物和图像有肺野外异物三类图像的召回率分别为94%、75%和82%,精确率分别为88%、98%和70%;总准确率为86%。对各类异物检测的AP范围为66%~100%,mAP为81%。结论 Faster RCNN可满足影像质量控制工作中自动检测DR图像异物的要求。Objective To explore the application value of faster region convolutional neural network(Faster RCNN) in the automatic detection of foreign objects position and type in chest digital radiology(DR) images. Methods The Faster RCNN training, verification and testing were performed on 960 chest DR images, and the datasets were randomly divided into training set(576 images),verification set(192 images) and testing set(192 images) in a 3∶1∶1 ratio. ImageJ,an open source segmentation tool, was used to mark the left and right lung fields and foreign objects(13 types),where the left and right lung fields were marked with regions, and the foreign objects were marked with bounding boxes.The precision of lung field segmentation was expressed by the measurement function IOU value of lung field prediction model.Foreign objects position labeling results were divided into three categories: images without foreign objects, images with foreign objects in the lung field, images with foreign objects out the lung field.The detection results were expressed by a three-class confusion matrix, and the recall rate, precision rate and final classification accuracy of the three types of images were calculated.The average precision(AP) and mean AP(mAP) of various foreign objects detection were outputted. Results There were 4655 foreign objects in 645 of the 768 chest DR images in the training set and verification set.In testing set, the mean IOU of the left lung segmentation was 0.959,and the mean IOU of the right lung segmentation was 0.958.The test for an image took 5 seconds.The recall rates of images without foreign objects, images with foreign objects in the lung field and images with foreign objects out the lung field were 94%,75% and 82% respectively, and the precision rates of them were 88%,98% and 70% respectively, the final classification accuracy was 86%.The range of AP for various foreign objects was 66% ~ 100%,and the mAP was 81%. Conclusion Faster RCNN can meet the requirements of automatic detection of foreign objects in DR

关 键 词：区域卷积神经网络 胸部数字X线摄影 异物 质量控制 

分 类 号：R816.4[医药卫生—放射医学]