摘要
目的基于深度学习的方法建立腕关节直接数字平板X线成像系统(DR)自动质控模型生成系统,并对质控性能进行初步研究。方法采用人工智能深度学习的方法建立腕关节正位和侧位DR图像质控系统模型,回顾性收集重庆大学附属中心医院临床怀疑腕关节病变的1315张图像,按6︰4的比例划分训练集和验证集。在MobileNet V2分类模型和Global Universal U-Net(GU2Net)关键点检测模型上进行训练,然后分别使用模型准确率、精准度、召回率和曲线下面积(AUC),以及平均径向误差(MRE)和成功检出率(SDR)进行评估。结果根据验证数据集的实验所得到的伪影分类模型在伪影识别方面具有较高的性能,AUC=0.9701,95%可信区间(95%CI)0.9700~0.9703,其准确率、精准度、召回率分别为0.93、0.88和0.97。正位和侧位影像中关键点检测模型的MRE也在合理水平,分别达到(0.7944±3.2535)mm和(3.8134±7.4087)mm。距离10.0 mm下正位和侧位关键点检测模型的SDR分别为99.64%、92.51%。结论基于深度卷积神经网络开发的全自动腕关节DR质控系统模型,能够对腕关节正位和侧位片自动生成图像质量控制报告,且效果较好。
Objective To establish an automatic quality control system for wrist joint direct digital flat panel X-ray imaging system(DR)based on deep learning methods and conduct preliminary studies on quality control performance.Methods This study employed artificial intelligence deep learning techniques to develop a quality control system model for anteroposterior and lateral wrist joint DR images.A retrospective collection of 1315 images from patients clinically suspected of having wrist joint lesions from Central Hospital Affiliated to Chongqing University was performed.The dataset was divided into a training set and a validation set at a ratio of 6∶4.Training was conducted on the MobileNet V2 classification model and the Global Universal U-Net(GU2Net)keypoint detection model,followed by evaluation using model accuracy,precision,recall rate,area under the curve(AUC),mean radial error(MRE),and successful detection rate(SDR).Results Experimental results on the validation dataset showed that the artefact classification model achieved high performance in artefact recognition,with an AUC=0.9701,95%confidence interval(95%CI)0.9700-0.9703,and its accuracy,precision and recall rate were 0.93,0.88 and 0.97,respectively.The MRE of the keypoint detection model in anteroposterior and lateral images was also within a reasonable range,with MRE values of(0.7944±3.2535)mm and(3.8134±7.4087)mm,respectively.The SDRs of the forward and lateral key point detection models at a distance of 10.0 mm were 99.64%and 92.51%,respectively.Conclusion The fully automatic wrist joint DR quality control system model,developed based on deep convolutional neural networks,can automatically generate image quality control reports for anteroposterior and lateral wrist joint images,with favourable results.
作者
彭超
张剑
刘欢
黄英
刘羽
PENG Chao;ZHANG Jian;LIU Huan;HUANG Ying;LIU Yu(Department of Medical Imaging,Chongqing University Central Hospital/Chongqing Emergency Medical Center,Chongqing 400014,China;CEC D-COMMERCE TECHCo Ltd,Shanghai 200131,China;Department of Medical Imaging,Chongqing Public Health Medical Center/Southwest University Public Health Hospital,Chongqing 400030,China)
出处
《现代医药卫生》
2024年第6期907-912,917,共7页
Journal of Modern Medicine & Health
基金
重庆市卫生健康委员会医学科研项目(2023WSJK114)。
关键词
腕关节
直接数字平板X线成像系统
质量模型
深度学习
模型
卷积神经网络
Wrist joint
Direct digital flat panel X-ray imagine system
Quality model
Deep learning
Model
Convolutional neural network
