首页|基于神经网络深度学习模型的踝关节X线片标志点自动定位研究

基于神经网络深度学习模型的踝关节X线片标志点自动定位研究

扫码查看
原文链接
  • 万方数据
  • 维普
目的 探索基于神经网络深度学习模型的踝关节X线片标志点自动定位方法及其应用价值.方法 选取陕西省人民医院2019年1月至2022年11月间行X线检查的360例成年人正常左踝关节正、侧位片影像资料为研究对象,将其随机分配至训练集(210例)、验证集(90例)和测试集(60例).以人工标注作为参考,对图像预处理后分别建立基于神经网络Unet架构的踝关节X线片标志点预测模型,生成对应的热力图,并用测试集数据进行验证.结果 在踝关节X线正位片6个标志点的预测中,2 mm阈值的平均正确估计比例(Percentage of Correct Keypoints,PCK)可达99.7％,总体平均径向误差(Mean Radial Errors,MRE)为0.411,总体标准差(Standard Deviation,SD)为0.290.距骨顶端内点的预测准确度最高,1 mm阈值时的PCK可达100％,同时其MRE及SD在正位片6个点中最小,分别为0.290和0.178.在踝关节X线侧位片9个标志点的预测中,2mm阈值的平均PCK达到95.0％,总体MRE为0.669,总体SD为0.710.胫骨下段最前点的预测准确度最高,1 mm阈值时的PCK可达100％,同时其MRE及SD在侧位片9个点中最小,分别为0.334和0.173.正位片和侧位片所有标志点的预测位置坐标与对应参考标准标志点位置坐标差异均无统计学意义(P＞0.05).结论 基于神经网络深度学习模型能够实现对踝关节X线片标志点的有效自动定位,对辅助踝关节X线片形态学自动测量和疾病诊疗具有应用价值.
Research on Automatic Positioning of Ankle Joint X-Ray Landmarks Based on Neural Network Deep Learning Model
Objective To explore the automatic positioning method of ankle joint X-ray landmarks based on neural network deep learning model and its application value.Methods The normal left ankle joint anterior-lateral X-ray images of 360 adults between January 2019 and November 2022 were obtained from Shaanxi Provincial People's Hospital as research objects,and were randomly assigned to the training set(210 cases),validation set(90 cases)and test set(60 cases).With manual annotation as a reference,the prediction models of ankle joint landmarks based on neural network Unet architecture were established after image preprocessing,and the corresponding thermal maps were generated,and verified with the test set data.Results In the prediction of the 6 landmarks of the ankle joint X-ray anterior images,the average percentage of correct keypoints(PCK)of 2 mm threshold reached 99.7％,the total mean radial errors(MRE)was 0.411,and the total standard deviation(SD)was 0.290.The prediction accuracy of inner point at the top of talus(IPTT)among the 6 points was the highest,and the PCK of this point at 1 mm threshold reached 100％,and its MRE and SD were also the smallest among the 6 points of X-ray anterior images,which were 0.290 and 0.178 respectively.In the prediction of the 9 landmarks of the ankle joint X-ray lateral images,the PCK of this point at 1 mm threshold reached 95％,and the total MRE was 0.669,and the total SD was 0.710.The prediction accuracy of anterior point of lower tibia(APLT)among the 9 points was the highest,and the PCK of this point at 1 mm threshold reached 100％,and its MRE and SD were also the smallest among the 9 points of X-ray lateral images,which were 0.334 and 0.173 respectively.There was no statistical difference between the predicted position coordinates and the corresponding reference standard position coordinates of all the landmarks in the ankle joint X-ray anterior images and lateral images(P＞0.05).Conclusion The neural network deep learning model can realize the effective automatic positioning of ankle joint X-ray landmarks,which has application value in assisting the automatic measurement of ankle X-ray morphology and disease diagnosis and treatment.

ankle jointlandmark automatic positioningX-ray imagingdeep learning modelneural networkUNet architecturemorphological automatic measurement

刘沁峰、胡师尧、张宇琛、常健、刘辉、孙正明、凌鸣、王涛

展开 >

陕西省人民医院医学装备部,陕西西安 710068

西安电子科技大学雷达信号处理国家重点实验室,陕西西安 710071

西安工程大学电子信息学院,陕西西安 710048

西安交通大学计算机科学与技术学院,陕西西安 710049

西北工业大学物理科学与技术学院,陕西西安 710129

陕西省人民医院放射科,陕西西安 710068

陕西省人民医院骨科,陕西西安 710068

西安交通大学口腔医院陕西省颅颌面精准医学研究重点实验室,陕西西安 710004

西安交通大学口腔医院医学影像科,陕西西安 710004

展开 >

踝关节 标志点自动定位 X线成像 深度学习模型,神经网络 UNet架构 形态学自动测量

陕西省重点研发计划社会发展项目西安交通大学基本科研业务费自由探索与创新-教师类项目陕西省重点研发计划社会发展项目陕西省人民医院科技发展孵化基金项目

2021SF-173xzy0120210672024SF-YBXM-4432023YJY-18

2024

10.3969/j.issn.1674-1633.2024.10.008

中国医疗设备
中国整形美容协会

中国医疗设备

CSTPCD
影响因子：0.825
ISSN：1674-1633
年,卷(期)：2024.39(10)