目的 探讨基于深度学习方法训练模型用于双髋正位X线图像全髋关节置换术(THA)后自动测量外展角的可行性.方法 收集行THA后X线检查资料,共得到381例有效数据用于训练模型.由2位影像科医师标注双侧髋关节假体关键点(髋臼外上缘、内下缘和泪滴点),将其分为训练集(304例)、调优集(38例)和测试集(39例),训练2D U-net模型分割关键点并自动测量外展角.训练完成后,再收集合格图像共143例,用于验证模型的效能.在测试集和外部验证集中,以Dice相似性系数(DSC)、平均绝对误差(MAE)评价模型分割关键点的效能,并以Bland-Altman分析评价模型自动测量外展角的定量值与医师测量的一致性.结果 测试集和外部验证集中,关键点DSC为0.870~0.905和0.690~0.750,MAE为0.311~0.561和0.951~1.310.Bland-Altman 分析显示测试集和外部验证集中,模型测量与医师测量外展角的一致性较高,分别只有6.52%(3/46)、2.08%(3/144)的点在95%一致性界限(LoA)以外.在测试集和外部验证集中,模型对外展角的定性判断与医师的一致率为97.8%和90.3%.结论 基于深度学习模型可在双髋关节正位X线图像上自动测量THA后外展角,与医师测量水平基本一致.
Objective To explore the feasibility of automating the measurement of abduction angle after total hip arthroplasty(THA)on postoperative radiographs by using deep learning algorithms.Methods The data were retrospectively collected.A total of 381 cases were used to develop deep learning model.Two radiologists annotated the key points on the images(lateral-superior point and medial-inferior point of acetabular cups,tear drops).The data was split into training dataset(304 cases),tuning dataset(38 cases),and test dataset(39 cases).A 2D U-net model was trained to segment the key points and the abduction angle were automatically meas-ured.After development of the model,an external validation dataset was collected(143 cases).Dice similarity coefficient(DSC)and mean absolute error(MAE)were used to evaluate the prediction efficiency of the model in the test dataset and the external validation dataset.Bland-Altman test was used to analyze the agreement between the abduction angle measured automatically by the model and the physician measurement.Results The DSC were 0.870-0.905 and 0.690-0.750 in the test dataset and the external validation dataset,and the corresponding MAE were 0.311-0.561 and 0.951-1.310.For the result of Bland-Altman analysis,only 6.52%(3/46)and 2.08%(3/144)of the abduction angle measurements in the test dataset and external validation dataset were outside the 95%limit of agreement(LoA).In the qualitative evaluation of the abduc-tion angle,the agreement of the model with the physician were 97.8%and 90.3%in the test dataset and the external validation dataset.Conclusion It is feasible to use deep learning algorithms to automatically measure the abduction angle after THA on X-ray images,achieving similar accuracy to that of physician.