髋关节关键角度人工智能测量对成人发育性髋关节发育不良的辅助诊断作用  

AI-assisted diagnosis of hip dysplasia:accuracy and efficiency in measuring key radiographic angles

作　　者：李瑞欣 王肖 张蓓蓓 李天然[1] 刘晓鸣 隋琪瑞 李文华 Li Ruixin;Wang Xiao;Zhang Beibei;Li Tianran;Liu Xiaoming;Sui Qirui;Li Wenhua(Department of Radiology,The Fourth Medical Center of Chinese PLA General Hospital,Beijing 100048,China;Beijing Lianying Intelligent Imaging Technology Research Institute,Beijing 100048,China)

机构地区：[1]解放军总医院第四医学中心放射科,北京100048 [2]北京联影智能影像技术研究院,北京100094

出　　处：《中华骨科杂志》2024年第22期1464-1473,共10页Chinese Journal of Orthopaedics

摘　　要：目的探讨人工智能(artificial intelligence,AI)模型在骨盆X线片上测量髋关节关键角的准确性,并评价AI模型对成人发育性髋关节发育不良(developmental dysplasia of the hip,DDH)和临界型DDH(borderline developmental dysplasia of the hip,BDDH)的诊断效能。方法回顾性分析来源于解放军总医院第四医学中心放射科1029例可疑DDH患者的病历资料,男273例、女756例,年龄(57.01±18.16)岁(范围12~88岁)。随机分配到训练集720例、测试集206例和验证集103例。由两名放射科医生在骨盆正位X线片上确定并标记髋关节关键解剖点,应用训练集进行深度学习,构建定位髋关节关键解剖点的AI模型,AI模型基于髋关节关键解剖点自动测量并计算Sharp角、Tönnis角和中心边缘(center-edge angle,CE)角。将放射科医生测量结果与AI模型自动测量结果进行比较,用于评估AI模型对测试集测量结果的准确性。验证集用于优化模型参数,测试集用于评估DDH的诊断性能。绘制受试者工作特征(receiver operating characteristic curve,ROC)曲线,计算ROC曲线下面积(area under roc curve,AUC)评价AI模型对DDH和BDDH的诊断效能。结果AI模型测量髋关节左侧Sharp角、Tönnis角及CE角诊断DDH的准确率分别为89.8%、86.8%、90.1%,右侧的准确率为93.7%、80.5%、92.2%,人工测量平均值和AI模型测量的Sharp角、Tönnis角和CE角的差异均无统计学意义(P>0.05)。AI模型与人工测量Sharp角、Tönnis角和CE角的相关性检验及一致性分析结果显示r值及组内相关系数(intraclass correlation coefficient,ICC)均>0.75。AI模型测量用时(1.7±0.1)s,较放射科医生人工测量用时的(88.1±8.4)s和(90.3±7.4)s更短(P<0.05)。AI模型测量得到的Sharp角、Tönnis角、CE角诊断DDH的AUC分别为0.883、0.908、0.922(左侧)和0.924、0.922、0.871(右侧);AI模型测量左、右侧CE角诊断BDDH的AUC分别为0.787和0.676。AI模型和人工测量诊断DDH和BDDH与临床最终诊断�ObjectiveTo evaluate the accuracy of an artificial intelligence(AI)model in measuring key angles on pelvic radiographs of the hip and assess its effectiveness in diagnosing developmental dysplasia of the hip(DDH)and borderline developmental dysplasia of the hip(BDDH).MethodsA retrospective analysis was conducted using anteroposterior pelvic X-ray films from 1,029 patients with suspected DDH.The data were collected from the Department of Radiology,Fourth Medical Center of the Chinese PLA General Hospital.Among the patients,273 were male,and 756 were female,with an average age of 57.01±18.16 years(range,12-88 years).The dataset was randomly divided into a training set(720 cases),a test set(206 cases),and a validation set(103 cases).Two radiologists identified and marked key anatomical points of the hip joint to establish the training dataset,which was then used to develop a deep learning-based AI model capable of locating these key anatomical positions.Using the identified anatomical points,the AI model automatically measured and calculated the Sharp angle,center-edge(CE)angle,and Tönnis angle in the test dataset.The measurement results from the AI model were compared with those of the radiologists to evaluate the model's accuracy.The validation set was used to optimize model parameters,and the test dataset was used to evaluate the diagnostic performance of DDH.Receiver operating characteristic(ROC)curves were employed to assess the diagnostic efficacy of the AI model for DDH and BDDH.ResultsThe accuracy rates of the AI model in measuring the left Sharp angle,CE angle,and Tönnis angle for diagnosing DDH were 89.8%,90.1%,and 86.8%,respectively.For the right side,the accuracy rates were 93.7%,92.2%,and 80.5%,respectively.There were no statistically significant differences in the mean values of the Sharp,Tönnis,and CE angles between manual and AI measurements(P>0.05).Pearson correlation tests and intraclass correlation coefficient(ICC)analyses revealed high consistency between AI and manual measurements of the Sha

关 键 词：发育性髋关节发育不良 骨盆 放射摄影术 人工智能 深度学习 

分 类 号：R687.4[医药卫生—骨科学]