基于膝关节正位X线应用深度学习构建胫股关节骨关节炎的自动诊断模型  

Application of Deep Learning to Develop Models for Automatic Diagnosis and Grading of Tibiofemoral Osteoarthritis Based on Knee Radiographs

作　　者：额·图娅 李晓庆 孙兆男 张耀峰 孙玉梦 张晓东[1] 王霄英[1] E Tuya;LI Xiaoqing;SUN Zhaonan(Department of Radiology,Peking University First Hospital,Beijing 100034,R.R.China)

机构地区：[1]北京大学第一医院医学影像科,100034 [2]北京赛迈特锐医疗科技有限公司,100011

出　　处：《临床放射学杂志》2023年第8期1298-1303,共6页Journal of Clinical Radiology

摘　　要：目的探索深度学习方法基于膝关节正位X线构建自动诊断并分级胫股关节骨关节炎(TFOA)诊断模型。方法搜集5837幅膝关节前后位的X线图像,按照8∶1∶1的比例随机分为训练集、调优集和测试集。以两位医师依据凯尔格伦-劳伦斯(K-L)分级及国际骨关节炎研究协会(OARSI)分级系统共同阅片结果分别作为TFOA及其影像特征[骨赘(OST)及关节间隙狭窄(JSN)]分类模型的参考标准。利用高分辨率网络(HRNet)算法建立上述对应的二分类诊断[K-L.0~1(无TFOA)vs.K-L.2~4(有TFOA);OARSI 0(无OST/JSN)vs.OARSI 1~3(有OST/JSN)]及多分类分级模型(K-L.0~4,分别代表无、可疑、轻度、中度及重度TFOA;OARSI 0~3,依次代表无、轻度、中度及重度OST/JSN)。以受试者工作特征曲线(ROC)、查准率-查全率曲线(P-R)及混淆矩阵评价模型的分类效能。结果在测试集中,二分类模型诊断TFOA、OST及JSN的曲线下面积(AUC)为0.95~0.99、P-R的AUC为0.86~0.98;多分类模型在分级TFOA、OST及JSN中的宏准确率为0.90~0.98、敏感度为0.09~0.99、特异度为0.69~1.00。结论以HRNet为基础架构的模型可自动诊断并分级TFOA、OST及JSN,其二分类诊断效能较高,多分类则对病变程度分级提供了可能性。Objective To explore applying deep learning-based models for diagnosing and grading tibiofemoral osteoarthritis(TFOA)from knee radiographs.Methods A total of 5837 anterior-posterior knee radiographs were retrospectively collected and randomly divided into the train set,validation set,and test set according to the ratio of 8∶1∶1.Two radiologists interpreted the reference standards of classification models in consensus according to the Kellgren-Lawrence(K-L)grading system for TFOA and Osteoarthritis Research Society International(OARSI)grading system for osteophyte(OST)and joint space narrowing(JSN),respectively.We used a high-resolution net(HRNet)algorithm to establish models.The binaryclass models were developed to diagnose the absence or presence of TFOA,OST,and JSN,i.e.,absence(K-L.0-1/OARSI 0)vs.presence(K-L.2-4/OARSI 1-3).The multi-class models were developed to grade the severity of TFOA,OST,and JSN.K-L.0 to K-L.4 represented no TFOA,doubtful TFOA,mild TFOA,moderate TFOA,and severe TFOA in order.OARSI 0 to 3 represented no OST/JSN,doubtful OST/JSN,mild OST/JSN,moderate OST/JSN,and severe OST/JSN in order.Model performances were evaluated using the receiver operating characteristic(ROC),precision-recall(P-R)curve,and confusion matrix.Results In the test sets,the area under the curve(AUC)of ROCs and P-R curves of the binary-class models for TFOA,OST,and JSN were 0.95-0.99,0.86-0.98,respectively;the macro accuracy,sensitivity,and specificity of the multi-class models in grading TFOA,OST and JSN were 0.90-0.98,0.09-0.99,and 0.69-1.00,respectively.Conclusion The HRNet-based models could automatically diagnose and grade TFOA,OST,and JSN.Its binary-class diagnostic efficacy is high.The multi-class models may provide the possibility to grade the severity of OA of the tibiofemoral joint.

关 键 词：膝关节骨关节炎 X线摄影 深度学习 分类 

分 类 号：R684.3[医药卫生—骨科学] TP18[医药卫生—外科学] TP391.41[医药卫生—临床医学]