基于深度学习算法从X线图像识别手关节炎的诊断研究  

A diagnostic study of hand arthritis recognized from X-ray images based on deep learning algorithm

作　　者：杨丽 王欢[1] 王婷婷[2] 吴建红[2] YANG Li;WANG Huan;WANG Tingting;WU Jianhong(Department of Clinical Medicine,North Sichuan Medical College,Nanchong 637000,China;Department of Rheumatology and Immunology,Dazhou Central Hospital,Dazhou 635000,China)

机构地区：[1]川北医学院临床医学院,四川南充637000 [2]达州市中心医院风湿免疫科,四川达州635000

出　　处：《现代医学》2024年第7期1043-1049,共7页Modern Medical Journal

基　　金：四川省医学科技创新研究会基金资助项目(YCH-ZZ2023-011)。

摘　　要：目的:基于深度学习算法构建一种从手X线图像中识别类风湿关节炎(RA)和手骨关节炎(OA)的诊断模型。方法:回顾性纳入2017年1月至2023年4月在达州市中心医院被诊断为RA的509例患者的960张单手X线图像和2016年1月至2023年4月在达州市中心医院被诊断为手OA的112例患者的216张单手X线图像。利用人工智能中的深度学习算法构建模型,分别对RA和手OA患者X线图像的目标关节进行检测,并进行the modified Sharp/van der Heijde Score(SHS)和Kellgren&Lawrence(K-L)分级。通过测试集评估模型的性能,最终建立从X线图像中自动完成RA和手OA骨破坏分级的模型。结果:模型在RA目标关节检测及其骨破坏的关节间隙狭窄程度分类方面,健康关节、轻度骨破坏关节、重度骨破坏关节和所有关节的精度-召回率曲线下面积(PR-AUC)分别为90.7%、76.3%、76.6%和81.2%。模型在手OA的目标关节检测及其骨破坏的关节间隙狭窄和骨赘程度分类方面,健康关节、轻度骨破坏关节、重度骨破坏关节和所有关节的PR-AUC分别为94.5%、93.8%、86.9%和91.7%。结论:本研究构建的深度学习诊断模型,能快速准确地识别RA和手OA患者X线图像中的目标关节,同时做出骨破坏的分级,具有良好的诊断效能,能辅助医生诊断RA和手OA。Objective:Construction of a diagnostic model for recognizing rheumatoid arthritis(RA)and hand osteoarthritis(OA)from hand X-ray images based on a deep learning algorithm.Methods:960 single hand X-ray images of 509 patients diagnosed with RA at Dazhou Central Hospital from January 2017 to April 2023 and 216 single hand X-ray images of 112 patients diagnosed with hand OA at Dazhou Central Hospital from January 2016 to April 2023 were included retrospectively.Deep learning algorithms in artificial intelligence were utilized to construct model to detect the target joints in X-ray images of patients with RA and hand OA,respectively,and to grade the target joints with the modified Sharp/van der Heijde Score(SHS)and Kellgren&Lawrence(K-L).The performance of the model was evaluated through a test set,culminating in the creation of a model that automates the grading of RA and OA bone destruction from X-ray images.Results:The area under the precision-recall curve(PR-AUC)of the model in terms of RA target joint detection and classification of the degree of joint space narrowing for bone destruction in the target joints was 90.7%,76.3%,76.6%,and 81.2%for healthy joints,mildly bone-damaged joints,severely bone-damaged joints,and all joints,respectively.The PR-AUC of the model in terms of hand OA target joint detection and classification of the degree of joint space narrowing for bone destruction in the target joints was 94.5%,93.8%,86.9%,and 91.7%for healthy joints,mildly bone-damaged joints,severely bone-damaged joints,and all joints,respectively.Conclusion:The deep learning diagnostic model constructed in this study can quickly and accurately identify the target joints in the X-ray images of patients with RA and hand OA,as well as make the grading of bone destruction,which has good diagnostic efficacy and can assist doctors in diagnosing RA and hand OA.

关 键 词：深度学习 X线图像 类风湿关节炎 手骨关节炎 辅助诊断 

分 类 号：R593.22[医药卫生—内科学]