基于膝关节MRI的人工智能重建软骨模型对膝关节软骨损伤的评价  

Evaluation of knee cartilage based on MRI artificial intelligence reconstruction model of knee joint

作　　者：高宏[1] 薛宾阁 吴厦 王亚魁 付鹏飞 沈乐[4] 娄佳旺 马琦 刘璞 蔡谞 Gao Hong;Xue Binge;Wu Sha;Wang Yakui;Fu Pengfei;Shen Le;Lou Jiawang;Ma Qi;Liu Pu;Cai Xu(Beijing Tsinghua ChanggungHospital,School of Clinical Medicine,Tsinghua University,Beijing 102218,China;Beijing Qinghe-Hospital,Beijing 100085,China;Beijing MEDERA Medical Group,Beijing 102200,China;Department of Engineering Physics,Tsinghua University,Beijing 100084,China)

机构地区：[1]清华大学附属北京清华长庚医院,清华大学临床医学院,北京102218 [2]北京清河医院,北京100085 [3]北京美德瑞骨科医生集团,北京102200 [4]清华大学工程物理系,北京100084 [5]河北北方学院,张家口075132

出　　处：《中华骨科杂志》2023年第5期316-321,共6页Chinese Journal of Orthopaedics

摘　　要：目的探讨基于膝关节薄层MRI数据的人工智能(artificial intelligence,AI)重建模型对膝关节软骨损伤评价的可行性。方法选取2021年5月至2022年4月在北京清华长庚医院以膝关节重度骨关节炎住院且拟行全膝关节置换术的33例患者(共41膝),男15例,年龄(71±5)岁;女26例,年龄(71±9)岁。左膝19例,右膝22例。术前对患侧膝关节进行薄层MR检查,并对膝关节薄层MRI数据进行AI建模,选取模型中软骨部分利用主成分分析(principal component analysis,PCA)进行模型摆正,将术中截取的膝关节胫骨平台软骨依据国际软骨修复协会(International Cartilage Repair Society,ICRS)软骨损伤分级进行分级,并与膝关节AI重建软骨模型及膝关节MRI人工识别的ICRS分级结果进行比较。结果AI重建软骨模型的软骨损伤分级与术中截取实物标本的软骨损伤分级相比较,AI重建软骨模型对于ICRS分级4级软骨损伤诊断的敏感性、特异性、阳性预测值和阴性预测值分别为93.1%、91.4%、92.2%和80.3%;ROC曲线下面积(AUC)值为0.92,AI重建软骨模型与实物标本ICRS分级的一致性良好,Kappa系数为0.81(P<0.001)。人工识别MRI的软骨损伤分级与实物标本相比,其诊断的敏感性、特异性、阳性预测值和阴性预测值分别为92.10%、91.60%、97.20%和78.8%,Kappa系数为0.79(P<0.001)。结论基于膝关节薄层MRI的AI重建软骨模型对于诊断ICRS分级4级软骨损伤具有较好的敏感性及特异性。Objective To explore the feasibility of the AI intelligent reconstruction model based on knee joint magnetic resonance data developed by Nuctech Company Limited for evaluating knee cartilage injury.Methods Thirty-three patients(a total of forty-one knees)who were hospitalized with severe knee osteoarthritis in Beijing Tsinghua Changgung Hospital from May 2021 to April 2022 were selected.All of them were planned to be performed total knee arthroplasty(TKA)for the treatment of knee osteoarthritis.Fifteen males with an average age of 71±5 years old and twenty six females with an average age of 71±9 years old were included in this study.There were 19 cases of left knee and 22 cases of right knee.Thin layer MRI examination on the patients'knee joints was performed before the surgery,and artificial intelligence model based on the thin layer MRI data of the knee joint was reconstructed.The cartilage part of the model was selected and corrected by Principal Component Analysis(PCA)in order to realize model straightening.The tibial plateau cartilage of knee joint which intercepted during operation was classified according to the International Cartilage Repair Society(ICRS).Finally the results were compared with the ICRS classification results of knee artificial intelligence reconstruction model and artificial recognition of knee joint MRI images.Results Compared with the grade of cartilage injury intercepted during our operation which was according to the ICRS classification,the sensitivity of artificial intelligence reconstruction model for the diagnosis of cartilage injury with ICRS classification grade four was 93.1%.The specificity of artificial intelligence reconstruction model was 91.4%.The positive predictive value(PPV)of artificial intelligence reconstruction model was 92.2%.And the negative predictive value(NPV)of artificial intelligence reconstruction model was 80.3%.The area under ROC curve(AUC)was 0.92.The ICRS classification consistency between artificial intelligence model and physical inspection results wa

关 键 词：膝关节 软骨 关节 磁共振成像 人工智能 模型 解剖学 

分 类 号：R687.4[医药卫生—骨科学] R445.2[医药卫生—外科学]