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。
ObjectiveTo 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.
MethodsThirty-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.
ResultsCompared 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 was good with kappa value 0.81 ( P<0.001) . In the aspect of artificial recognition of cartilage injury grading in MRI images, the sensitivity of artificial recognition was 92.10% compared with the manual identification of cartilage injury classification in MRI images. The specificity of artificial recognition was 91.60%. The positive predictive value (PPV) of artificial recognition was 97.20% and the negative predictive value (NPV) of artificial recognition was 78.8%. The kappa value of the cartilage injury classification in MRI images consistency between artificial recognition and manual identification was 0.79 ( P<0.001).
ConclusionBased on the evaluation of cartilage injury by AI reconstruction model of knee joint, the sensitivity and specificity of the diagnosis of ICRS grade IV cartilage injury can be acceptable, but still needs to be improved.
高宏,薛宾阁,吴厦,等. 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[J]. 中华骨科杂志,2023,43(05)：316-321.
DOI:10.3760/cma.j.cn121113-20221013-00615版权归中华医学会所有。
未经授权，不得转载、摘编本刊文章，不得使用本刊的版式设计。
除非特别声明，本刊刊出的所有文章不代表中华医学会和本刊编委会的观点。
高宏：研究实施、数据分析与解释、统计分析、论文撰写与修改；薛宾阁：数据整理、统计分析；吴厦：论文审阅；王亚魁：资料收集；付鹏飞、娄佳旺：资料收集、数据整理；沈乐：软件及电脑程序设计；马琦、刘璞：研究设计与指导、统计分析；蔡谞：手术操作、技术指导、方法学指导
文档即答
智问智答
机器翻译
你好，我可以帮助您更好的了解本文，请向我提问您关注的问题。
