基于半月板MRI的3D卷积神经网络模型预测膝骨关节炎发生的研究  

Predicting the occurrence of knee osteoarthritis based on MRI meniscus 3D convolutional neural network model

作　　者：蒋可欣 谢雨含 李勉文 张志勇 陈少龙 丘昌镇 张晓东 JIANG Kexin;XIE Yuhan#;LI Mianwen;ZHANG Zhiyong;CHEN Shaolong;QIU Changzhen;ZHANG Xiaodong(Department of Radiology,the Third Affiliated Hospital,Southern Medical University(Academy of Orthopedics,Guangzhou),Guangzhou 510630,China;School of Electronics and Communication Engineering,Sun Yat-sen University,Guangzhou 510275,China)

机构地区：[1]南方医科大学第三附属医院(广东省骨科研究院)影像科,广州510630 [2]中山大学电子与通信工程学院,广州510275

出　　处：《磁共振成像》2024年第2期103-107,121,共6页Chinese Journal of Magnetic Resonance Imaging

基　　金：南方医科大学第三附属医院院长基金项目(编号:YM2021012)。

摘　　要：目的探究基于自动分割半月板MRI的3D卷积神经网络(convolutional neural network,CNN)模型预测膝骨关节炎(knee osteoarthritis,KOA)发生的潜在价值。材料与方法本回顾性研究数据均来自于公开数据库骨关节炎倡议(Osteoarthritis Initiative,OAI)。随机选择基线时的130例膝关节MRI图像,由经过训练的肌骨诊断医师手动勾画半月板感兴趣区,训练半月板MRI分割模型。采用OAI骨关节炎发生队列的MRI图像分割半月板,并基于3D CNN构建KOA预测模型。该关节炎发生队列共纳入710例膝关节,基线时均无放射学KOA,即Kellgren-Lawrence(KL)分级均≤1。在48个月的随访期间,发生放射学KOA(KL分级≥2)为病例组,未发生放射学KOA为对照组。病例组与对照组以1∶1的比例进行匹配。分别利用基线和确定放射学KOA前1年随访时间点(P-1)的膝关节MRI图像构建KOA预测模型。采用Dice系数评估半月板MRI分割模型性能。采用受试者工作特征曲线下面积(area under the curve,AUC)评估基于半月板MRI、临床信息和MRI骨关节炎膝关节评分(MRI Osteoarthritis Knee Score,MOAKS)构建的预测模型的预测价值。结果本研究半月板分割模型在测试集的Dice系数达90.32%。在基线和P-1时间点,3D CNN KOA预测模型(基线时AUC:0.60;P-1时AUC:0.71)比基于临床信息的模型(基线时AUC:0.55;P-1时AUC:0.63)及MOAKS(基线时AUC:0.52~0.56;P-1时AUC:0.51~0.64)在测试集中表现出更好的预测能力,且差异存在统计学意义(P<0.05)。结论基于自动分割半月板MRI构建的3D CNN KOA预测模型较临床信息或MRI半定量评分能更好地预测放射学KOA的发生。Objective:To explore the potential value of a 3D convolutional neural network(CNN)model based on automatically segmented meniscus MRI in predicting the occurrence of knee osteoarthritis(KOA).Materials and Methods:This retrospective study used data from the Osteoarthritis Initiative(OAI),a publicly available database.A total of 130 baseline knee joint MRI images were randomly selected,and the meniscus regions of interest were manually delineated by trained musculoskeletal radiologists to train the meniscus MRI segmentation model.The meniscus segmentation was performed on the incident osteoarthritis cohort of OAI,and a 3D CNN model for KOA prediction was constructed.The incident osteoarthritis cohort included 710 knee joints with baseline Kellgren-Lawrence(KL)grading of≤1,and no radiographic KOA at baseline.During a 48-month follow-up,cases with radiographic KOA(KL grade≥2)were considered as the case group,while those without radiographic KOA served as the control group,matched in a 1∶1 ratio.KOA prediction models were built using baseline and the time point one year before the occurrence of radiographic KOA(P-1)knee joint MRIs.The Dice coefficient was used to evaluate the performance of the meniscus MRI segmentation model.The predictive value of models based on meniscus MRI,clinical information,and MRI Osteoarthritis Knee Score(MOAKS)was assessed using the area under the curve(AUC)of the receiver operating characteristic curve.Results:The meniscus segmentation model achieved a Dice coefficient of 90.32% on the test set.At baseline and P-1 time points,the 3D CNN KOA prediction model(baseline AUC:0.60;P-1 AUC:0.71)outperformed models based on clinical information(baseline AUC:0.55;P-1 AUC:0.63)and MOAKS(baseline AUC:0.52-0.56;P-1 AUC:0.51-0.64)in the test set,with statistically significant differences(P<0.05).Conclusions:The 3D CNN KOA prediction model based on automatically segmented meniscus MRI demonstrates superior predictive capabilities for the occurrence of radiographic knee osteoarthritis compared to c

关 键 词：膝骨关节炎 半月板 磁共振成像 卷积神经网络 分割 预测 

分 类 号：R445.2[医药卫生—影像医学与核医学] R684.3[医药卫生—诊断学]