基于机器学习算法的影像组学在MRI诊断早期股骨头坏死中的应用  

Machine Learning-Based MRI Radiomics for Predicting Early Femoral Head Necrosis

作　　者：朱士铭 滕伟[1] 李唯[1] 宣建新[1] ZHU Shiming;TENG Wei;LI Wei;XUAN Jianxin(Department of Radiology,Chengde Central Hospital,Chengde 067000,China)

机构地区：[1]承德市中心医院CT室,河北承德067000

出　　处：《中国医学影像学杂志》2021年第8期817-821,共5页Chinese Journal of Medical Imaging

摘　　要：目的探索基于机器学习的影像组学在MRI诊断早期股骨头坏死中的应用价值。资料与方法回顾性分析接受髋关节MRI检查的293例患者的临床及影像学资料。根据国际骨循环协会股骨头坏死分型,其中118例诊断为股骨头坏死早期。采用随机数字法,以7∶3抽取206例用于训练模型,87例用于验证模型。应用影像组学方法从MRI的T1WI影像中共提取319项影像学特征值。利用LASSO模型挑选9项最优特征值。应用Logistic回归法(LR)、随机森林法(RF)、支持向量机法(SVM)和K邻近法(KNN)4种机器学习算法构建基于影像组学的MRI诊断早期股骨头坏死模型。应用ROC曲线和曲线下面积(AUC)评估模型的诊断效能。结果RF模型AUC为0.86(95%CI 0.83~0.92),敏感度为86.3%、特异度为84.4%;LR模型AUC为0.81(95%CI 0.76~0.84),敏感度为78.3%、特异度为86.3%;KNN模型AUC为0.73(95%CI 0.67~0.75),敏感度为61.7%、特异度为84.8%;SVM模型AUC为0.82(95%CI 0.76~0.89),敏感度为81.3%、特异度为82.4%。验证集中,RF模型的诊断效能最高(P=0.017)。结论基于机器学习的影像组学在MRI诊断早期股骨头坏死中的诊断效能高,具有一定的临床应用前景。Purpose To develop a predictive model using machine learning-based radiomics on the MRI to diagnose early femoral head necrosis(ONFH).Materials and Methods 293 patients’clinical data and hip MRI were retrospectively analyzed.Of them,118 cases were defined as ONFH according to the Association Research Circulation Osseous classification.All cases were divided into training cohort(n=206)and validation cohort(n=87)at a 7∶3 ratio by using random numbers.319 radiomic features were totally extracted from T1 WI of MRI,and then 9 radiomic features were selected using 10-fold cross-validation LASSO model.The diagnostic models of radiomics for early ONFH were developed using four machine learning algorithm,including Logistic regression(LR),random forest(RF),support vector machine(SVM)and K-nearest neighbor(KNN).The predictive performances of models were evaluated and compared by receiver operating curve(ROC)and AUC.Results In the validation dataset,the AUC for RF was 0.86(95%CI 0.83-0.92),the sensitivity and specificity were 86.3%and 84.4%,respectively;the AUC for LR was 0.81(95%CI 0.76-0.84),the sensitivity and specificity were 78.3%and 86.3%,respectively;the AUC for KNN was 0.73(95%CI 0.67-0.75),the sensitivity and specificity were 61.7%and 84.8%,respectively;and the AUC for SVM was 0.82(95%CI 0.76-(15).89),the sensitivity and specificity were 81.3%and 82.4%,respectively.Of them,RF achieved the best predictive value(P=0.017).Conclusion Machine learningbased MRI radiomics may be promising in diagnosis of early ONFH,and has a certain clinical application prospect.

关 键 词：股骨头坏死 磁共振成像 影像组学 机器学习算法 诊断 

分 类 号：R445.2[医药卫生—影像医学与核医学] R681.8[医药卫生—诊断学]