基于IDEAL-IQ的影像组学模型预测腰椎椎体低骨量的价值研究  

Value of IDEAL IQ⁃Based Radiomics Model in Predicting Bone Mass Loss of Lumbar Vertebra

作　　者：成东亮 关炜 文戈[3] 冯红梅[1] 刘健萍[1] 张文强 吴耀忠 高明勇[1] CHENG Dongliang;GUAN Wei;WEN Ge(Department of Radiology,First People's Hospital of Foshan,Foshan,Guangdong Province,528000,P.R.China)

机构地区：[1]佛山市第一人民医院影像科,528000 [2]佛山市第一人民医院核医学科,528000 [3]南方医科大学南方医院影像教研室,广州510515

出　　处：《临床放射学杂志》2023年第4期645-650,共6页Journal of Clinical Radiology

基　　金：佛山市卫生健康局医学科研项目(编号:20220050);佛山市“十四五”高水平医学重点专科项目(编号:FSGSP145036)。

摘　　要：目的探讨基于MRI最小二乘法迭代水脂分离定量技术(IDEAL-IQ)序列的影像组学模型对识别腰椎椎体低骨量的价值。方法搜集2019年1月至2022年2月经双能X线吸收测定法(DXA)诊断腰椎椎体低骨量的患者40例,为低骨量组,另纳入同期经DXA诊断腰椎椎体骨密度在正常参考值范围内,且年龄及性别与低骨量组相匹配的健康志愿者66名,为正常对照组,两组均接受IDEAL-IQ序列扫描。根据DXA结果,最终530个椎体符合研究标准,包括正常椎体330个椎体,低骨量椎体200个。应用MaZda软件在每个腰椎椎体上进行感兴趣区(ROI)勾画,然后提取279个影像组学特征,将530个椎体按照分层随机抽样方法以7∶3的比例分为训练集和验证集,采用两样本t检验、Mann-Whitney U检验进行特征降维,后通过LASSO回归分析从保留的特征中选择最优特征子集,建立Logistic回归(LR)、随机森林(RF)、支持向量机(SVM)模型,采用受试者工作特征曲线(ROC)、曲线下面积(AUC)、准确率、敏感度和特异度评估模型预测效能,Delong检验比较不同模型的效能。结果在提取的279个影像组学特征中,通过降维后最终保留24个与腰椎低骨量显著相关的影像组学特征用于构建机器学习模型;在训练集中,LR、RF、SVM模型的AUC分别为0.928、0.926、0.95;在验证集分别为0.897、0.917、0.948。经Delong检验,训练集中LR、RF、SVM模型AUC值无统计学差异(P>0.05),验证集中SVM模型AUC值高于LR(P=0.003),RF与SVM、LR模型AUC值无统计学差异。结论基于MRI IDEAL-IQ的影像组学模型能无创、准确的评估低骨量状态,有利于早期干预,改善患者预后。Objective To explore the value of radiomics model based on MR IDEAL⁃IQ sequence in identifying bone mass loss of lumbar vertebra.Methods From January 2019 to February 2022,totally 40 patients with low bone mass of lumbar vertebrae diagnosed by DXA were divided into low bone mass group,and age and sex matched healthy volunteers di⁃agnosed by DXA within the normal reference in the same period were included as normal control group.Both groups received IDEAL⁃IQ sequence scanning.Finally,530 vertebrae including 330 normal and 200 osteopenia vertebrae met the study crite⁃ria.279 radiomics features in the FF map of IDEAL⁃IQ sequence by MaZda software of each vertebra ROI extracted.Then 530 vertebrae were randomly divided into training set and verification set at the ratio of 7∶3.Two⁃sample t test and Mann⁃Whitney U test were used to select the features from the extracted radiomics features with R software.Furthermore,LASSO regression analysis and 10⁃fold cross⁃validation method were further used to selected the optimal feature subset.LR,RF and SVM models were established,effectiveness of the three prediction modelswere evaluated by the area under the receiver op⁃erating characteristic curve(AUC),accuracy,sensitivity and specificity.Results Among theextracted 279 radiomicsfea⁃tures,24 optimal features significantly related to bone mass loss of lumbar vertebrawere finally retained and used to establish machine learning model.ROC curve showed that the AUC of LR,RF and SVM models in the training set were 0.928,0.926 and 0.95 respectively,and those in the verification set were 0.897,0.917 and 0.948,respectively.According to Delong test,there is no statistical difference in AUC values of LR,RF and SVM models in training set(P>0.05);AUC values of SVM models in verification set are higher than LR(P=0.003);AUC values of RF,SVM and LR models are not statisti⁃cally different.Conclusion Radiomics model based on MRI IDEAL⁃IQ could helpidentifythe state of bone mass lossaccu⁃rately and noninvasively,whic

关 键 词：低骨量 IDEAL-IQ 影像组学 机器学习 

分 类 号：R580[医药卫生—内分泌] R445.2[医药卫生—内科学]