基于影像组学的机器学习模型鉴别孤立性纤维性肿瘤与血管瘤型脑膜瘤  

Machine learning models based on radiomics in differentiating solitary fibrous tumor from angiomatous meningioma

作　　者：毕玉珍 白洁[2] 白培瑞 李向荣[4] 付圣莉 王键 任延德[1] BI Yuzhen;BAI Jie;BAI Peirui;LI Xiangrong;FU Shengli;WANG Jian;REN Yande(Department of Radiology,the Affiliated Hospital of Qingdao University,Qingdao 266000,China;Department of Magnetic Resonance,the Frist Affiliated Hospital of Zhengzhou University,Zhengzhou 450000,China;College of Electronic Information Engineering,Shandong University of Science and Technology,Qingdao 266555,China;Department of Radiology,the Frist Affiliated Hospital of Guangxi Medical University,Nanning 530000,China)

机构地区：[1]青岛大学附属医院放射科,青岛266000 [2]郑州大学第一附属医院磁共振科,郑州450000 [3]山东科技大学电子信息工程学院,青岛266555 [4]广西医科大学第一附属医院放射科,南宁530000

出　　处：《磁共振成像》2023年第9期50-55,共6页Chinese Journal of Magnetic Resonance Imaging

基　　金：青岛市医药卫生科研计划项目(编号:2021-WJZD192);青岛市市南区科技计划项目(编号:2022-4-010-YY)。

摘　　要：目的探讨基于MRI图像影像组学特征的机器学习模型鉴别孤立性纤维性肿瘤(solitary fibrous tumor,SFT)与血管瘤型脑膜瘤(angiomatous meningioma,AM)的价值。材料与方法回顾性分析两个中心经病理证实的SFT患者病例68例、AM患者病例41例。运用3D Slicer软件对T1加权成像(T1-weighted imaging,T1WI)、液体衰减反转恢复(fluid-attenuated inversion recovery,FLAIR)、T1WI增强轴位图像进行预处理、感兴趣区(region of interest,ROI)勾画及特征提取,应用独立样本t检验和最小绝对收缩和选择算子(least absolute shrinkage and selection operator,LASSO)进行特征选择。联合T1WI、FLAIR及T1WI增强筛选多参数MRI序列的最佳特征。按照7∶3的比例将患者随机分为训练集(76例)与测试集(33例),使用logistic回归(logistic regression,LR)、随机森林(random forest,RF)、支持向量机(support vector machine,SVM)构建预测模型,并绘制受试者操作特征(receiver operating characteristic,ROC)曲线,计算准确度、敏感度、特异度及ROC曲线下面积(area under the curve,AUC)。采用DeLong检验比较不同模型之间AUC的差异。结果AM组平均年龄明显高于SFT组,差异有统计学意义(P<0.001),两组性别组成差异无统计学意义(P>0.05)。基于T1WI、FLAIR、T1WI增强以及多参数MRI序列分别筛选出22、12、12、65个最佳影像组学特征。多参数MRI序列模型诊断效能优于单序列模型,其中SVM模型效能最好,测试集AUC为0.99。单序列模型中,T1WI和FLAIR模型的诊断效能优于T1WI增强模型。LR模型的AUC均大于0.9。结论基于MRI图像影像组学特征的机器学习模型可以鉴别SFT与AM,多参数MRI序列模型的效能较好,其中SVM模型的效能最高,LR模型具有较好的效能及稳定性。Objective:To investigate the value of machine learning models based on MRI radiomics features in differentiating solitary fibrous tumor(SFT)from angiomatous meningioma(AM).Materials and Methods:A total of 68 patients with SFT and 41 patients with AM confirmed by pathology from the Affiliated Hospital of Qingdao University and the First Affiliated Hospital of Guangxi Medical University were retrospectively enrolled.The pre-processing,delineation of the region of interest(ROI),and feature extraction of the T1-weighted images(T1WI),fluid-attenuated inversion recovery(FLAIR),and contrast-enhanced T1WI were performed in the 3D slicer software.The optimal feature set was selected by independent-samples t test and least absolute shrinkage and selection operator(LASSO).The optimal features of multi-parameter MRI were selected based on T1WI,FLAIR and contrast-enhanced T1WI.All patients were randomly divided into the training group(n=76)and the test group(n=33)at a ratio of 7∶3.The models were established by logistic regression(LR),random forest(RF),and support vector machine(SVM).Receiver operating characteristic(ROC)curves were drawn,respectively,and accuracy,sensitivity,specificity,and area under the curve(AUC)were calculated.The DeLong test was used to compare the differences in AUCs among different models.Results:The average age of the AM group was higher than that of the SFT group(P<0.001).There was no significant difference in gender composition between AM group and SFT group(P>0.05).Twenty-two,twelve,twelve and sixty-five radiomics features were extracted from T1WI,FLAIR,contrast-enhanced T1WI and multi-parameter MRI,respectively.The differentiation efficiency of models based on multi-parameter MRI between intracranial SFT and AM was better than that of models based on a single sequence.The SVM model based on multi-parameter MRI reached the highest performance of all models,and the AUC was 0.99.Among models based on a single sequence,differentiation efficiency of models based on T1WI or FLAIR was better than that

关 键 词：孤立性纤维性肿瘤 血管瘤型脑膜瘤 磁共振成像 影像组学 机器学习 

分 类 号：R445.2[医药卫生—影像医学与核医学] R730.262[医药卫生—诊断学] R739.45[医药卫生—临床医学]