目的 探讨多序列MRI影像组学特征联合常规征象鉴别诊断脑膜瘤与其他颅内脑膜起源肿瘤的价值.材料与方法 回顾性分析经病理证实的两个中心共360例患者的临床及术前MRI资料.中心1患者256例(脑膜瘤145例、非脑膜瘤111例),按7∶3的比例随机分为训练组(n=179)和测试组(n=77);中心2患者104例作为外部验证组(脑膜瘤53例,非脑膜瘤51例).评估肿瘤的生长部位、生长方式、数目等18项一般临床资料及MRI常规征象,采用单变量及多变量二元logistic回归分析筛选与鉴别诊断相关的指标.图像标准化后,利用3D Slicer软件于T2WI、扩散加权成像(diffusion-weighted imaging,DWI)、T1WI增强图像勾画感兴趣区(region of interest,ROI)并进行特征提取;采用5折交叉验证法和最小绝对收缩和选择算子(least absolute shrinkage and selection operator,LASSO)进行特征筛选.训练组、测试组采用逻辑回归(logistic regression,LR)、支持向量机(support vector machine,SVM)、K近邻(K-nearest neighbor,KNN)、轻量级梯度提升机(light gradient boosting machine,LightGBM)、自适应增强(adaptive boosting,AdaBoost)5种分类器进行建模,建立MRI常规模型、组学瘤内模型、组学瘤周模型、组学融合模型、全融合模型,筛选出其中效能最佳的模型进行外部验证.绘制受试者工作特征(receiver operating characteristic,ROC)曲线,评估模型的鉴别诊断效能;使用DeLong检验对模型曲线下面积(area under the curve,AUC)进行比较;使用决策曲线分析(decision curve analysis,DCA)评估模型的临床价值.结果 不同分类器构建的同一模型效能不同,其中SVM模型整体效能较高,测试组组学瘤内SVM模型AUC为0.889,除此之外训练组、测试组所有SVM模型AUC均大于0.900.组学瘤内模型、组学瘤周模型效能相当,二者均高于MRI常规模型;组学融合模型效能高于三者,但全融合模型效能最佳;其在外部验证组中亦表现良好,AUC为0.925,准确率为88.5%,DCA显示该模型在大范围阈值内可以为患者带来临床净收益.结论 基于多序列MRI影像组学特征模型可以在术前鉴别脑膜瘤与其他颅内脑膜起源肿瘤,联合常规征象可以提升模型效能;不同分类器对模型效能有影响,SVM模型效能高,稳健且泛化能力好.
The value of multi-sequence MRI-based radiomics in differential diagnosis of meningioma
Objective:To evaluate the value of multi-sequence MRI features combined with routine signs in differentiating meningioma from other other intracranial meningeal tumors.Materials and Methods:Clinical and preoperative MRI data of 360 patients confirmed by pathology in two centers were retrospectively analyzed.A total of 256 patients(145 meningiomas and 111 non-meningiomas)in center 1 were randomly divided into the training group(n=179)and the test group(n=77)at a ratio of 7∶3.A total of 104 patients in Center 2 served as the external validation group(53 meningiomas and 51 non-meningiomas).The tumor growth site,growth pattern,number and other 18 general clinical data and MRI routine signs were evaluated.Univariate and multivariate binary logistic regression analysis was used to screen the indicators related to differential diagnosis.After image standardization,3D Slicer software was used to outline region of interest(ROI)and extract features on T2WI,diffusion-weighted imaging(DWI)and enhanced T1WI images.The feature screening was performed by using the method of 5-fold cross-validation and least absolute shrinkage and selection operator(LASSO).The training group and the test group were modeled by five classifiers:logistic regression(LR),support vector machine(SVM),K-nearest neighbor(KNN),light gradient boosting machine(LightGBM)and adaptive boosting(AdaBoost).MRI conventional model,radiomics intratumoral model,radiomics peritumoral model,radiomics fusion model,and full fusion model were established,and the models with the best performance were selected for external verification.The receiver operating characteristic(ROC)curve was plotted to evaluate the differential diagnostic performance of the model.The area under the curve(AUC)of the model was compared by DeLong test.Decision curve analysis(DCA)was used to assess the clinical value of the model.Results:The effectiveness of the same model constructed by different classifiers was different.The overall efficiency of the SVM models was higher,and the AUC of the radiomics intratumoral SVM model in the test group was 0.889.In addition,the AUC of all SVM models in the training group and the test group was greater than 0.900.The efficacy of the radiomics intratumoral model and the radiomics peritumoral model were similar,both of which were higher than the MRI conventional model,while the efficacy of the radiomics fusion model was higher than that of the three,but the efficacy of the full fusion model was the best,and it also performed well in the external validation group,with an AUC of 0.925 and an accuracy of 88.5%.DCA showed that this model could bring clinical net benefits to patients within a wide range of thresholds.Conclusions:Multi-sequence MRI-based radiomics model can be used to distinguish meningioma from other intracranial meningeal tumors before surgery,and combined with conventional signs can improve the effectiveness of the model.Different classifiers have influence on model efficiency,SVM model has high efficiency,robustness and good generalization ability.
国家科技期刊平台
NSTL
万方数据
