MRI影像组学无创预测脑胶质母细胞瘤MGMT启动子甲基化状态的价值

Value of MRI radiomics in the noninvasive prediction of O 6-methylguanine-DNA methyltransferase promoter methylation status in brain glioblastomas

目的:探讨MRI影像组学模型术前预测脑胶质母细胞瘤(GBM)O 6-甲基鸟嘌呤-DNA甲基转移酶启动子甲基化(MGMT-PM)状态的价值。方法:回顾性分析2018年1月-2021年10月在本院经病理证实的130例脑GBM患者的临床资料和MRI图像(ADC和对比增强3D-T 1WI)。其中,MGMT-PM阳性组(PM率≥8%)58例,MGMT-PM阴性组(PM率<8%)72例。按7:3的比例将所有患者随机分为训练集(91例)和验证集(39例)。由2位放射科医师独立在ADC和CE-3D-T 1WI图像上逐层勾画ROI,获得病灶的全域容积感兴趣区(VOI),分别提取851个组学特征。然后,采用最小绝对收缩和选择算法(LASSO)进行特征降维,将保留下来的特征与其对应的系数进行线性组合,构建影像组学模型并计算每例患者的影像组学评分(Radscore),得到Radscore ADC、Radscore CE-T 1WI和Radscore联合三组评分。采用ROC曲线评估各组学模型的诊断效能,将最优模型的Radscore和临床特征(年龄、性别)纳入logistic回归分析构建预测MGMT-PM状态的临床-组学综合模型,并绘制其诺模图。采用ROC曲线评价综合模型的预测效能,并采用校准曲线和决策曲线分别评估此模型的校准度和临床实用价值。结果:在训练集中,Radscore联合预测MGMT-PM状态的AUC为0.872,优于单一序列(Radscore ADC:AUC=0.798,P<0.05;Radscore CE-T 1WI:AUC=0.840,P<0.05);在验证集中得到了一致的结论。在影像组学模型中加入临床特征后,可提高预测效能,临床-组学综合模型的AUC、敏感度和特异度分别为0.904、92.50%和78.43%。校准曲线显示临床-组学综合预测模型在训练集和验证集中预测概率与实际概率之间的差异均无统计学意义(P=0.051、0.284)。决策曲线分析表明综合预测模型具有一定的临床实用价值。结论:MRI影像组学模型有助于术前无创性预测GBM的MGMT启动子甲基化状态,多序列结合及引入临床特征能提高模型的预测效能。

Objective:To investigate the value of MRI radiomics in the preoperative noninvasive prediction of O6-methylguanine-DNA methyltransferase(MGMT)promoter methylation(PM)status in glioblastomas(GBM).Methods:The clinical data and MRI images(including ADC and 3D contrast-enhanced T 1-weighted imaging)of 130 GBM patients confirmed by pathology from January 2018 to October 2021 were retrospectively analyzed.There were 58 cases with positive MGMT-PM status(PM ratio≥8%)and 72 cases with negative MGMT-PM status(PM ratio<8%).All patients were randomly assigned into a training dataset(n=91)and validation dataset(n=39)at a ratio of 7:3.Region of interest(ROI)was delineated on all slices of the lesions by two radiologists independently,and whole volume of interest(VOI)of lesion was obtained.Then,851 radiomics features were extracted based on ADC and 3D-CE-T 1WI sequences,respectively.The least absolute shrinkage and selection operator(LASSO)method was used for feature dimension reduction.Radiomics models were built and radscores were calculated by linear combination of retained features and their corresponding coefficients.And three Radscores named Radscore ADC,Radscore CE-T 1WI and Radscore union were obtained.The receiver operating characteristic(ROC)curves were plotted and AUCs were calculated for evaluation of the diagnostic efficacy of the three radiomics models,and the radscore of the optimal model and clinical characteristics(age and gender)were incorporated to perform logistic regression analysis for establishing the MGMT-PM status clinical-radiomics comprehensive prediction model.A nomogram was plotted for realizing the visualization of the prediction model.ROC curve was plotted to evaluate the prediction performance of the comprehensive prediction model.The calibration curve and decision curve analysis were applied to evaluate the calibration and clinical utility of the comprehensive model.Results:Radscore union based on combination of two sequences for predicting MGMT-PM status with AUC of 0.872 in the training set was superior to the radscores from single sequence(Radscore ADC:AUC=0.798,P<0.05;Radscore CE-T 1WI:AUC=0.840,P<0.05),which was consistent with the result in the validation set.The addition of clinical characteristics to the radiomics prediction model improved the predictive efficacy with AUC,sensitivity and specificity of 0.904,92.50%and 78.43%,respectively.The calibration curve showed that whether in training or validation set,there were no significant difference between the predicted probability of the clinical-radiomics comprehensive prediction model and the actual probability(P=0.051 and 0.284,respectively).The analysis of decision curve concluded that the comprehensive prediction model showed certain clinical value.Conclusion:MRI radiomics prediction models contribute to preoperative noninvasive prediction of MGMT-PM status in brain glioblastomas.The combination of multi-sequences and the addition of clinical characteristics can improve the prediction performance.