Abstract:

Objective: To explore the potential value of expert - identified CT and MRI imaging features in predicting the MGMT promoter methylation status in glioma. Methods: A retrospective analysis was conducted in 188 patients in The First Medical Center of Chinese PLA General Hospital from January 2019 to December 2020 with pathologically confirmed glioma. Imaging features were extracted, including calcification, clear lesion margins, peritumoral edema, T₂WI/T₂ - FLAIR mismatch, cortical involvement, subventricular zone involvement, insular involvement, homogeneous signal on T₂WI, and enhanced lesions. Pyrosequencing was used to detect the MGMT promoter methylation status. Univariate and multivariate Logistic regression analyses were used to find the imaging feature factors that affect the MGMT promoter methylation status. Then, by plotting the receiver operating characteristic (ROC) curve, verify the predictive efficacy of the imaging features. For the prediction task, further train and test 4 machine learning (ML) models, namely Logistic regression (LR), support vector machine (SVM), random forest (RF), and gradient boosting (GB). Results: Logistic regression analysis showed that homogeneous signal on T₂WI (OR = 2.843, 95%CI: 1.055-7.658; P = 0.039) and enhanced lesions (OR = 0.146, 95%CI: 0.069-0.308; P = 0.000) were imaging feature factors affecting the MGMT promoter methylation status. The comprehensive parameters combining both had higher prediction ability compared with homogeneous signal on T 2WI (Z = 3.961, P = 0.000) and enhanced lesions (Z = 2.233, P = 0.026). The prediction accuracies rates of LR, SVM, RF and GB models were 0.84, 0.76, 0.68 and 0.76, respectively. However, there were no statistically significant differences in prediction efficacy when comparing the models pairwise (P > 0.05, for all). Conclusions: Imaging features based on preoperative CT and MRI show promise for non - invasive prediction of MGMT promoter methylation status in glioma.

Key words: Glioma, Magnetic resonance imaging, O(6)-methylguanine-DNA methyltransferase, Logistic models, Machine learning

摘要：

目的: 探讨基于人眼识别的CT和MRI影像学特征预测胶质瘤MGMT启动子甲基化状态的潜在价值。方法: 回顾分析2019年1月至2020年12月解放军总医院第一医学中心收治的188例首次确诊胶质瘤患者的临床资料, 纳入病灶钙化、病灶边界清晰、T₂WI信号均匀性、病灶呈强化改变等9个人眼可识别的术前常规CT或MRI影像学特征, 通过焦磷酸测序分析MGMT启动子甲基化状态。采用单因素和多因素Logistic回归分析筛查MGMT启动子甲基化状态的影像学特征影响因素, 并绘制受试者工作特征（ROC）曲线验证影像学特征的预测效能。针对预测任务进一步训练并测试Logistic回归、支持向量机、随机森林和梯度提升4种机器学习模型。结果: Logistic回归分析显示, T₂WI信号均匀性（OR = 2.843, 95%CI: 1.055 ~ 7.658;P = 0.039）和病灶呈强化改变（OR = 0.146, 95%CI: 0.069 ~ 0.308;P = 0.000）是MGMT启动子甲基化状态的影像学特征影响因素。T₂WI信号均匀性和病灶呈强化改变联合的综合参数较T₂WI信号均匀性（Z = 3.961, P = 0.000）和病灶呈强化改变（Z = 2.233, P = 0.026）的预测效能更高。Logistic回归、支持向量机、随机森林和梯度提升4种机器学习模型的预测准确度分别为0.84、0.76、0.68和0.76, 但预测效能两两比较差异无统计学意义（均P > 0.05）。结论: 基于术前CT和MRI的影像学特征具备无创预测胶质瘤MGMT启动子甲基化状态的潜力。

关键词: 神经胶质瘤, 磁共振成像, O(6)-甲基鸟嘌呤DNA甲基转移酶, Logistic模型, 机器学习