基于多模态MRI影像组学的胶质母细胞瘤生境亚区预测MGMT启动子甲基化表达

Prediction of habitat subregions of the glioblastoma microenvironment based on multimodal MRI radiomics for MGMT promoter methylation expression

目的探讨胶质母细胞瘤(glioblastoma,GBM)不同肿瘤生境亚区多模态影像组学模型用于术前预测GBM O6-甲基鸟嘌呤-DNA甲基转移酶(O6-methylguanine-DNA methyltransferase,MGMT)启动子甲基化的效能。材料与方法回顾性分析来自湖北医药学院附属太和医院、宾法尼西亚大学、加州大学弗朗西斯科分校的600例GBM患者的术前MRI图像、临床和基因信息。自动分割预处理后的图像,得到GBM的三个肿瘤生境亚区(增强区域、坏死区域和水肿区域)。从术前MRI图像[对比增强T1加权成像(contrast enhanced T1-weighted imaging,T1WI-CE)序列、T2液体衰减反转恢复(T2 fluid attenuation inversion recovery,T2-FLAIR)序列和弥散张量成像(diffusion tensor imaging,DTI)的各向异性分数(fractional anisotropy,FA)参数图]上分别提取三个亚区的2153个影像组学特征。通过相关性分析、最小冗余最大相关算法(minimum redundancy maximum relevance,MRMR)和Boruta算法进行特征筛选,使用XGBoost算法构建模型。以受试者工作特征(receiver operating characteristic,ROC)曲线及其曲线下面积(area under the curve,AUC)、准确度、敏感度和特异度等指标评价模型诊断效能,DeLong检验对比模型差异性。结果在训练集和测试集两个亚型的不同临床特征的组间比较差异无统计学意义(P>0.05)。特征筛选后获得10个来自多模态生境亚区的特征。多模态生境亚区影像组学模型在训练集和测试集上的AUC分别为0.874和0.899。结论术前MRI影像组学模型可以预测GBM患者MGMT基因启动子甲基化状态,多序列组合模型的诊断效能更具鲁棒性,肿瘤生境亚区的研究为GBM患者的分子分型的精确诊断、替莫唑胺(temozolomide,TMZ)使用的决策、生存期预测提供了重要的临床辅助价值。

Objective:To explore the efficacy of multimodal imaging radiomics models of different tumor microenvironment subregions in predicting the methylation status of the O6-methylguanine-DNA methyltransferase(MGMT)promoter in glioblastoma before surgery.Materials and Methods:A retrospective analysis was conducted on preoperative MRI images,clinical,and genetic information of 600 glioblastoma patients from Taihe Hospital,Hubei University of Medicine,University of Pennsylvania,and University of California,San Francisco.The preprocessed images were automatically segmented to obtain three subregions of the tumor microenvironment.From the preoperative MRI images[contrast enhanced T1-weighted imaging(T1WI-CE),T2 fluid attenuation inversion recovery(T2-FLAIR)sequence,and diffusion tensor imaging(DTI)fractional anisotropy(FA)maps],2153 radiomics features were extracted from three habitat subregions,including enhanced region,necrotic region and edema region.Feature selection was performed using correlation analysis,minimum redundancy maximum relevance(MRMR),and Boruta algorithm,and the XGBoost algorithm was used to build classification model.The diagnostic performance of the models was evaluated using receiver operating characteristic(ROC)curves,area under the curve(AUC),accuracy,sensitivity,specificity,and DeLong test for model comparison.Results:There were no statistically significant differences in the intergroup comparisons of clinical features between the two subtypes in the training and testing sets(P>0.05).The multimodal imaging radiomics model for the enhanced region had AUCs of 0.842 and 0.935 in the training and validation sets,respectively.Ten features from the multimodal habitat subregions were obtained after feature selection.The AUC of the imaging omics model in the multimodal habitat subregion was 0.874 and 0.899 on the training and test sets,respectively.Conclusions:The preoperative MRI radiomics models can predict the MGMT promoter methylation status in glioblastoma patients,and the multimodal combination models showed more robust diagnostic performance.The study of tumor microenvironment subregions provides important clinical utility for accurate molecular subtyping,decision-making for temozolomide(TMZ)use,and survival prediction in glioblastoma patients.