Magnetic Resonance Imaging-Based Radiomics For Preoperative Prediction Of MGMT Promoter Methylation Status And Risk Of Recurrence Within One Year In IDH1 Wild-Type Glioblastoma

Objective:To investigate the value of MRI-based radiomics to non-invasively predict 0~6-methylguanine-DNA methyltransferase(MGMT)promoter methylation status in IDH1 wild-type glioblastoma,and to construct and evaluate the ability of different models to predict the risk of recurrence within one year in patients with IDH1 wild-type glioblastoma treated with temozolomide(TMZ).Methods:Data from 122 patients with IDH1 wild-type glioblastoma were retrospectively collected,and 3404 radiomics features were extracted from preoperative MRI.Feature dimension,feature selection and classifier were performed using multiple machine learning models randomly combined by Fe Ature Explorer(FAE),and then the best performing prediction model was selected.For MGMT promoter methylation prediction,single-layer radiomics signature based on single MR sequence and single ROI were constructed respectively,and then fusion MR radiomics signature were constructed based on significant single-layered radiomics signature.Select the features in the optimal hyperparameters to calculate the radiomics score(Rad-score)to construct a model based on multi-sequence MRI radiomics to predict the risk of glioblastoma recurrence.A total of six models were constructed using different feature screening methods and classifiers:a clinical-only model,a radiomics-only model,a two-fusion model of radiomics,clinical or gene,and a three-fusion model of radiomics,clinical and gene.Combining Rad-score with other factors using Cox proportional risk regression models to construct a nomogram which can predict patient’s risk of recurrence within one year.Results:1.In terms of MGMT promoter methylation status prediction,the fusion model based on the nine radiomics features achieved the best performance in validation set(AUC,0.77;95%CI:0.58-0.93),using the Kruskal-Wallis test as method screening features,least absolute shrinkage and selection operator(LASSO)as the classifier.2.The radiomics-gene fusion model had the best performance in predicting patients’risk of recurrence within one year,with an AUC value of0.895(95%CI:0.814-0.960)in validation set.Relief selected radiomics features and the classifier was logistic regression.The nomogram can individually predict the probability of patients’risk of recurrence within one year.Conclusion:Radiomics based on MRI and machine learning can noninvasively predict the MGMT promoter methylation status of IDH1 wild-type glioblastoma and the probability of early recurrence in glioblastoma patients who were treated with Temozolomide.