RADIOMIC FEATURES FROM PHYSIOLOGICAL MRI SHOWS IMPROVED ACCURACY OVER STRUCTURAL MRI IN PREDICTING MGMT PROMOTER METHYLATION IN GLIOBLASTOMA

O6-methylguanine-DNA methyltransferase (MGMT) promoter methylation is widely used as an independent favorable prognostic factor in glioblastoma patients. However, the evaluation depends on the tissue obtained via surgery. The purpose of this study is to determine whether radiomic features of magnetic resonance imaging (MRI) can be used for predicting MGMT prompter methylation status, and to compare the accuracies of physiological and structural sequences in prediction. Pre-operative MRI scan was performed on 111 primary glioblastoma patients. MRI sequences include structural sequence (T1-weighted, T2-weighted, post-contrast T1-weighted and T2-weighted fluid attenuated inversion recovery [FLAIR]) and physiological sequence (dynamic susceptibility contrast-enhancement [DSC] and diffusion tensor imaging [DTI]). All images were co-registered to T2-weighted images. DTI and DSC were processed as previously. Mean diffusivity (MD), fractional anisotropy (FA), DTI-p and DTI-q were generated from DTI. The relative cerebral blood volume (rCBV), mean transit time (MTT) and relative cerebral blood flow (rCBF) maps were generated from the DSC. Radiomic features which describe the shape, margin, intensity histogram and texture of the tumor were extracted from each imaging modality. The least absolute shrinkage and selection operator (LASSO) regularization was used for feature selection. Support Vector Machines (SVM) was used for the prediction of MGMT prompter promoter methylation status, which was determined using pyrosequencing. A total of 15 features were selected from the structural imaging and 21 features were selected from the physiological imaging. An accuracy of 0.71 was achieved from the selected structural imaging features, while an accuracy of 0.79 was obtained from the selected physiological features. Our findings showed that radiomic features are useful in predicting MGMT promoter methylation status of glioblastoma using supervised machine learning schemes. Physiological imaging features may have the potential to make more accurate prediction than structural imaging features. Texture analysis may extract the most useful features for the prediction.