Radiogenomic Prediction of MGMT Using Deep Learning with Bayesian Optimized Hyperparameters

Glioblastoma (GBM) is the most aggressive primary brain tumor. The standard radiotherapeutic treatment for newly diagnosed GBM patients is Temozolomide (TMZ). O6-methylguanine-DNA-methyltransferase (MGMT) gene methylation status is a genetic biomarker for patient response to the treatment and is associated with a longer survival time. The standard method of assessing genetic alternation is surgical resection which is invasive and time-consuming. Recently, imaging genomics has shown the potential to associate imaging phenotype with genetic alternation. Imaging genomics provides an opportunity for noninvasive assessment of treatment response. Accordingly, we propose a convolutional neural network (CNN) framework with Bayesian optimized hyperparameters for the prediction of MGMT status from multimodal magnetic resonance imaging (mMRI). The goal of the proposed method is to predict the MGMT status noninvasively. Using the RSNA-MICCAI dataset, the proposed framework achieves an area under the curve (AUC) of 0.718 and 0.477 for validation and testing phase, respectively.