The possibility to perform virtual biopsies through mpMRI is pivotal especially for invasive procedures such as brain biopsies. The assessment of the methylation status of MGMT promoter in patients affected by glioblastoma is crucial to determine therapy and prognosis. We present our preliminary results about the prediction of the MGMT promoter methylation performed on a publicly available dataset, UPENN-GBM. The dataset contains pre-processed T1, TIGd, T2, T2-FLAIR and DTI images, segmentations of different areas of the tumor (necrosis, enhancing tumor and edema) for a subset of patients and the information on MGMT promoter methylation status. As the number of available sequences is quite high, finding a trade-off between images to be used and computational burden is strictly needed. After extracting radiomic features on brain-stem normalized images, we trained and evaluated in a 5-fold stratified cross-validation fashion:

• a Random Forest (RF) and a Multi-Layer Perceptron (MLP) only with radiomic features extracted from the different tumor areas;
• a 3D Convolutional Neural Network (CNN) with only Fractional Anisotropy (FA) and Axial Diffusivity (AD);
• a multi-branch 3D CNN merging radiomic features taken from structural MRI images, FA and AD images;

The best result is obtained by the multi-branch 3D CNN with an accuracy (AUC) over the folds of 0.61±0.06 (0.63±0.09). As only a small part of the dataset has been explored, we plan to continue to search for a possible correlation between MRI signal and MGMT promoter methylation status since very contrasting results can be found in the literature.