Improved Prediction of MGMT Methylation Status in Glioblastoma using a Deep Attention Network

Glioblastoma multiforme (GBM) is the most aggressive brain tumor with a survival rate of 18 months. Research shows that the status of MGMT gene determines the effectiveness of chemotherapy treatment in GBM patients. In this work, we propose an attention based deep learning network for predicting MGMT methylation status in GBM patients. Current techniques utilize radiomic analyses, but these methods may not extract features required for accurate prediction. Recently deep learning techniques have been proposed, but these methods require looking at the entire data to predict the methylation status. To account this, we built squeeze and sequence attention for prioritizing the slices and regions respectively. The proposed model was evaluated on several binary classification metrics, with the best AUC of 70.59. We demonstrate a robust and automatic method to capture important features from MRI images that performs substantially better on predicting methylation status of the brain compared to existing methods.