K-fold ensemble 3D convolutional neural network for predicting MGMT promoter methylation in glioblastoma

Abstract

Medical intervention is necessary for brain tumors, which represent a critical health threat. Chemotherapy response and patient survival outcomes depend on the methylation status of the O-methylguanine-DNA methyltransferase (MGMT) promoter. Biopsy and laboratory testing currently provide the only method to obtain this specific information. This study investigates a non invasive technique for measuring MGMT promoter methylation through magnetic resonance imaging (MRI) scanning. The BraTS 2021 dataset provided fluid attenuated inversion recovery (FLAIR) and contrast-enhanced T1 (T1ce) MRI data to develop a 3D convolutional neural network (CNN) system. The model used five-fold stratified cross-validation for training and testing to create a reliable assessment method. Prediction accuracy improved through the use of an ensemble that combined the best models from each cross-validation fold. The model achieved an average accuracy of 0.718 and an area under the curve (AUC) of 0.727 on the validation data. The results demonstrate that MRI features can provide essential molecular details despite using restricted imaging techniques. The proposed framework shows that deep learning enables early non-invasive detection of MGMT promoter methylation status in glioblastoma (GBM). The methods help doctors with treatment planning while also identifying patients who will benefit from temozolomide-based therapies.