Multimodal MRI-Based Glioma Segmentation and MGMT Promoter Methylation Status Prediction Using Multitask Learning Architecture

Accurate identification of brain tumor regions (glioma segmentation) and prediction of O⁶-methylguanine-DNA methyltransferase (MGMT) promoter methylation are essential for guiding therapy in glioma patients. Typically, these tasks are conducted separately, limiting performance by neglecting the relationship between tumor localization and methylation status. To address this gap, we propose TAUM-Net, a multitask learning model that simultaneously performs glioma segmentation and MGMT promoter methylation prediction from MRI scans. TAUM-Net merges convolutional neural networks (CNNs), which capture local tumor details, with a Transformer architecture modeling global structural features. It uses two branches: one refines tumor boundaries, while the other aggregates multi-scale information to predict MGMT promoter methylation, both of which share a unified representation that optimizes the two tasks in tandem. Evaluations on the BraTS2021 and TCGA-GBM datasets demonstrate TAUM-Net's effectiveness, attaining a Dice score of 0.9210 for glioma segmentation and 63.48% accuracy for MGMT promoter methylation prediction. This performance underscores the value of multitask learning in leveraging shared features, maintaining high segmentation quality, and providing moderate predictive accuracy for methylation status. Although TAUM-Net's current accuracy does not yet replace standard clinical tests, it highlights the potential of integrated approaches for guiding diagnosis and treatment planning. Our code is freely available at https://github.com/smallboy-code/TAUM-Net.