A multi-task deep learning framework incorporating tumor morphology for MRI-based MGMT methylation prediction in glioblastoma

Methylation of the O 6 -methylguanine-DNA methyltransferase (MGMT) promoter is a clinically important biomarker associated with prognosis and treatment response in glioblastoma. However, its assessment typically requires invasive tissue sampling. This study aimed to develop and validate a non-invasive, MRI-based deep learning model for MGMT methylation prediction, incorporating tumor morphological information through a multi-task learning (MTL) framework. We developed a deep learning model that jointly performs tumor segmentation and MGMT methylation classification. The model was trained on a publicly available cohort from The Cancer Imaging Archive (TCIA) and independently validated in an institutional cohort of patients with glioblastoma. Multiparametric MRI sequences, including contrast-enhanced T1-weighted and T2-FLAIR images, were used. Three input strategies were evaluated: image-only, image combined with tumor mask, and an MTL framework integrating segmentation and classification. Model performance was assessed using the area under the receiver operating characteristic curve (AUC). Among the evaluated convolutional neural network architectures, the ConvNeXt-V2-based MTL model demonstrated the best performance. It achieved an AUC of 0.862 in development-cohort cross-validation and 0.769 in independent institutional validation. The MTL approach incorporating tumor morphological information consistently outperformed image-only and image-plus-mask models, particularly in the independent institutional validation cohort, suggesting improved generalizability. An MRI-based multi-task deep learning model integrating tumor segmentation and classification shows promising performance for non-invasive MGMT methylation prediction in glioblastoma. Incorporation of morphological information through an MTL framework may enhance model robustness across datasets, particularly under limited data conditions. This approach may provide complementary information for non-invasive molecular profiling and support clinical decision-making in glioblastoma.