MM-3DAttNet: Multi-Modal 3D Attention Network for MGMT Methylation Prediction

The methylation status of the$O^{6}$-methylguanine-DNA methyltransferase (MGMT) gene promoter is a well-established prognostic and predictive biomarker in glioma, particularly for assessing responsiveness to alkylating agents such as temozolomide. However, current radiogenomic approaches for MGMT status prediction are limited by reliance on invasive biopsy, 2D slice-wise modelling, suboptimal exploitation of multi-modal MRI, and weak model interpretability, all of which hinder robust clinical adoption. To address these gaps, we introduce a multi-modal 3D attention network (MM-3DAttNet) for non-invasive prediction of MGMT promoter methylation status from pre-operative multi-parametric brain MRI. The architecture comprises four modality-specific 3D CNN encoder branches for T1, T1ce, T2, and FLAIR, which are fused via a cross-modality attention mechanism to capture complementary diagnostic information.The model was trained and validated on the BraTS 2021 cohort, including 585 glioma cases annotated with MGMT methylation labels, and achieved an average accuracy of$91.6\%$, an F$_{1}$-score of$89.9\%$, and an area under the ROC curve of 0.925 across five-fold cross-validation. Model interpretability was enhanced through Grad-CAM-based saliency maps, which consistently highlighted clinically meaningful regions, including enhancing tumour margins and peritumoural oedema. A comprehensive ablation study confirmed the importance of both multi-modal integration and the attention-based fusion: removing individual modalities reduced performance relative to the full four-modality model, with the largest drops observed when T1ce or FLAIR was excluded, underscoring their dominant contribution to MGMT-related signal. Taken together, these results demonstrate the efficacy and robustness of MM-3DAttNet as an interpretable radiogenomic framework for MGMT methylation assessment and motivate its prospective evaluation in multi-centre studies and integration into MRI-based decision-support pipelines for glioma management.