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

The methylation status of the O⁶-methylguanine-DNA methyltransferase (MGMT) promoter is an established prognostic and predictive biomarker in glioma, particularly for estimating response to alkylating chemotherapy such as temozolomide. However, many existing radiogenomic methods remain constrained by invasive biopsy dependence, slice-wise 2D modelling, limited use of multi-modal MRI, and insufficient interpretability, which collectively impede clinical translation. We propose MM-3DAttNet, a multi-modal 3D attention network for noninvasive prediction of MGMT promoter methylation status from pre-operative multiparametric brain MRI. The model employs four modality-specific 3D CNN encoder branches (T1, T1ce, T2, and FLAIR) and integrates them using a cross-modality attention fusion module to capture complementary diagnostic cues. MM-3DAttNet was trained and evaluated on the BraTS 2021 cohort comprising 585 glioma cases with MGMT labels, achieving an average accuracy of 91.6%, ₁-score of 89.9%, and AUC of 0.925 under five-fold cross-validation. Interpretability was supported using Grad-CAM saliency maps, which consistently emphasized clinically relevant regions such as enhancing tumour boundaries and peritumoural oedema. Ablation experiments verified the importance of multi-modal learning and attention-based fusion, with the most pronounced performance reductions observed when excluding T1ce or FLAIR. Overall, MM-3DAttNet provides an accurate and interpretable radiogenomic framework for MGMT methylation assessment and supports future validation in multi-centre settings and integration into MRI-based decision-support workflows for glioma management.