Problem

Glioblastoma (GBM) is a highly aggressive primary brain tumor with poor prognosis and limited treatment options. The methylation status of the O6-methylguanine-DNA methyltransferase (MGMT) gene promoter is a critical molecular biomarker that determines a patient's response to temozolomide chemotherapy. Current methods for assessing MGMT status rely on invasive surgical biopsies, which are associated with risks, are time-consuming, expensive, and prone to sampling errors due to the inherent heterogeneity of tumors. There is an urgent clinical need for non-invasive, accurate, and efficient alternatives to predict MGMT status to improve therapeutic decision-making.

Method

The study proposes a novel radiogenomic molecular imaging analysis framework for the non-invasive prediction of MGMT promoter methylation using multi-parametric magnetic resonance imaging (mpMRI). This approach integrates radiomics, deep learning, and explainable artificial intelligence (XAI). Radiomic features, including shape, first-order statistics, and texture descriptors, are extracted from FLAIR, T1-weighted, T1-contrast-enhanced, and T2-weighted MRI sequences. Simultaneously, a 3D convolutional neural network (EnhancedMGMTNet) learns deep representations from these same modalities, with an augmented residual U-Net (ResUNetVSA) used for accurate tumor segmentation. These complementary features are then fused using both early and attention-based strategies and classified to predict MGMT methylation status. To enhance clinical interpretability, XAI methods such as Grad-CAM and SHAP are applied to visualize and explain model decisions. The framework was trained on the RSNA-MICCAI Radiogenomic Classification dataset (BraTS 2021 cohort) and externally validated on the independent UCSD-PTGBM dataset.

Results

The proposed ResUNetVSA model demonstrated strong performance in tumor subregion segmentation on the BraTS 2021 dataset, achieving a macro Dice score of 0.912 and low 95th percentile Hausdorff Distances across enhancing tumor, tumor core, and whole tumor regions. For MGMT promoter methylation prediction, the attention-gated radiomics fusion model consistently outperformed both radiomics-only and deep-learning-only baselines across all metrics. On the independent UCSD-PTGBM external validation cohort (109 patients), the hybrid radiomics–deep fusion model achieved a robust ROC-AUC of 0.82 and an overall accuracy of 0.80. These results highlight the strong transferability and discriminative capability of the framework across different institutions and MRI acquisition protocols, without requiring any fine-tuning.

Implications

This work significantly advances the field of molecular imaging by demonstrating the potential of AI-driven radiogenomics for precision oncology. The framework offers a non-invasive, accurate, and interpretable method for predicting clinically actionable molecular biomarkers like MGMT methylation status in glioblastoma. This could potentially eliminate the need for risky and invasive surgical biopsies, reduce patient risk, and accelerate therapeutic decision-making, especially for inoperable tumors or in resource-limited settings. The hybrid approach, fusing deep representations with biologically meaningful radiomic features, enhances cross-site generalization and provides transparent explanations, fostering trust and collaboration between clinicians and AI systems. The conceptual framework is also generalizable to other diagnostic challenges, such as IDH mutation prediction, and diverse imaging contexts, positioning it as a foundational algorithm for interpretable biomedical imaging.