Multi-View MRI Approach for Classification of MGMT Methylation in Glioblastoma Patients

The presence of MGMT promoter methylation significantly affects how well chemotherapy works for patients with Glioblastoma Multiforme (GBM). Currently, confirmation of MGMT promoter methylation relies on invasive brain tumor tissue biopsies. In this study, we explore radiogenomics techniques, a promising approach in precision medicine, to identify genetic markers from medical images. Using MRI scans and deep learning models, we propose a new multi-view approach that considers spatial relationships between MRI views to detect MGMT methylation status. Importantly, our method extracts information from all three views without using a complicated 3D deep learning model, avoiding issues associated with high parameter count, slow convergence, and substantial memory demands. We also introduce a new technique for tumor slice extraction and show its superiority over existing methods based on multiple evaluation metrics. By comparing our approach to state-of-the-art models, we demonstrate the efficacy of our method. Furthermore, we share a reproducible pipeline of published models, encouraging transparency and the development of robust diagnostic tools. Our study highlights the potential of non-invasive methods for identifying MGMT promoter methylation and contributes to advancing precision medicine in GBM treatment.

💡 Research Summary

This paper presents a novel deep learning framework for the non-invasive classification of O6-methylguanine–DNA methyltransferase (MGMT) promoter methylation status in patients with Glioblastoma Multiforme (GBM), a critical biomarker for predicting response to temozolomide chemotherapy. The work addresses the limitations of current 3D deep learning models, which are computationally expensive and memory-intensive, by proposing an efficient “multi-view” 2.5D approach.

The core methodology involves four key stages. First, a standardized pre-processing pipeline is applied to the heterogeneous BraTS 2021 dataset (583 patients), involving format conversion, reorientation, and registration to a common brain atlas. T2-weighted images (T2wi) are used as the primary input modality due to their effectiveness in highlighting tumor extent. Second, tumor subregions (edema, necrosis, enhancing tumor) are automatically segmented using a pre-trained nnU-Net model. Third, a novel slice selection technique is introduced: instead of using all slices or the slice with the largest tumor area, the authors select a single, most representative slice from each of the three anatomical planes (axial, sagittal, coronal) based on the maximum Feret diameter of the tumor in that slice. This method proves superior in identifying slices that best represent the tumor’s spatial extent. Fourth, these three selected 2D slices are processed through three separate pre-trained MONAI DenseNet-121 feature extractor branches. The resulting feature vectors from each view are then concatenated and fed into a fully connected classifier network to predict the binary MGMT methylation status.

The proposed multi-view model achieves an AUC of 0.662 on the test set, outperforming not only single-view models using only axial (AUC 0.556), sagittal (0.568), or coronal (0.553) slices but also a 3D ResNet model using the entire volume (AUC 0.551). This demonstrates that the multi-view strategy effectively captures complementary spatial information across planes while significantly reducing computational complexity compared to full 3D models.

A significant additional contribution of this work is the authors’ commitment to reproducibility and benchmarking. They re-implemented several key state-of-the-art methods for MGMT prediction (e.g., Korfiatis et al., Han & Kamdar) within the same pre-processing pipeline and have publicly released the entire codebase, including executable notebooks. This provides a transparent and accessible benchmark for future research in this challenging domain.

In summary, this study introduces an efficient and effective multi-view 2.5D deep learning model for non-invasive MGMT status prediction in GBM, balancing performance and computational cost. By also providing a reproducible pipeline for existing approaches, it advances the field towards more robust, transparent, and clinically applicable diagnostic tools in neuro-oncology and precision medicine.