Hybrid Deep Learning Architecture for Non-Invasive MGMT Methylation Prediction in Glioblastoma Using Multi-Modal MRI

Glioblastoma is still the most invasive type of primary brain tumor in adults, accounting for highly elevated mortality rates, and median survival times hardly ever go above 15 months with current therapies. The methylation status of the MGMT promoter is a major biomarker for predicting the response of a patient to a temozolomide, based chemotherapy, thus it allows for the tailoring of treatment plans and providing precise prognosis. One of the ways to carry out MGMT tissue biopsy determination has been through surgery which comes with substantial risks for patients, therefore causing diagnostic delays and a sampling problem due to intratumoral heterogeneity. Hence, a novel hybrid architecture is proposed in this paper which combines different paradigms of representation learning such as: a two-dimensional slice analysis based on the Swin Transformer, a volumetric three-dimensional ResNet encoding, and traditional handcrafted radiomic descriptors, which are then combined through an attention-driven fusion mechanism. The technique uses advanced training methods such as mixup regularization, smooth label calibration, focal loss optimization, test, time augmentation, and stratified k-fold cross-validation. In total, 585 glioblastoma cases were used for a thorough experimental study that showed the way our combined multipath method achieves an overall accuracy of 87.16%, which is way more than the accuracies of standalone radiomics pipelines and deep learning architectures. The model showed remarkable performance stability in cross-validation, thus it demonstrated the robustness that is required for clinical usage. This multi-scale representation model successfully merges the visual abstractions learned and the domain-specific engineered features, thus it can be considered a promising clinical virtual assistant.