Unimodal vs. multimodal deep learning for non-invasive MGMT promoter methylation prediction in glioblastoma: A systematic evaluation on the BraTS 2021 dataset.
Authors
Affiliations (5)
Affiliations (5)
- BIGR, Inserm, U1134, University Paris City, Paris, France.
- Faculty of Sciences and Technology, DSIMB, BIGR, Inserm, U1134, University of Reunion, Saint-Denis, France.
- Department of Computer Science, University of Management and Technology, Lahore, Pakistan.
- Faculty of Sciences and Technology, ENERGYLab, University of Reunion, Saint-Denis, France.
- PEACCEL, AI for Biologics, Paris, France.
Abstract
Glioblastoma multiforme (GBM) is the most aggressive primary brain tumor in adults, with a median survival of 14.6 months under standard radiotherapy and temozolomide (TMZ) chemotherapy. The methylation status of the O⁶-methylguanine-DNA methyltransferase (MGMT) promoter is a critical biomarker predicting TMZ response; however, its determination currently requires invasive tissue sampling. Non-invasive prediction of MGMT promoter methylation from multiparametric MRI (mpMRI) through deep learning represents a compelling alternative, yet its clinical feasibility remains unresolved. Using the BraTS 2021 dataset (582 patients, four MRI sequences: FLAIR, T1w, T1wCE, T2w), we conducted a systematic comparative study of unimodal and multimodal deep learning approaches based on VGG-16, exploring 1,380 experimental configurations (unimodal: 192; multimodal: 1,188) across three imaging planes, eight slice counts, and three multimodal fusion strategies (early, intermediate, and late fusion). In the unimodal setting, the best model trained on T2w coronal images (32 slices, no transfer learning) achieved an accuracy of 0.6458 and an AUC of 0.6422 on the validation set, but dropped to 0.5586 and 0.5533 on the independent test set, revealing substantial overfitting attributable to limited dataset size. Strikingly, multimodal fusion consistently failed to outperform the best unimodal model, with all three fusion strategies plateauing at ~0.64 accuracy and ~0.64 AUC on validation data. Transfer learning improved generalization across train/test distributions at the cost of peak performance. These findings suggest, for the tested framework in this study, that MGMT methylation status prediction from mpMRI remains fundamentally constrained by dataset heterogeneity and size, irrespective of modality combination strategy, and that T2w coronal acquisitions could be more interesting in future data collection efforts.