Evaluation of ResNet-50 and EfficientNet-B0 for MGMT Promoter Methylation Classification from Multi-Modal MRI
Information · 07.11.2026 · 17 (11) · 1116
DOI: 10.3390/info17111116 ↗Kurz gefasst
Fachartikel in der Zeitschrift Information, online erschienen am 07.11.2026. Verfasst von Nada Nofal und 2 weiteren Autor:innen. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0).
Abstract (Originalsprache)
Glioblastoma is an aggressive adult brain tumor in which O6-methylguanine-DNA methyltransferase (MGMT) promoter methylation is a clinically important biomarker for predicting response to temozolomide-based therapy. This study presents a comprehensive, reproducible, and multi-cohort deep learning framework for patient-level MGMT promoter methylation classification using multi-modal preoperative MRI. We systematically investigated two complementary architectures, ResNet-50 and EfficientNet-B0 with bidirectional long short-term memory (BiLSTM) aggregation, to assess the influence of imaging modalities, feature representation, and patient-level aggregation strategies. The study used the public RSNA–MICCAI Brain Tumor Radiogenomic Classification dataset, comprising 585 labeled patients (307 methylated and 278 unmethylated), of whom 582 were retained after comprehensive, quality-controlled preprocessing across all four MRI modalities (FLAIR, T1w, T1Gd, and T2w). The proposed evaluation framework incorporated physical-space rigid registration to a common T1Gd reference, foreground and brain masking, N4 bias-field correction, intensity normalization, and systematic quality control. To ensure robust, reproducible evaluation, we performed repeated nested patient-level cross-validation using three random seeds, three outer folds, two inner folds, and all 15 possible MRI modality combinations. In the primary nested evaluation, ResNet-50 achieved a mean ROC-AUC of 0.543 ± 0.028 and accuracy of 0.535 ± 0.022, while EfficientNet-B0 + BiLSTM achieved a mean ROC-AUC of 0.525 ± 0.036 and accuracy of 0.530 ± 0.031. In a complementary fixed-configuration out-of-fold analysis involving the same 582 patients, EfficientNet-B0 + BiLSTM achieved a ROC-AUC of 0.599 compared with 0.566 for ResNet-50. Paired DeLong testing (p=0.268) and exact McNemar testing (p=0.231) indicated that the observed differences were not statistically significant. A controlled four-arm backbone-by-aggregation experiment further systematically assessed architectural and sequence-aggregation effects. At the same time, calibration analysis showed lower Brier scores and calibration error for ResNet-50 in the secondary out-of-fold comparison. External evaluation across two independent cohorts, comprising 20 and 95 patients, respectively, extended the assessment beyond the development dataset and provided valuable evidence regarding cross-cohort model behavior and transportability. Collectively, this study establishes a reproducible and extensively evaluated radiogenomic benchmarking framework integrating anatomical alignment, multi-modal fusion, nested validation, statistical comparisons, calibration assessment, and independent external testing. The findings provide a methodological foundation for advancing reliable MRI-based MGMT promoter methylation prediction and guiding future developments in clinically oriented radiogenomic deep learning.
Zitieren
APA
Nofal, N., Nasr, M., & Defrawi, M. E. E. (2026). Evaluation of ResNet-50 and EfficientNet-B0 for MGMT Promoter Methylation Classification from Multi-Modal MRI. Information, 17(11), 1116. https://doi.org/10.3390/info17111116
Vancouver
Nofal N, Nasr M, Defrawi MEE. Evaluation of ResNet-50 and EfficientNet-B0 for MGMT Promoter Methylation Classification from Multi-Modal MRI. Information. 2026;17(11):1116. doi:10.3390/info17111116
BibTeX
@article{Nofal2026,
author = {Nofal, Nada and Nasr, Mona and Defrawi, Mai El El},
title = {Evaluation of ResNet-50 and EfficientNet-B0 for MGMT Promoter Methylation Classification from Multi-Modal MRI},
journal = {Information},
year = {2026},
volume = {17},
number = {11},
pages = {1116},
doi = {10.3390/info17111116}
}
EUVIMED