GlioVision: A multimodal MRI framework for non-invasive prediction of glioma molecular biomarkers.
Authors
Affiliations (4)
Affiliations (4)
- Department of Health Data Science and Artificial Intelligence, D. Bradley McWilliams School of Biomedical Informatics, The University of Texas Health Science Center, Houston, TX, 77030, USA. Electronic address: [email protected].
- Department of Health Data Science and Artificial Intelligence, D. Bradley McWilliams School of Biomedical Informatics, The University of Texas Health Science Center, Houston, TX, 77030, USA.
- Department of Neurosurgery, The University of Texas Health Science Center, Houston, TX, 77030, USA.
- Department of Neurosurgery, Baylor College of Medicine, Houston, TX 77030, USA.
Abstract
Gliomas are primary brain tumors for which molecular biomarker assessment plays an important role in clinical evaluation and treatment planning. Traditional assessment relies on analysis of surgical tumor specimens, which carries procedural risks and sampling bias due to tumor heterogeneity. Existing deep learning methods can be computationally demanding and are often trained on narrowly represented datasets, which may limit their clinical deployment. In this study, we propose GlioVision, a multi-modal deep learning framework for non-invasive prediction of glioma molecular biomarkers using multi-parametric MRI, including T1, T1-Gd, T2, and FLAIR sequences. The framework employs a 3D DenseNet-based architecture with channel and spatial feature recalibration, combined with an adaptive processing module that handles variability across multi-institutional datasets. In addition, a confidence-based filtering mechanism (CFPM) is introduced to identify low-confidence predictions and quantify prediction uncertainty. The model is trained and evaluated on a multi-cohort dataset compiled from publicly available sources, comprising 2270 subjects with heterogeneous imaging protocols. Subject-level splitting is used to prevent data leakage, and performance is reported using AUC, accuracy, sensitivity, and specificity across all biomarkers. Independent public cohorts were used for testing. GlioVision achieves AUC values of 0.94 for IDH, 0.87 for 1p/19q, 0.86 for MGMT, and 0.92 for tumor grade, indicating consistent performance across biomarkers. The effect of confidence-based filtering is also analyzed to examine the trade-off between prediction coverage and accuracy. Finally, we investigate a simple model obfuscation approach based on controlled image perturbations and evaluate its impact on prediction performance. This study presents a multi-institutional evaluation of a deep learning pipeline for predicting glioma biomarkers from MRI. It may support further research in radiogenomic analysis and non-invasive tumor characterization.