Glioma Grade and Molecular Markers: Comparing Machine-Learning Approaches Using VASARI (Visually AcceSAble Rembrandt Images) Radiological Assessment

Objectives: This study aimed to leverage Visually AcceSAble Rembrandt Images (VASARI) radiological features, extracted from magnetic resonance imaging (MRI) scans, and machine-learning techniques to predict glioma grade, isocitrate dehydrogenase (IDH) mutation status, and O6-methylguanine-DNA methyltransferase (MGMT) methylation. Methodology: A retrospective evaluation was undertaken, analyzing MRI and molecular data from 107 glioma patients treated at a tertiary hospital. Patients underwent MRI scans using established protocols and were evaluated based on VASARI criteria. Tissue samples were assessed for glioma grade and underwent molecular testing for IDH mutations and MGMT methylation. Four machine learning models, namely, Random Forest, Elastic-Net, multivariate adaptive regression spline (MARS), and eXtreme Gradient Boosting (XGBoost), were trained on 27 VASARI features using fivefold internal cross-validation. The models' predictive performances were assessed using the area under the curve (AUC), sensitivity, and specificity. Results: For glioma grade prediction, XGBoost exhibited the highest AUC (0.978), sensitivity (0.879), and specificity (0.964), with f6 (proportion of non-enhancing) and f12 (definition of enhancing margin) as the most important predictors. In predicting IDH mutation status, XGBoost achieved an AUC of 0.806, sensitivity of 0.364, and specificity of 0.880, with f1 (tumor location), f12, and f30 (perpendicular diameter to f29) as primary predictors. For MGMT methylation, XGBoost displayed an AUC of 0.580, sensitivity of 0.372, and specificity of 0.759, highlighting f29 (longest diameter) as the key predictor. Conclusions: This study underscores the robust potential of combining VASARI radiological features with machine learning models in predicting glioma grade, IDH mutation status, and MGMT methylation. The best and most balanced performance was achieved using the XGBoost model. While the prediction of glioma grade showed promising results, the sensitivity in discerning IDH mutations and MGMT methylation still leaves room for improvement. Follow-up studies with larger datasets and more advanced artificial intelligence techniques can further refine our understanding and management of gliomas.