Artificial intelligence pipeline predicts the integrated molecular-morphologic risk score of meningiomas from routine pre-operative MRI

Abstract Background Meningiomas are the most common primary intracranial neoplasms. A significant unmet need exists for non-invasive biomarkers to risk-stratify patients and guide decision-making. The integrated risk score (IRS), which combines histopathological grade, DNA methylation class, and copy number variation, is a powerful predictor of risk of progression. We aimed to predict the molecular-morphologic IRS from pre-operative T1-weighted contrast-enhanced MRI scans. Methods We retrospectively analysed two multi-institutional prospectively compiled datasets, including patients with histologically confirmed intracranial meningiomas that had matched DNA methylation and copy number variation profiles. We first developed a radiomics model and trained a support vector machine (SVM) classifier. We then developed a convolutional neural network based on ResNet101, and a vision transformer (ViT). Model performance was assessed using accuracy, area under the curve, precision, recall, and F1-scores. Results Seventy-six patients met the inclusion criteria. Distributions per WHO class and IRS were 53 (I), 22 (II), 1 (III), and 53 (low), 18 (intermediate) and 5 (high), respectively. The SVM model achieved an AUC of 80.4% and accuracy of 78.2%. The ResNet101 model achieved 85.3% accuracy. The ViT model outperformed, achieving 89.4% accuracy and 89.1% precision in predicting low versus intermediate/high IRS tumours. Conclusions This study represents the first application of the ViT architecture to predict, with high accuracy, the prognostically highly relevant IRS from routine pre-operative neuroimaging. This pipeline provides a biologically informed and clinically relevant model with potential for future use in early risk stratification and decision support in the management of patients with meningiomas.