2026 Proffered Presentations
S249: ARTIFICIAL INTELLIGENCE PREDICTION OF MENINGIOMA GRADE USING PREOPERATIVE MRI: A LARGE MULTICENTER STUDY
Syed M Adil1; Tanner J Zachem1; Pranav Warman2; Jihad Abdelgadir1; Nirav Patel1; Benjamin D Wissel1; Kyle M Walsh1; Christopher J Tralie3; Andreas M Rauschecker4; Patrick J Codd1; Ali Zomorodi1; Anoop Patel1; Allan Friedman1; Timothy W Dunn1; Evan Calabrese1; C. Rory Goodwin1; 1Duke University; 2Johns Hopkins University; 3Ursinus College; 4University of California San Fransisco
Introduction: Accurate preoperative determination of meningioma grade could substantially enhance surgical planning, yet no reliable noninvasive method currently exists. This study aims to create and externally validate a machine learning model that integrates conventional radiomics with novel fractal and topological metrics to classify WHO meningioma grade from preoperative MRI in a large, multi-institutional cohort with a true test set.
Methods: In this retrospective study, we utilized data from the BraTS Meningioma Challenge. A total of 699 patients from six academic centers were used for model development and internal validation. An additional 210 patients, inaccessible to the study team conducting model development and validation, served as an external test set, representing the largest study to date. Tumors were segmented into three regions and radiomic, fractal, and topologic features were extracted across four MRI sequences. A total of 2,889 extracted features served as the inputs for eight different machine learning architectures tuned via bayesian optimization to predict binarized WHO grade (1 vs 2/3). Area under the receiver operating curve (AUROC) was used as the primary metric.
Results: The final model selected is a support vector machine with 20 features that achieved an AUROC of 0.824 (95% CI: 0.755-833) on the completely unseen test data and demonstrated appropriate calibration (Figure 1). Feature analysis with SHAP values demonstrated that topological features provided additional information to conventional radiomic features alone. The final trained model is openly available for use.
Figure 1: Left: AUROC Cruve for Internal Validation (Grey) and Held-Out Test Set (Blue). Right: Calibration Curve
Conclusion: This externally validated machine learning model demonstrates robust performance in predicting meningioma grade from standard preoperative MRI. By combining traditional radiomics with novel topologic and fractal predictors, this approach provides a foundation for future imaging-driven surgical planning and tumor classification frameworks.