2025 Poster Presentations
P137: DISTINCT RADIOGRAPHIC PHENOTYPE OF LOW-GRADE MENINGIOMAS
Abraham Dada; Aymen Kabir; Rithvik Ramesh, BS; Daniel Quintana, BA; Christian Jimenez, BS; Wesley Shoap, MD; Ezequiel Goldschmidt, MD, PhD; University of California, San Francisco
Introduction: The WHO classification system for meningiomas, despite frequent challenges, remains a robust predictor of tumor recurrence. Accurate pre-operative identification of Grade 1 meningiomas is crucial for informed management decisions and effective patient counseling. This study aimed to establish clinically sensitive and specific radiomic markers for low-grade meningiomas.
Method: We retrospectively reviewed patients with newly diagnosed WHO 1-3 meningioma from a single center between 2021 and 2024. Radiographic features were manually curated from CT and MRI sequences, including location, size (maximal tumor dimension), T1 enhancement pattern, T2 signal change, bone involvement, presence of osteolysis, and hyperostosis type (none, Type 1, Type II). Type 1 hyperostosis was defined as hyperostosis with destruction of cortical architecture, while Type 2 referred to the preservation of cortical structure. Univariate analyses and multivariate logistic regression analyses compared differences between low- and high-grade meningiomas. Three machine learning models—Random Forest, Support Vector Machine, and Gradient Boost Machine—were employed to assess nonparametric relationships between features.
Results: 230 low-grade meningiomas were compared to 91 high-grade meningiomas. Compared to high-grade meningiomas, low-grade meningiomas were more likely in females (73.9% vs. 51.6% p<0.001), located at the skull base (68.3% vs. 46.2% p<0.001), had smaller size (3.43±1.50 vs 4.5±1.85 p<0.001), were homogenous enhancing (78.3% vs 53.8% p<0.001), and exhibited T2-signal change (42.2% vs 74.7% p<0.001). Multivariate logistic regression analysis showed low-grade meningiomas were significantly associated with skull base location (OR: 7.63 P<0.011), smaller size (OR 0.76 p=0.004), and absence of T2 signal change (OR:0.46, P=0.02). Among machine learning models, the RF model achieved the highest performance on the test set (AUC:0.96, Accuracy: 0.88) compared to the GBM (AUC:0.88, Accuracy: 0.84) and SVM (AUC:0.71, Accuracy: 0.71). The top features of the RF model included tumor size, enhancement pattern, and T2 enhancement. In subgroup analysis 92.9% of meningiomas located at the skull base with no signal change and homogenous enhancement were low-grade. This phenotype was present in 40% of grade I meningiomas.
Conclusion: In addition to the known skull base location and smaller size, low-grade meningioma pathology is associated with the absence of T2 signal change and homogenous enhancement pattern. The presence of all these features is highly predictive of a low-grade meningioma and can help making decisions regarding surgical removal, counseling and follow-up.