455 Radiographic Features Predict Genomic Subtype in Meningioma

Abstract

INTRODUCTION: Meningiomas are the most common primary intracranial neoplasms. While most meningiomas demonstrate benign behavior with low rates of recurrence, a large percentage are more aggressive and are currently most widely prognosticated by histological grade. Several groups have published classification systems grouping meningiomas into distinct subtypes based on transcriptomic, cytogenetic, and methylomic signatures that better characterize their activity compared with traditional grading. Methods to preoperatively predict genomic subtype are desired. METHODS: Resected tumor samples that were genomically classified through a protocol described in previous publications were included in this analysis. Corresponding preoperative MRIs were gathered. The tumors were characterized using a standardized sheet of semantic variables. Chi-square and Wilcoxon rank-sum tests followed by univariable generalized linear regression were used for feature selection. Random forest algorithms were trained to classify genomic subtype based on radiographic features alone and tested using an 80/20 pseudorandom split. RESULTS: One-hundred tumors were included (n = 69 MenG-A/B, 31 MenG-C). Males were significantly more likely to develop MenG-C tumors than females (OR 3.44 [1.43, 8.52] p = 0.006). MenG-C tumors had significantly lower recurrence-free survival than MenG-A/B tumors (log-rank p value <0.01). Supratentorial status, parafalcine location, larger volume, moderate-severe edema, irregular enhancement, hypointensity on T1 non-contrast sequence, and indistinct tumor margins were significant predictors of more aggressive genomic class. A random forest model based on these features blindly predicted MenG-A/B class with 85% accuracy (sensitivity 83%, specificity 100%). Tumor volume, location grouping, and T1 hypointensity were the most important features in accurate classification. CONCLUSIONS: Aggressive genomic status in meningioma can be predicted on routine preoperative imaging. These results may aid in preoperative counseling and treatment planning and should be validated across multi-institutional datasets.