MNGI-23. PREOPERATIVE QUANTITATIVE IMAGING FEATURES ARE PROGNOSTIC FOR MENINGIOMA OUTCOMES. (5th November 2018)

• Record Type: Journal Article
• Authors: Morin, Oliver; Chen, William; Gennatas, Efstathios; Magill, Stephen; Wu, Ashley; Valdes, Gilmer; Perry, Arie; Sneed, Penny; McDermott, Michael; Solberg, Timothy; Bush, Nancy Ann Oberheim; Braunstein, Steve; Villanueva-Meyer, Javier; Raleigh, David
• Is Part Of: Neuro-oncology. Volume 20(2018)Supplement 6
• Page Start: vi153
• Page End: vi154
• Publication Date: 2018-11-05
• Subjects: Brain Neoplasms -- Periodicals; Brain -- Tumors -- Periodicals; Brain -- Cancer -- Periodicals; Nervous system -- Cancer -- Periodicals; 616.99481
• DOI: 10.1093/neuonc/noy148.639
• Languages: English
• ISSNs: 1522-8517

Abstract

Abstract: OBJECTIVES: Quantitative radiologic and radiomic features can identify brain tumors at risk for poor outcomes. Here, we investigate prognostic models for meningioma grade, local failure (LF) and overall survival (OS) based on demographic, radiologic, radiomic and therapeutic data. METHODS: We developed a database that was enriched for high grade meningiomas from 219 patients who underwent surgery for 229 meningiomas from 1990 to 2015 who had comprehensive clinical, pathologic and radiologic information available for retrospective review. The median imaging follow up was 4.3 years, and there were 112 WHO grade I (49%), 93 grade II (41%) and 24 grade III (10%) meningiomas. Two neuro-radiologists independently annotated 17 radiologic features, and 154 radiomic features were extracted from preoperative post-contrast 3D SPGR MR images for each meningioma. Random forest models were trained using nested resampling, and the performance of each model was assessed by calculating feature importance, mean balanced accuracy (BA) and area under the curve (AUC). RESULTS: Models restricted to preoperative demographic information and quantitative imaging features had superior BA (0.60–0.67) and AUC (0.60–0.76) for LF or OS as compared to models based on meningioma grade and extent of resection (BA 0.65, AUC 0.64). Integrated models incorporating all available data provided the most accurate estimates of LF and OS (BA 0.67, AUC 0.76). Radiomic features alone or in combination with other variables showed moderate and marginal predictive value for grade (BA 0.63, AUC 0.72) and LF (BA 0.61, AUC 0.65), respectively. Among radiologic features, meningioma diffusion characteristics significantly strengthened prognostication of grade and LF (RR 25.6, P = 0.001). Recursive partitioning analysis identified tumor size, primary versus recurrent presentation, grade, sphericity, apparent diffusion coefficient, location, extent of resection and T2 signal as the most important features for LF. CONCLUSIONS: Models using clinical and quantitative imaging data can accurately predict meningioma outcomes.