NIMG-01. RADIOMIC DATA ANALYSIS WITH ENSEMBLE MACHINE LEARNING TECHNIQUES PREDICT GRADE, HISTOPATHOLOGIC SUBTYPE, AND 1P/19Q CO-DELETION STATUS IN PATIENTS WITH LOW-GRADE GLIOMAS

Abstract Low-grade gliomas (LGGs) are a diverse group of primary brain tumors characterized by their variable prognosis and treatment response. Traditionally, the management of LGGs has relied on histopathological analysis, which sometimes offers limited insight into tumor behavior and treatment outcomes. Recent advances in radiomics have enabled the extraction of detailed quantitative features from routine imaging, providing a non-invasive method to assess tumor characteristics. These radiomic features can capture underlying tumor heterogeneity more comprehensively than conventional methods. Integrating radiomic data with ensemble machine learning techniques offers a promising approach to enhance the prediction of tumor grade, histopathologic subtype, and genetic alterations such as 1p/19q co-deletion status in LGGs, potentially leading to more tailored therapeutic strategies. We combined radiomic-genomic data from 159 magnetic resonance imaging (MRI) scans from the updated LGG-1p19qDeletion dataset (Erickson 2020) and 114 MRI scans from the ReMIND dataset (Juvekar 2024). Expert radiologists performed segmentations. A total of 2446 extracted radiomic features were initially extracted. Feature selection was performed using wrapper- and filter-based techniques with cross-validation. Prediction models were built using machine learning techniques, including random forest and extreme gradient boosting. Using Random Forest, we achieved high accuracy in predicting histopathologic subtype (88.14%, AUC 96.68%) with precision of 87.86% and Dice score of 87.88%. For grade prediction (accuracy 76.19%, AUC 91.84%), precision was 73.91%, and the Dice score was 77.27%. For 1p/19q co-deletion status (accuracy 75.61%, AUC 82.74%), precision was 77.78%, and the Dice score was 73.68%. Our study establishes a robust radiomic workflow to predict tumor grade, histopathologic subtype, and 1p/19q co-deletion status in low-grade gliomas, enhancing diagnostic accuracy and treatment personalization. This methodology can be applied clinically to improve the stratification of treatment plans for low-grade gliomas, potentially leading to effective, personalized therapy options. Using multicentric data potentially improved external validity, something which past research lacked. Future research should focus on validating and deploying this model using observational multicentric study designs.