IMGP-06. Radiomic Risk Stratification for Pediatric High-Grade Glioma

Abstract Pediatric high-grade gliomas (pHGGs) are highly aggressive brain tumors responsible for significant morbidity and mortality in children. Current diagnostic approaches rely on invasive neurosurgery to obtain tissue for molecular and histopathological testing—an often high-risk procedure in cases where the tumor is inoperable or located near critical brain structures. Consequently, there is a pressing need for accurate non-invasive methods to predict tumor subtype, genetic mutations, and prognosis to guide clinical decision-making. In this study we have developed and validated a fully automated, MRI-based workflow that can accurately segment pHGG tumor sub-compartments and extract radiomic features predictive of clinical outcomes and genetic markers. The project has focused on two central aims: (1) to train and refine a deep learning-based segmentation algorithm, leveraging state-of-the-art architectures including nnUNet and MedNeXt, using retrospective clinical MRI datasets and transfer learning; and (2) to extract radiomic features from these automated segmentations to build machine learning models capable of predicting one-year progression-free survival, overall survival, and key genetic mutations (e.g., H3K27M, IDH1, p53). Data collection and analysis occurred in two phases. Retrospectively, pHGG cases were used to train and test the segmentation algorithm and radiomics-based predictive models. Our preliminary results in a cohort of 46 children with pHGG demonstrate strong performance of the radiomic model in predicting progression free survival (PFS), with a c-index of 0.9 in the test set, using a 70:30 split (training:testing). Statistical analyses will include standard segmentation quality metrics (Dice scores, Hausdorff distance), and machine learning performance measures (accuracy, AUC-ROC), as well as Kaplan-Meier survival curves and hazard ratios for outcome prediction. This integrated automated segmentation-radiomics framework could serve as a transformative clinical tool for non-invasive risk stratification, enabling more personalized treatment strategies for pediatric patients with high-grade gliomas and opening avenues for broader implementation in neuro-oncology.