IMGP-01. Radiomics Analysis in MRI Glioma for Histologic Grade and BRAF Mutation Classification: A Public and Robust Pipeline from Segmentation to Explainability

INTRODUCTION Low-Grade Gliomas (LGG) are the most common primary brain tumors in pediatric patients, presenting diagnostic and therapeutic challenges due to their heterogeneity. Radiomics and Machine Learning (ML) can support classification of histological grades, according to WHO-2021 criteria, and prediction of molecular alterations such as BRAF mutations. MATERIALS & METHODS A robust pipeline was developed and tested on both a public dataset (BRATS-2020) and a private dataset from Meyer Children’s Hospital IRCCS, including pediatric patients with known BRAF-V600E and KIAA1549 alterations, as well as BRAF wild type cases. T1w, T2w and FLAIR images were analyzed using Python 3.6 and Pyradiomics. Brainstem was extracted using Synthseg for normalization. To enhance robustness, each original ROI was artificially augmented to generate two additional ROIs. Feature selection was performed to retain only robust features. Random Forest and XGBoost models were trained using stratified 5-fold cross-validation. The pipeline was applied for both histologic grade and BRAF mutation classification. Explainability was achieved using LORE and SHAP algorithms. RESULTS On the BRATS-2020 (n=369) test set, the pipeline achieved a mean accuracy of 0.91 and an AUC of 0.93. On the private dataset from Meyer Children’s Hospital IRCCS (n=46; 14 BRAF-V600E, 11 KIAA1549, 21 BRAF-wildtype), the pipeline achieved an accuracy of 0.76 and an AUC of 0.75 for KIAA1549 prediction, for BRAF-V600E an accuracy of 0.63 and an AUC of 0.66. SUMMARY This study presents a reproducible and interpretable radiomics pipeline validated on both public and private pediatric datasets. The integration of voxel normalization, multi-ROI enhancement, and explainability tools supports accurate LGG subtyping and BRAF mutation status prediction. The model’s consistent performance across datasets underscores its potential generalizability in clinical applications. All code used in the pipeline is available on GitHub to facilitate accessible and reproducible radiomic analysis of MRI images.