IMG-105. Prediction of 3-Month Mortality In Glioblastoma Using Interpretable Radiomics and Machine Learning

Le résumé fourni par la source

Abstract INTRODUCTION Anticipated survival time is a key factor in determining the appropriateness of surgical resection for glioblastoma (GBM). When life expectancy is <3 months, risks and recovery associated with surgery may outweigh benefits. Although preoperative risk stratification remains challenging, the incorporation of interpretable radiomic features (RFs) from GBM MRIs may enhance prediction of short-term survival. OBJECTIVE To develop an interpretable RF model for 3-month survival prediction in GBM. METHODS This study used the BraTS 2017 dataset (PMID:25494501), which includes T1CE, T2, and FLAIR images with annotations for tumor, peritumoral edema, and necrosis. RFs and heatmaps were extracted with PyRadiomics, and region-to-region RF ratios (tumor:edema, tumor:necrosis, edema:necrosis) were computed. Two-sample tests with FDR correction identified RFs significantly associated with 3-month survival, and a logistic regression (LR) model was optimized with stratified 3-fold cross-validation, hyperparameter tuning, and feature selection. RESULTS This study analyzed 112 GBMs, including eighteen with 3-month mortality. A total of 110 RFs differed significantly by survival, with enrichment in RF ratios highlighting the importance of region interfaces. An optimized logistic regression model achieved a mean AUC of 0.795 using eleven selected feature. Patients with 3-month mortality showed increased tumor:necrosis major axis length (q=0.005), suggesting a higher enhancing tumor-to-necrosis ratio predicts worse prognosis. Additionally, higher tumor textural complexity (q=0.029) and lower necrosis intensity (q=0.046) were observed in poor outcomes. CONCLUSIONS RFs capturing spatial and intensity relationships between tumor subregions can effectively predict 3-month survival in GBM. Upon validation in a larger dataset, the proposed model and RF visualizations may enhance prediction of short-term survival while facilitating more personalized treatment strategies.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.