ESTRO 2025 - Abstract Book

S3806

Physics - Radiomics, functional and biological imaging and outcome prediction

ESTRO 2025

2965

Digital Poster Voxel-Based Analysis for Predicting Recurrence in Post-Operative Glioblastoma Using Magnetic Resonance Spectroscopy Imaging: Beyond the Cho/NAA Ratio Wafae Labriji 1 , Jean-Yves Tourneret 1 , Elizabeth Moyal Cohen-Jonathan 2,3 , Lotfi Chaari 1 , Soléakhéna Ken 2,3 1 Toulouse INP, University of Toulouse, IRIT, Toulouse, France. 2 Institut Universitaire du Cancer, Oncopole Claudius Regaud, Toulouse, France. 3 Centre de Recherches en Cancérologie de Toulouse, INSERM UMR1037, Equipe RADOPT, Toulouse, France Purpose/Objective: Notwithstanding an arsenal of treatment combination, glioblastoma (GBM) always relapse in 85% within the surgical margin, despite complete removal of the contrast-enhancing (CE) tumor [1]. There is a growing interest in understanding the tumoral metabolism underlying behind the relapse mechanisms. This study aims to predict high risk recurrence areas using Magnetic Resonance Spectroscopy Imaging (MRSI) alongside with anatomical Magnetic Resonance Images (MRI). Material/Methods: Post-operative and pre-radiotherapy anatomical MRI (T1w, T1w-c, T2w, FLAIR) and MRSI from 44 GBM patients treated in our institute with relapse ≤ 9 months were studied. Metabolite quantification was performed using a Bayesian method [2] with a sparsity constraint on concentration estimates enabling to study high-interest metabolites not conventionally investigated. A voxel was classified as relapse if it changed from non-enhanced to CE during the post-treatment follow-up. MRI volumes were downsampled to MRSI resolution, with MRSI voxels classified as recurrence-positive if ≥50% of corresponding MRI voxels were positive. Thirty percent of subjects were used for testing; the rest for training and validation using leaving-one out method. SMOTE [3] was used to address imbalanced classes. Elasticnet, Support Vector Machines (SVM), one-class SVM and Random Forests classification methods were performed on two sets: set #1 contains anatomical MRI features and quantification of main metabolites of interest conventionally studied and set #2 expands the number of quantified metabolites with less commonly studied and so-called secondary ones.

Results: