Overall survival prediction in glioblastoma patients using multimodal MRI and radiomic feature extraction

Details

Supervisors

Macq, Benoît

Faculty

Ecole polytechnique de Louvain

Degree label

Master [120] : ingénieur civil biomédical, à finalité spécialisée

Abstract

enGlioblastoma (GBM) is the most common and most aggressive primary brain tumour in adults, with a median survival of approximately 15 months. As the most appropriate treatment differs from one patient to another, the ability to predict survival is valuable for tailoring patient management. Magnetic resonance imaging (MRI) is widely recognised as a powerful, non-invasive tool for visualising tumour structures within the brain, and radiomics offers a way to extract quantitative information from these images. This work aims to model overall survival (OS) in GBM patients and to identify the imaging biomarkers that carry the prognostic signal. A cohort of 234 pre-operative MRI examinations from the BraTS2020 dataset was preprocessed, from which 10,203 radiomic features were extracted per patient across three tumour sub-regions, together with the patient’s age at diagnosis. This high-dimensional set was reduced to a consensus of nine interpretable features through a selection procedure combining three complementary methods. Eight survival models were then trained and compared, in cross-validation and on a held-out test set kept strictly isolated throughout, and the one best suited to a clinical context was retained. The Random Survival Forest was retained as the best overall model, reaching a test-set concordance index of 0.671, a mean time-dependent AUC of 0.739 and an integrated Brier score of 0.193, performance on a par with the state of the art for conventional-MRI radiomics. It also separated patients into two significantly distinct risk groups (log-rank p = 1.43e-2, a 192-day difference in restricted mean survival time), with the retained signature concentrating in the tumour core. These results show that a model built solely on conventional multimodal MRI can stratify GBM patients by risk at a level competitive with the literature, while preserving the biological interpretability required for clinical translation.