When extended metabolites profiling helps the prediction of local relapse in glioblastoma: a retrospective study using MRI and MRSI

OBJECTIVES: This study investigates whether integrating postoperative MRI with advanced MR spectroscopic imaging (MRSI) improves voxel-wise prediction of future local recurrence in glioblastoma (GBM). METHODS: Fifty-eight GBM patients included in the clinical trial (XXXXXXXXXXX) were retrospectively analyzed with post-operative baseline for MRI and MRSI processed to generate voxel-level features. MRSI metabolites were quantified using a Bayesian sparse model, enabling reliable estimation of both conventional and extended metabolite panels. Ground-truth relapse labels were derived from longitudinal follow-up MRI to track voxel-wise tissue class changes over time. Voxel-wise relapse prediction was performed using three supervised classifiers (Elastic Net, Random Forest, RBF-SVM) evaluated with leave-one-patient-out cross-validation. Feature sets combining MRI intensities, spatial descriptors, and MRSI metabolites were systematically compared. RESULTS: MRI-only models showed moderate discrimination (AUC ≤ 0.70). Using MRSI features alone significantly improved prediction performance, with further improvements when extending the metabolite basis beyond conventional markers (AUC 0.77 to 0.82, p < 0.001). Spatial descriptors and MRSI-based models provided the largest performance gains (AUC up to 0.95). SHAP analysis consistently identified both conventional and secondary metabolites among the most influential predictors, particularly in the absence of spatial features. CONCLUSIONS: Postoperative MRSI and spatial descriptors significantly enhances voxel-wise prediction of GBM relapse, particularly when using an extended metabolite basis beyond conventional markers.