ESTRO 2025 - Abstract Book

S3807

Physics - Radiomics, functional and biological imaging and outcome prediction

ESTRO 2025

Elasticnet achieves higher scores when including secondary metabolites, as shown in Table 1 (AUC=0.82 vs. 0.80, Sensitivity=0.81 vs. 0.77, p<0.05 over validation set). Among the six most important features (NAA, CE, PCh, FLAIR, Gln, and GSH (reduced form of glutathione)), four were extracted from MR spectroscopy, while two were secondary metabolites (Figure 1a). The model's prediction probability map demonstrates good spatial overlap with CE at relapse (Figure 1b). However, low F1-scores highlight challenges in balancing precision and recall, likely due to class imbalance, while score variability arises from GBM heterogeneity or differences in relapse timing.

Conclusion: The results are consistent with previous findings relating an increase in Choline and a decrease in NAA to relapse risk. In addition, including secondary metabolites improves performance. The importance of GSH, linked to glutamine, highlights its role in tumor survival, chemotherapy resistance, and relapse [4]. This study therefore opens avenues for uncovering specific metabolic changes involved in GBM relapse using in-vivo MRI and