Abstract BACKGROUND Even for those undergoing gross total resection (GTR), patients with WHO grade 4 glioblastoma (GBM) have poor overall survival (OS). Stratifying those who will have better OS vs. those that will not remains challenging, with conventional imaging and radiomic models falling short. This study utilizes machine learning (ML)-generated radio-pathomic models based on autopsy tissue as ground-truth to enhance prognostication through tumor probability maps (TPM) within preoperative contrast-enhancing (CE) regions of GBM. METHODS Data from the UCSF-PDGM-v2 and UPenn-GBM databases were analyzed, focusing on patients who underwent GTR with WHO 4 IDH-wildtype glioblastoma (n = 528). Preoperative tumor cellularity within CE was quantified using ML-generated TPMs, with a median threshold of 0.7242. Kaplan-Meier survival curves were employed to compare OS between patients with low (≤ 0.7242) and high (> 0.7242) TPM values in CE regions. Statistical significance was evaluated using log-rank tests. RESULTS Kaplan-Meier curves demonstrated significant stratification in survival outcomes based on TPM values. Patients with lower TPM values in CE regions had a median OS of 567 days, compared to 457 days for those with higher TPM values (log-rank p-value: < 0.001). These findings suggest that lower tumor cellularity in CE regions as assessed by TPMs correlates with improved survival, offering a more nuanced stratification than current MRI-based methods. CONCLUSION This study underscores the potential of TPM metrics as a powerful prognostic tool in GBM. The significant correlation between lower TPM values in CE regions and better median OS indicates that TPM can effectively stratify patients, aiding in clinical decision-making and personalized treatment planning. Future research is necessary to validate these findings in larger, diverse patient populations and to explore the biological mechanisms underlying these survival differences.
Read full abstract