Abstract BACKGROUND Glioblastoma (GBM) remains the most devastating central nervous system malignancy with poor overall survival (OS), even with gross total resection (GTR) and chemoradiation. Identifying factors associated with long-term survival (LTS) in GBM patients is challenging but crucial for improving prognostic models. This study compares an internal dataset of the top 10% long-term survivors of GBM (n=14) against a combined external cohort of 528 GBM patients who underwent GTR. METHODS We used a previously published machine learning-generated radio-pathomic model based on autopsy tissue as ground-truth to generate whole-brain maps of cell density and tumor probability maps (TPM) using conventional preoperative MRI sequences in both our internal dataset and a combined external cohort of GBM patients who underwent GTR from the UCSF-PDGM-v2 and UPenn-GBM databases. We then analyzed clinical and radio-pathomic metrics, including cell density and TPM values within contrast-enhancing (CE) and FLAIR hyperintense regions. Statistical comparisons were performed using t-tests. RESULTS Cell density metrics showed significant differences, with lower cell density in contrast-enhancing regions (t-value: 4.709, p<0.001; median: 1559 vs. 1959) and in FLAIR hyperintense regions (t-value: 3.585, p<0.001; median: 1356 vs. 1568) within the LTS group. TPM metrics also showed significant differences, with lower TPM values in contrast-enhancing regions (t-value: 3.312, p<0.001; median: 0.71 vs. 0.72) and in non-enhancing regions (t-value: 2.606, p<0.01; median: 0.68 vs. 0.69) in the LTS group. CONCLUSION Our comparative analysis reveals distinct radio-pathomic characteristics associated with long-term survival in GBM patients. Reduced cell density and lower TPM values in both contrast-enhancing and FLAIR hyperintense regions are significantly correlated with long-term survival. These findings suggest that integrating radio-pathomic data may enhance prognostic models and warrants further studies evaluating the relationship of these markers and OS.
Read full abstract