MRI-derived radiomic risk stratification for overall survival in glioma

MRI radiomics may capture tumor phenotypes associated with glioma aggressiveness, but many radiomic survival signatures lack independent validation. We developed a UCSF-PDGM-derived MRI radiomic score for overall survival stratification in glioma and externally tested the locked score in UPenn-GBM.

This retrospective study used UCSF-PDGM as the development/internal cohort and UPenn-GBM as the independent external validation cohort. The UCSF dataset included 494 patients. A locked radiomic score derived from T1 contrast-enhanced/T1GD and FLAIR features was applied directly to UPenn-GBM without feature reselection, coefficient refitting, or retraining. External validation used automated segmentations as the primary analysis and manual segmentations as a sensitivity analysis. Cox regression, Kaplan-Meier analysis, concordance-index and model comparisons, bootstrap resampling, segmentation correlation and agreement analyses, calibration, and decision curve analysis were performed.

In UCSF-PDGM, radiomic score was significantly associated with overall survival (HR approximately 1.66, p = 0.000176). In UPenn-GBM, the score was applied to 581 patients and was associated with worse overall survival (HR 1.63, 95% CI 1.30-2.05, p < 0.005; C-index 0.56). Kaplan-Meier analysis showed worse survival in the high-risk group (log-rank p < 0.001). In multivariable analysis, radiomic score remained independently associated with survival after adjustment for age, IDH mutation status, and extent of resection (adjusted HR 1.35, 95% CI 1.04-1.77, p = 0.03). Automated- and manual-derived scores were strongly correlated among overlapping cases (Pearson r = 0.84; Spearman ρ = 0.83) and showed strong absolute agreement (ICC = 0.837).

The locked UCSF-derived MRI radiomic score showed a reproducible association with overall survival in the independent UPenn cohort and remained independently associated with survival after clinical adjustment. However, standalone and incremental discrimination were modest. These findings support further study of the score as a potential noninvasive prognostic biomarker rather than a stand-alone clinical prediction tool.