Radiomics Outperforms Conventional Tumor Volume in Prognostic Stratification Across Independent Glioma Cohorts

Tumor volume is routinely assessed on MRI and is often used as a surrogate measure of glioma burden. However, conventional volume measurements may not capture the spatial heterogeneity and phenotypic complexity visible on routine MRI. To address this potential gap in clinical assessment, we compared the prognostic performance of segmentation-derived tumor volume and that of an MRI-derived radiomic risk score across independent glioma cohorts.

This retrospective multi-cohort study included four publicly available glioma MRI cohorts comprising 1,447 patients: MU-Glioma-Post (n = 194), University of California San Diego post-treatment glioblastoma (UCSD-PTGBM; n = 178), University of California San Francisco preoperative diffuse glioma MRI (UCSF-PDGM; n = 494), and the University of Pennsylvania glioblastoma (UPenn-GBM; n = 581). MU-Glioma-Post served as the discovery cohort for post-treatment volumetric progression-free survival (PFS) analysis. UCSD-PTGBM, UCSF-PDGM, and UPenn-GBM were used for external volume assessment, radiomic validation, and head-to-head radiomics-versus-volume testing.

In the MU discovery cohort, 194 patients were included, and 152 experienced progression. Median PFS/follow-up was 159 days (IQR 77-283). Active tumor volume was independently associated with shorter PFS after adjustment for age, grade, isocitrate dehydrogenase status, and O6-methylguanine-DNA methyltransferase (MGMT) methylation (HR 1.045 per 10,000 mm³, 95% CI 1.015-1.076, p = 0.003; C-index 0.676). In external volume analyses, UCSD enhancing volume was not significantly associated with PFS (HR 1.144, 95% CI 0.935-1.398, p = 0.191), while UCSF fluid-attenuated inversion recovery volume was significant on univariable analysis (HR 0.917, 95% CI 0.875-0.961, p = 0.00031) but not after clinical adjustment (HR 1.034, 95% CI 0.983-1.087, p = 0.196). In UPenn-GBM, automated tumor volume was not associated with overall survival (OS) on univariable analysis (HR 0.993, 95% CI 0.978-1.009, p = 0.409) or adjusted analysis (HR 0.998, 95% CI 0.982-1.014, p = 0.768). In contrast, the radiomic score was associated with worse OS in UPenn (HR 1.634, 95% CI 1.302-2.050, p < 0.001). In direct head-to-head modeling, the radiomic score remained significant while tumor volume did not, both unadjusted (radiomics HR 1.626, p < 0.001; volume HR 0.996, p = 0.655) and adjusted for age, isocitrate dehydrogenase 1, and MGMT (radiomics HR 1.580, 95% CI 1.246-2.005, p < 0.001; volume HR 0.998, p = 0.806).

Conventional tumor volume showed inconsistent prognostic performance across external glioma cohorts and failed to retain independent significance in UPenn head-to-head models. MRI radiomics demonstrated greater prognostic reproducibility and remained independently associated with OS beyond that explained by tumor volume. These findings suggest that radiomic phenotyping captures clinically relevant imaging information beyond simple tumor burden measurements.