ASL perfusion radiomic heterogeneity is associated with early mortality in IDH-wildtype glioblastoma, beyond tumor volume: a time-varying effect with tumor-aware DTI-ALPS.

Background: Glioblastoma exhibits significant spatial heterogeneity, with perfusion variability that may reflect angiogenesis, hypoxia, and biological aggressiveness. Arterial spin labeling (ASL) offers non-contrast perfusion imaging, and radiomics quantifies tumor texture beyond basic ROI metrics. However, radiomic texture features can be confounded by tumor volume -- rarely tested directly. Diffusion tensor imaging along the perivascular space (DTI-ALPS) may capture perivascular/neurofluid dynamics, but its added prognostic value in glioblastoma is uncertain. Methods: We retrospectively included 322 patients with IDH-wildtype WHO grade 4 glioblastoma. ASL radiomic features (shape features excluded) were tested against tumor volume, clinical covariates, and DTI-ALPS in volume-adjusted Cox proportional hazards models. A machine-learning classification analysis of 12-month mortality (N=259; ALPS-valid subset N=100) compared clinical, volume-aware, and ASL-augmented models across six classifiers, with proportional-hazards assumptions formally tested. Results: Higher ASL heterogeneity was associated with mortality within 365 days of imaging, after adjustment for tumor volume and clinical variables (HR 1.39, 95% CI 1.07 -1.81, p=0.015), with proportional hazards confirmed within this window. This association persisted when the candidate feature pool was widened from the pre-specified 26-feature family to all 1,209 non-shape ASL features screened within training folds (HR 1.33, 95% CI 1.05 -1.69). The association was unchanged under flexible modeling of tumor volume (HR 1.40 -1.49 across spline, polynomial, and quantile-indicator specifications), showed no heterogeneity across volume strata (I²=0%), and persisted after orthogonalizing the score with respect to volume (HR 1.25 -1.27, all p<0.03). The association was time-varying -- strongest in the first 180 days (HR 1.37) and attenuating beyond one year. In contrast, the unadjusted association with overall survival did not remain significant after accounting for tumor volume, and machine-learning classification showed no incremental value from ASL radiomics beyond a volume-aware clinical baseline. DTI-ALPS did not improve classifier performance or show independent survival association. Conclusions: ASL perfusion heterogeneity was associated with early, but not overall, mortality in this cohort, with a modest component not explained by tumor volume -- a signal specific to the first year after imaging. Findings for overall survival and machine-learning classification were substantially attributable to tumor volume rather than radiomic texture, underscoring the importance of volume-adjusted testing in radiomics research. Tumor-aware DTI-ALPS provided no additional prognostic value. These findings are hypothesis-generating and require prospective, multi-center validation.