Diffusion and Perfusion Heterogeneity for Survival Stratification in Post-Treatment Glioblastoma

Purpose: The prognostic value of diffusion- and perfusion-derived tumor-mask heterogeneity for overall survival in post-treatment glioblastoma was evaluated using a public MRI dataset. Materials and Methods: The University of California San Diego Post-Treatment Glioblastoma (UCSD-PTGBM) dataset was used to construct a first-timepoint cohort of 133 subjects. Twenty tumor-mask features were extracted from high b-value apparent diffusion coefficient (ADC) and dynamic susceptibility contrast (DSC) perfusion maps. Prognostic associations were assessed using univariate and adjusted Cox regression. A benchmark compared clinical, diffusion, perfusion, and combined models using cross-validated concordance indices and permutation testing. Results: ADC standard deviation (ADCstd) showed the strongest univariate prognostic association (hazard ratio 1.56, false discovery rate q = 0.0003, concordance index 0.621) and remained independently significant after clinical adjustment (HR 1.48, p < 0.001). Mean transit time standard deviation (MTTstd) was the strongest perfusion-derived feature (HR 1.38, q = 0.025, concordance index 0.578). ADCstd and MTTstd showed low correlation (Spearman r = 0.24). In cross-validation, neither imaging feature alone significantly improved discrimination over the clinical baseline (clinical plus ADC, {Delta}C = +0.058, p = 0.071; clinical plus MTT, {Delta}C = +0.035, p = 0.194). Only the model combining clinical variables, ADCstd and MTTstd achieved a significant improvement (concordance index 0.619; {Delta}C = +0.072, p = 0.029). Conclusion: ADC heterogeneity was the numerically strongest imaging signal, while DSC perfusion heterogeneity was weaker and less consistent. Only the combined model significantly outperformed the clinical baseline, but not ADC alone, leaving perfusion's contribution unproven.