Validating Radiomic Feature Comparisons with Patient-Specific Image Surrogates

Radiomic features may discriminate outcomes without identifying the image property responsible for that discrimination. We evaluated a patient-specific surrogate framework that tests feature responses to explicitly defined image modifications and diagnoses limitations of the surrogate procedure itself. Four public brain tumor magnetic resonance imaging collections contributed 1699 studies passing initial quality control. Core validation used smaller subsets, including 40 depth-stratified control evaluations from 38 glioma subjects, alongside a constructive counterexample and Gaussian-null diagnostics. Thirteen quantile-based morphological and topological descriptors were compared with conventional radiomics. Binary image pairs with exactly equal distance-1 co-occurrence counts differed in Euler characteristic beyond an independent ensemble's range in 16/16 replicates, versus 0/16 negative controls. This establishes a specific limitation of those co-occurrence measurements. The Gaussian spectrum estimator was parameter-sensitive: only three of five evaluated settings satisfied an operational negative-control diagnostic, which does not establish statistical calibration. Under the depth-stratified null, the median fraction of descriptors detecting planted objects was 61.5% for the topological set, versus 13.7% and 31.5% for conventional features at 24 and 64 bins. These are descriptive, configuration-dependent comparisons; dependent features, control-based selection, and incomplete spectral convergence limit inference. Secondary clinical applications did not establish biological transferability or prognostic benefit. The framework provides reproducible tests of specified feature-null combinations, while exposing the controls and assumptions needed to interpret them.