Characterizing the Stability of Radiomics-Derived Tumor Habitats Using Image Perturbation in Head and Neck Cancer

Tumor habitat imaging aims to capture intratumoral heterogeneity by grouping voxels with similar radiomic properties into spatially coherent subregions. However, radiomic features are known to be sensitive to small variations in image acquisition and processing, which can affect the stability of the resulting habitat maps. Feature repeatability is usually evaluated using test-retest scans, but such data are rarely available in clinical practice. To overcome this, we adopted an image perturbation framework, which simulates test-retest conditions by applying small, controlled changes to a single image. In head and neck cancer (HNC), where imaging is further complicated by complex anatomy, dental artifacts, and variability in tumor delineation, dedicated stability analyses are still missing. In this study, we evaluated how the repeatability of radiomic features affects habitat stability in 390 oropharyngeal cancer patients (discovery cohort). For each patient, 11 perturbed CT volumes were generated using small in-plane rotations, sub-voxel translations, and tumor-adaptive Gaussian noise. Ninety-three radiomic features were extracted from each image set, and their repeatability was assessed using the lower confidence limit of the intraclass correlation coefficient (ICC-LCL), grouped into poor, moderate, good, and excellent categories. Tumor habitats were then generated using K-means clustering (H = 3) for each feature subset, and habitat stability was measured by the Dice similarity coefficient (DSC) between habitat maps obtained from original and perturbed images. Overall, 48.4% of features were poorly repeatable and only 6.5% reached the excellent category, with first-order features being more stable than texture-based ones. Habitat stability followed a clear monotonic trend with feature repeatability: the median DSC was 0.93 for habitats generated from excellent features, 0.84 for good features, 0.75 for moderate features, and dropped to 0.41 for poorly repeatable features. Habitats generated using all features (without any repeatability-based filtering) yielded an intermediate median DSC of 0.52. All pairwise comparisons between feature subsets were statistically significant (p < 0.001). To evaluate the generalizability of these findings, the analysis was repeated in an independent external validation cohort of 372 oropharyngeal cancer patients treated at the H. Lee Moffitt Cancer Center. The stability classification showed substantial feature-level concordance between the discovery and validation cohorts (overall agreement 67.7%, quadratic-weighted Cohen's kappa = 0.78), with no feature shifting by more than two stability classes. The habitat-stability hierarchy was fully preserved in the validation cohort (median DSC of 0.87, 0.73, 0.69, and 0.39 for excellent, good, moderate, and poor features, respectively; all pairwise p < 0.001). These results show that selecting features with higher repeatability clearly improves the spatial consistency of habitat maps in HNC and support the use of perturbation-based stability analysis as a routine step in habitat imaging studies.