Feature-Spectral Fragility in Segmentation: Dataset Dependence, Architecture-Specific Localization, and Spectral Correlates
The study investigated how robust segmentation models are to changes in input features versus learned representations. Findings show that feature-domain fragility is much worse on the CVC dataset than ISIC for all architectures. Sensitivity is not uniform; CNN and SSM models show sensitivity at different depths, and high-frequency energy correlates inversely with fragility on CVC, suggesting specific architectural vulnerabilities.