Physics-Grounded Adversarial Stain Augmentation with Calibrated Coverage Guarantees

Stain variation across hospitals degrades histopathology models at deployment. Existing augmentation methods perturb color spaces with arbitrary hyperparameters, lacking both a principled budget and coverage guarantees for unseen centers. We propose $\mathbf{\text{C}}$alibrated $\mathbf{\text{A}}$dversarial $\mathbf{\text{S}}$tain $\mathbf{\text{A}}$ugmentation ($\mathbf{\text{CASA}}$), which performs adversarial augmentation in the Macenko stain parameter space with a budget calibrated from multi-center statistics via the DKW inequality. On Camelyon17-WILDS (5 seeds), CASA achieves $93.9\mathrm{\%}\pm 1.6\mathrm{\%}$ slide-level accuracy – outperforming HED-strong ($88.4\mathrm{\%}\pm 7.3\mathrm{\%}$), RandStainNA ($85.2\mathrm{\%}\pm 6.7\mathrm{\%}$), and ERM ($63.9\mathrm{\%}\pm 11.3\mathrm{\%}$) – with the highest worst-group accuracy ($84.9\mathrm{\%}\pm 0.9\mathrm{\%}$) among all 10 compared methods.