CASA boosts histopathology AI to 93.9% with calibrated stain augmentation
New method CASA achieves 93.9% accuracy on Camelyon17-WILDS, beating all baselines.
Histopathology AI models often fail when deployed across hospitals due to stain variations in tissue slides. Existing augmentation methods rely on arbitrary hyperparameters, lacking principled budgets and coverage guarantees for unseen centers. Mingi Hong's new paper introduces CASA (Calibrated Adversarial Stain Augmentation), which performs adversarial augmentation in the Macenko stain parameter space. The budget is calibrated using multi-center statistics via the DKW inequality, providing theoretical coverage guarantees for new hospitals.
On the challenging Camelyon17-WILDS benchmark (5 seeds), CASA achieves 93.9% ± 1.6% slide-level accuracy – a dramatic improvement over HED-strong (88.4% ± 7.3%), RandStainNA (85.2% ± 6.7%), and standard ERM (63.9% ± 11.3%). Crucially, CASA also delivers the highest worst-group accuracy (84.9% ± 0.9%) among all 10 compared methods, indicating robust performance across difficult subgroups. This work provides a principled and reliable solution for deploying histopathology AI across diverse medical centers, with calibrated guarantees that reduce the risk of unexpected failures.
- CASA uses Macenko stain parameter space with DKW inequality to calibrate augmentation budget, providing theoretical coverage guarantees for unseen hospitals.
- Achieves 93.9% slide-level accuracy on Camelyon17-WILDS, surpassing HED-strong (88.4%), RandStainNA (85.2%), and ERM (63.9%).
- Highest worst-group accuracy (84.9%) among 10 methods, demonstrating robustness across challenging subgroups.
Why It Matters
Enables reliable histopathology AI across hospitals by tackling stain variation with calibrated guarantees.