Medical AI has a nagging habit of falling apart when it meets images from a scanner it never trained on, which can widen health gaps rather than close them. Most fixes try to anticipate the shift by simulating it or making pseudo-labels, but the real world throws up an essentially endless variety you cannot list in advance.

CRISP leans on a quieter regularity: even when a model's confidence numbers drift on unfamiliar images, the ranking of which regions it trusts most tends to hold. From that it carves a high-precision core that survives repeated perturbations to the model's internal features and a wider high-recall region that might be foreground under some perturbation, then squeezes the two together over several rounds. No target data, no test-time tuning, weights frozen. On cardiac MRI and lung-vessel CT, it reports better boundary accuracy than prior methods.

That comes from the abstract, so see the paper for the full assumption and exact reductions.