CRISP: Constrained Refinement via Iterative Squeezing Process for Robust Medical Image Segmentation under Domain Shift

Medical image segmentation models often suffer from performance degradation when deployed in environments different from their training data, a phenomenon known as distribution shift. Existing methods typically rely on either simulating potential shifts during training or adapting model parameters during inference. This paper introduces CRISP (Constrained Refinement via Iterative Squeezing Process), a model-agnostic, target-free framework designed to maintain segmentation accuracy under unpredictable, real-world domain shifts without requiring parameter updates at deployment.

The authors propose the "Rank Stability of Positive Regions" as a core working assumption: while absolute probability values may fluctuate under domain shift, the relative ranking of foreground voxels remains largely consistent. CRISP exploits this by using latent feature perturbation to simulate domain shifts, identifying two types of spatial hints: a high-precision (HP) core (voxels that are consistently foreground) and a high-recall (HR) support (voxels that are potentially foreground). These hints are then used in a recursive self-evolution training framework that iteratively "squeezes" the uncertainty gap between the HP and HR regions, forcing the model to produce more decisive and accurate segmentations.

CRISP demonstrates significant robustness across diverse clinical scenarios, including multi-center cardiac MRI and CT-based lung vessel segmentation. By operating with frozen weights and requiring no target-domain data, the method consistently outperforms state-of-the-art domain generalization techniques. Experimental results show substantial improvements in HD95 metrics, with reductions of up to 38.9% across various modality and demographic shifts, confirming that rank-based spatial priors are a highly effective mechanism for handling domain-induced uncertainty.