CRISP: Constrained Refinement via Iterative Squeezing Process for Robust Medical Image Segmentation under Domain Shift
Merged summary
TL;DR - CRISP is a source-only medical image segmentation framework that stays robust under domain shift without test-time updates or target-domain data, addressing a key barrier to clinical AI deployment.
- Built on a "Rank Stability of Positive Regions" assumption: probability rankings stay stable under distribution shift, yielding robust spatial priors rather than direct mask outputs.
- Uses latent feature perturbation to derive dual priors — a high-precision (HP) core from perturbation-invariant regions and a high-recall (HR) support from voxels foreground under any perturbation.
- An iterative training framework recursively "squeezes" HP and HR toward the final segmentation; it's model-agnostic, target-free, and runs with frozen weights.
- Evaluated on multi-center cardiac MRI and CT lung vessel segmentation, reporting HD95 reductions up to 0.14 (7.0%), 1.90 (13.1%), and 8.39 (38.9%) across multi-center, demographic, and modality shifts.
Sources (1)
CRISP: Constrained Refinement via Iterative Squeezing Process for Robust Medical Image Segmentation under Domain Shift
TL;DR - CRISP is a source-only medical image segmentation framework that stays robust under domain shift without test-time updates or target-domain data, addressing a key barrier to clinical AI deployment.
- Built on a "Rank Stability of Positive Regions" assumption: probability rankings stay stable under distribution shift, yielding robust spatial priors rather than direct mask outputs.
- Uses latent feature perturbation to derive dual priors — a high-precision (HP) core from perturbation-invariant regions and a high-recall (HR) support from voxels foreground under any perturbation.
- An iterative training framework recursively "squeezes" HP and HR toward the final segmentation; it's model-agnostic, target-free, and runs with frozen weights.
- Evaluated on multi-center cardiac MRI and CT lung vessel segmentation, reporting HD95 reductions up to 0.14 (7.0%), 1.90 (13.1%), and 8.39 (38.9%) across multi-center, demographic, and modality shifts.