CRISP: Calibrated Robust Interface-aware Segmentation for Lightweight Polyp Segmentors via Amortized Boundary Posterior Projection under Domain Shift

Nguyen Vo Ngoc Khue, Thanh Trung Huynh, Huy-Hieu Pham

Keywords: Polyp segmentation, uncertainty calibration, domain shift, boundary-aware.

TL;DR: CRISP improves lightweight polyp segmentors under domain shift by turning calibration into amortized boundary posterior projection.

Registration Requirement: Yes

Abstract: Polyp segmentation under domain shift fails most critically at lesion boundaries, where geometric errors and unreliable confidence co-occur. We present CRISP, a framework for strengthening lightweight polyp segmentors through amortized boundary posterior projection. CRISP constructs a boundary-local posterior target from supervision, teacher evidence, and a soft boundary field, then fits it through a calibrated family tied to the student logit. This yields a one-dimensional local projection amortized by a small projector head. Under source-only training on Kvasir-SEG and cross-dataset evaluation on ColonDB and ETIS, CRISP consistently improves boundary geometry and boundary-local calibration on U-Net and PraNet.

Reproducibility: none

Visa & Travel: Yes

Read CFP & Author Instructions: Yes

Originality Policy: Yes

Single-blind & Not Under Review Elsewhere: Yes

LLM Policy: Yes

Submission Number: 22