SegWithU Introduces Lightweight Uncertainty Head for Single-Pass Medical Image Segmentation
SegWithU, introduced in arXiv:2604.15271, is a post-hoc framework that adds a lightweight uncertainty head to frozen pretrained segmentation backbones. It leverages intermediate features and models uncertainty as perturbation energy using rank-1 posterior probes in a compact probe space. The method outputs two voxel-wise maps: one for calibration via probability tempering and one for ranking errors and selective prediction. Evaluated on ACDC, BraTS2024, and LiTS, it outperforms all single-forward-pass baselines in both strength and consistency. The approach addresses the demand for reliable uncertainty without repeated inference.
Key facts
- SegWithU is a post-hoc framework augmenting frozen segmentation backbones with a lightweight uncertainty head.
- It uses rank-1 posterior probes to model uncertainty as perturbation energy in a compact probe space.
- The method produces two voxel-wise uncertainty maps: calibration-oriented and ranking-oriented.
- It was evaluated on ACDC, BraTS2024, and LiTS datasets.
- SegWithU is the strongest and most consistent single-forward-pass baseline across these benchmarks.
- The paper is available on arXiv under identifier 2604.15271.
- It is a revised version (v4) with announcement type 'replace-cross'.
- The approach targets reliable uncertainty estimation for medical image segmentation without repeated inference.
Entities
—