Abstract
Typically, learning-based image registration identifies a deformation field deterministically in a single pass through the neural network, with no test-time adaption or quantification of confidence in the prediction. We propose Recursive Uncertainty-Gated Image Registration (RUGI), a model-agnostic, inference-time algorithm, to recursively refine an initial deformation field while focusing updates on regions of high uncertainty. RUGI produces uncertainty estimates which can be propagated to downstream tasks, and improves registration performance. We evaluate on two datasets spanning cardiac MRI and ultrasound, and demonstrate statistically significant improvements over single-step inference. Mean squared error (MSE) reductions exceed 20% relative to the single-step baseline, with gains in anatomical alignment confirmed by Dice similarity within labelled datasets. RUGI can also be applied directly to existing pretrained models; applied to VoxelMorph, TransMorph, and CycleMorph in a third dataset, it yields MSE reductions of 27-37% with no modification to the original training procedure. Therefore, we expect RUGI to improve performance and drive interpretability in any learning-based image registration scenario.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/iMIMIC_009.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to Open Review
Open Review Page: Not Available
BibTex
@InProceedings{RodCla_Recursive_MICCAISAT2026,
author = { Rodrigo González, Clara AND Bates, Oscar AND Ng, Fu Siong AND Tang, Meng-Xing},
title = { { Recursive Uncertainty-Gated Image Registration for Learning-Based Algorithms } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026 Workshops and Challenges},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 17279},
month = {pending},
page = {pending}
}