Abstract
Despite advances in self-supervised learning for robust end-to-end medical image registration models, instance-based optimisation for the alignment of geometric or semantic features still remains highly effective for complex clinical tasks.
Recently, the fitting of SIREN-based neural fields has challenged the direct optimisation of transformation parameters.
In this work, we provide insights into these underlying mechanisms by drawing comparisons to explicit strategies for over-parameterisation.
By evaluating a unified explicit over-parameterisation scheme that combines 12-degree-of-freedom (DoF) log-Euclidean polyaffine matrix prediction with spatially adaptive multi-scale smoothing, we empirically demonstrate a substantial positive impact on first-order optimisers (Adam), whereas second-order methods (L-BFGS) remain vulnerable to local minima in sparse formulations.
The SIREN-based neural field in contrast benefits less, because it inherently captures local coarse-to-fine smoothness as well as global rotation, shears, or scaling.
A symmetric objective aligning multi-scale MIND features in a mutual midpoint space enforces inverse consistency throughout optimisation.
Combining these ideas sets new state-of-the-art performance for unsupervised inter-subject abdominal CT registration and improves upon previous continuous-optimisation-based approaches for large-deformation COPD inspiration-expiration alignment.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/Off_Grid_033.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to Open Review
Open Review Page: https://openreview.net/forum?id=gPoTNCCA4X
BibTex
@InProceedings{HeiMat_Overparameterising_MICCAISAT2026,
author = { Heinrich, Mattias P.},
title = { { Over-parameterising of optimisation for sparse 3D medical image registration } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026 Workshops and Challenges},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 17277},
month = {pending},
page = {pending}
}