Polaffini: A feature-based approach for robust affine and polyaffine image registration

Antoine Legouhy; Cosimo Campo; Ross Callaghan; Hojjat Azadbakht; Hui Zhang

cs.CVFeb 19, 2026

Polaffini: A feature-based approach for robust affine and polyaffine image registration

Authors: Antoine Legouhy, Cosimo Campo, Ross Callaghan, Hojjat Azadbakht, Hui Zhang

Organizations: Hawkes Institute & Department of Computer Science, University College London, London, UK · Institut Pasteur, Universit´e Paris Cit´e, Unit´e de Neuroanatomie Appliqu´ee et Th´eorique, France · AINOSTICS ltd., Manchester, UK

Abstract

In this work we present Polaffini, a robust and versatile framework for anatomically grounded registration. Medical image registration is dominated by intensity-based registration methods that rely on surrogate measures of alignment quality. In contrast, feature-based approaches that operate by identifying explicit anatomical correspondences, while more desirable in theory, have largely fallen out of favor due to the challenges of reliably extracting features. However, such challenges are now significantly overcome thanks to recent advances in deep learning, which provide pre-trained segmentation models capable of instantly delivering reliable, fine-grained anatomical delineations. We aim to demonstrate that these advances can be leveraged to create new anatomically-grounded image registration algorithms. To this end, we propose Polaffini, which obtains, from these segmented regions, anatomically grounded feature points with 1-to-1 correspondence in a particularly simple way: extracting their centroids. These enable efficient global and local affine matching via closed-form solutions. Those are used to produce an overall transformation ranging from affine to polyaffine with tunable smoothness. Polyaffine transformations can have many more degrees of freedom than affine ones allowing for finer alignment, and their embedding in the log-Euclidean framework ensures diffeomorphic properties. Polaffini has applications both for standalone registration and as pre-alignment for subsequent non-linear registration, and we evaluate it against popular intensity-based registration techniques. Results demonstrate that Polaffini outperforms competing methods in terms of structural alignment and provides improved initialisation for downstream non-linear registration. Polaffini is fast, robust, and accurate, making it particularly well-suited for integration into medical image processing pipelines.

Figures & tables

Figure 1: Fusion of multiple affine transformations. For each quadrant: a. local affine transformations and their associated weight maps; b. the resulting transformation obtained by naïve fusion through Euclidean average; and c. the resulting transformation obtained using the log-Euclidean polyaffine framework of Arsigny et al., 2009 (illustration inspired by that paper).

Figure 2: Illustration of several steps of the polyaffine estimation. a) Two sets of paired points (plain disc: moving, circle: reference). b) Delaunay tetrahedralisation performed on the reference set (blue) and pattern reproduced on the moving one (red). c) Local affine regression between two homologous neighbourhoods. d) Set of estimated local affine transformations. e) Colour gradient representing all weight maps combined. f) Overall polyaffine transformation. Black rectangles represent the frame of the reference image.

Figure 3: For a given set of local affine transformations (a), illustration of the influence of the choice of the background transformation for the construction of the overall polyaffine transformation: b) no background transformation, c) identity background transformation. d) affine background transformation. Black rectangles represent the frame of the reference image.

Figure 4: Usual (top) and proposed (bottom) paths for non-linear registration.

Database	Training set	Validation set	Testing set
IXI	20	5	100
UK Biobank	20	5	100
ADNI (HC/MCI/AD)	60 (20/20/20)	15 (5/5/5)	150 (50/50/50)

Table 1: Distribution of the subjects for the training, validation and testing sets.

Figure 5: Average Dice scores after registration with Polaffini with various values of the smoothness parameter $\sigma$ , for subject-to-template (top) and subject-to-subject (bottom).

Panel of Figure 6 — Figure 6: Example of failure cases. Images are shown in the grid of the reference image.

	Flirt	ANTs-aff	Anima-aff	Aladin	Polaffini-aff	Polaffini-polyaff
subject-to-template	4.86	0	0.29	0	0	0
subject-to-subject	11.71	0	0.29	0	0	0

Table 2: Failure rate (%) for each registration method.

Figure 7: Average Dice scores after registration with Polaffini and various affine registration methods, for subject-to-template (top) and subject-to-subject (bottom). colours are linked to the similarity metric types (reddish: image-wise, bluish: block-wise, greenish: Polaffini ).

	regions	rank	Flirt	ANTs-aff	Anima-aff	Aladin	Polaffini-aff	Polaffini-polyaff
subject-to-template	sub-cortical	1 st	0.3		0.9	1.2	0.6	96.99
	sub-cortical	2 nd	1.51	3.61	34.04	15.66	43.37	1.81
	cortical	1 st	0	0			0	100
	cortical	2 nd	2.11	1.2	6.02	12.35	78.31	0
	white matter	1 st	0	0	0	0.3	2.11	97.59
	white matter	2 nd	0.9	1.81	22.59	16.27	56.93	1.51

Table 3: Frequency (%) at which each method performs 1 st and 2 nd in terms of Dice score, for each registration, averaged by label.

methods		subject-to-template			subject-to-subject
A	B	sub-cort	cort	WM	sub-cort	cort	WM
ANTs-aff	Flirt	1.0e-23	1.1e-18	4.0e-20	3.5e-10	6.0e-22	5.5e-17
Anima-aff	Flirt	6.5e-52	1.8e-31	1.1e-48	1.0e-42	3.8e-42	4.9e-44
Anima-aff	ANTs-aff	4.5e-46	6.0e-17	3.1e-36	6.6e-43	2.8e-33	6.8e-40
Aladin	Flirt	8.3e-35	6.8e-49	1.9e-46	8.4e-43	2.1e-40	1.1e-46
Aladin	ANTs-aff	1.3e-19	3.1e-45	4.4e-38	2.5e-41	2.6e-28	4.7e-44

Table 4: Pairwise statistical comparison of average Dice scores after image registration using Polaffini and various affine registration methods. Raw p-values for Wilcoxon signed-rank tests (

\alpha=0.05

). Significant comparisons after Bonferroni correction in favor of method A (resp. method B) are highlighted in blue (resp. red); non-significant ones in gray.

component	set	geometric distance
linear part	$\operatorname{GL}(d,\mathbb{R})$	$\mathrm{dist}(L_{i},L_{j})$	=	$\left\\|\log\left(L_{i}^{-1}L_{j}\right)\right\\|_{F}$
rotation	$\operatorname{SO}(d)$	$\mathrm{dist}(R_{i},R_{j})$	=	$\left\\|\log\left(R_{i}^{T}R_{j}\right)\right\\|_{F}$
stretch	$\operatorname{Sym}^{+}(d,\mathbb{R})$	$\mathrm{dist}(S_{i},S_{j})$	=	$\left\\|\log\left(S_{i}^{-1/2}S_{j}S_{i}^{-1/2}\right)\right\\|_{F}$
translation	$\mathbb{R}^{d}$	$\mathrm{dist}(t_{i},t_{j})$	=	$\left\\|t_{i}-t_{j}\right\\|_{2}$

Table 5: Geometric distances for the affine matrix components.

\log

is the matrix logarithm and

\|.\|_{F}

is the Frobenius norm.

Figure 8: Geometric differences between the affine transformations estimated by Polaffini-aff and the ones from various affine registration methods, for subject-to-template (top) and subject-to-subject (bottom).

	affine init.	polyaffine init.	p-value	Cohen’s d
sub-cortical regions	affine	polyaffine	$<10^{-115}$	1.864
	affine + ANTs	polyaffine + ANTs	$<10^{-14}$	0.447
	affine + deep	polyaffine + deep	$<10^{-108}$	1.762
cortex	affine	polyaffine	$<10^{-107}$	1.745
	affine + ANTs	polyaffine + ANTs	$<10^{-38}$	0.798
	affine + deep	polyaffine + deep	$<10^{-176}$	3.019

Table 6: Statistics about differences between overlap scores after registration with affine vs proposed polyaffine initialisation. Reported p-values are for paired t-tests, Cohen’s d for paired samples (with polyaffine

-

affine on numerator).

Figure 9: Dice scores after non-linear registration with traditional (ANTs) and deep-learning approaches, initialised with intensity-based affine (affine) and Polaffini-polyaff (Polaffini).

Figure 10: Evolution of image similarity (LCC) and segmentation (average Dice on all regions) losses during training of deep-learning models with affine (blue) and proposed polyaffine (red) initialisation for training (light) and validation (dark) samples.

Figure 11: Dice scores after deep-learning non-linear registration for different configurations of initialisation with Polaffini-aff and Polaffini-polyaff at training and testing.

Explore similar work

Oct 13, 2026

Registration is a Representation Problem: Solving Whole-Body PET/CT Registration

Raw image intensities are a poor surrogate for similarity across images in correspondence problems. Further, while auxiliary segmentation labels can provide semantic identification and improve matching, they can assign only coarse identities to fine-grained anatomy and often cover only a subset of structures, particularly in whole-body settings. We present an approach that instead establishes correspondence using complementary representations of anatomy and demonstrate its utility for longitudinal whole-body PET/CT registration in the Learn2Reg 2026 PSMAReg Challenge. Our 3D foundation model, anatomix, is used to derive fine-grained, per-voxel semantic representations that are consistent across volumes, and we further employ off-the-shelf segmentation models to obtain coarse anatomical identities. Together, these representations drive a GPU-accelerated iterative registration solver (FireANTs) in a hybrid multistage framework: the first stage estimates alignment from dense multi-channel features, after which a second refines the deformation using both anatomix- and segmentation-based losses. Lastly, because registration should not introduce non-biological tumor volume changes across time, we develop lesion-specific component-wise volume preservation schemes. Without any task-specific training, our system won first place in the PSMAReg Challenge leaderboard. Code and interactive visualizations are available at http://www.neeldey.com/reps-for-reg.

Neel Dey, Adrian V. Dalca

A. A. Martinos Center for Biomedical Imaging, Massachusetts General Hospital · Department of Radiology, Harvard Medical School · Computer Science and AI Lab, Massachusetts Institute of Technology

Medical Image Registration Deformable Image Registration Longitudinal Medical Image Analysis

Ask PDF

Aug 23, 2026

Architectural and Regularization Components in Deep Learning Medical Image Registration: Systematic Ablation Study

Deep learning registration methods routinely stack two kinds of enhancement on a base network: architectural additions such as affine pre-alignment stages, and training-objective additions such as regularization losses. Papers tend to adopt both at once, so it is unclear which is doing the work. I ran a controlled ablation to separate them. Using the OASIS brain MRI dataset (394 training subjects, 20 test subjects), I trained four variants of the same registration pipeline: a baseline 3D U-Net with basic similarity losses, the same U-Net with a full regularization suite, an affine-plus-deformable architecture with basic losses, and the affine architecture with the full suite. I evaluated registration accuracy (MSE, NCC, SSIM), deformation quality (Jacobian determinant preservation, displacement statistics, an anatomical plausibility score), and computational cost. Regularization alone accounted for most of the gain: a 21.3% relative gain on the MSE-improvement metric (1.78% to 2.16%, P<.001) and a 21.8% relative gain in NCC improvement, while cutting maximum deformation from 53.1 to 0.51 units, a 99.0% reduction, at essentially no computational cost (-0.06% inference time). The combined model produced the largest accuracy gain, 25.8% (1.78% to 2.24%), and raised anatomical plausibility from 0.596 to 0.930, at a moderate +9.8% inference-time cost. Gradient correlation rose from 0.742 at baseline to 0.980 for the fully enhanced model. All enhanced variants reached sub-voxel accuracy under plausible deformation constraints. Regularization losses are the primary driver in this setting, delivering the accuracy gains and almost all of the deformation control for free at inference time, while the affine architecture adds a smaller complementary benefit at acceptable cost. The 99% reduction in unrealistic deformations addresses a known barrier to clinical deployment.

Nabira Rashid

Carnegie Mellon University · Harvard Medical School

Medical Image Registration Deformable Image Registration

Ask PDF

Feb 27, 2026

GLIDE-Reg: Global-to-Local Deformable Registration Using Co-Optimized Foundation and Handcrafted Features

Deformable registration is crucial in medical imaging. Several existing applications include lesion tracking, probabilistic atlas generation, and treatment response evaluation. However, current methods often lack robustness and generalizability across two key factors: spatial resolution and differences in anatomical coverage. We jointly optimize a registration field and a learnable dimensionality reduction module so that compressed VFM embeddings remain registration-relevant, and fuse these global semantic cues with MIND local descriptors. GLIDE-Reg achieves average dice similarity coefficients (DSC) across 6 anatomical structures of 0.859, 0.862, and 0.901 in two public cohorts (Lung250M and NLST) and one institution cohort (UCLA5DCT), and outperforms the state-of-the-art DEEDS (0.834, 0.858, 0.900) with relative improvements of 3.0%, 0.5%, and 0.1%. For target registration errors, GLIDE-Reg achieves 1.58 mm on Lung250M landmarks (compared to 1.25 mm on corrField and 1.91 mm on DEEDS) and 1.11 mm on NLST nodule centers (compared to 1.11 mm on DEEDS). The substantiated performance on the nodule centers also demonstrates its robustness across challenging downstream tasks, such as nodule tracking, which is an essential prior step for early-stage lung cancer diagnosis.

Yunzheng Zhu, Aichi Chien, Kimaya kulkarni +5

Medical & Imaging Informatics, Department of Radiological Sciences, David Geffen School of Medicine at UCLA, USA · Magnetic Resonance Research Labs, Department of Radiological Sciences, David Geffen School of Medicine at UCLA, USA · Department of Radiation Oncology, David Geffen School of Medicine at UCLA, Los Angeles, USA

Medical Image Registration Deformable Image Registration Medical Image Representation Learning

Ask PDF