ACorD: Adaptive Correspondence Discovery for Medical Image Registration

Arghavan Rezvani; Kun Han; Xiaohui Xie

cs.CVOct 19, 2026

ACorD: Adaptive Correspondence Discovery for Medical Image Registration

Authors: Arghavan Rezvani, Kun Han, Xiaohui Xie

Organizations: University of California, Irvine

Abstract

Medical image registration is a fundamental task for aligning anatomical structures across images. Exhaustively searching for point-wise correspondences is prohibitively expensive in both computation and memory, particularly for high-resolution 3D medical images. Most recent learning-based methods directly predict correspondences in a coarse-to-fine manner, but often propagate intermediate estimates across scales without explicitly weighting correspondence evidence, which can allow coarse-scale errors to persist during refinement. To address these challenges, we propose ACorD, an adaptive correspondence discovery framework designed to mitigate error propagation with an attention-guided process. ACorD first predicts multiple potential correspondence candidates through Correlation-guided Cross-scale Offset Prediction (CCOP), which combines correlation maps as spatial matching priors with cross-scale features as global deformation context. Then, Multi-Head Deformable Aggregation (MHDA) aggregates these candidate correspondences using attention, producing a learned, non-uniform weighting over competing matches rather than treating them equally. The multi-head design further captures complementary deformation patterns, ensuring both global alignment and local precision. By integrating adaptive offset prediction with selective, attention-based aggregation of correspondence evidence, ACorD enables accurate deformation refinement. Extensive experiments on four public medical registration benchmarks demonstrate that adaptive correspondence discovery within a multi-scale refinement framework improves registration robustness across diverse anatomical structures.

Figures & tables

Figure 1: Adaptive Correspondence Discovery. ACorD calculates attention between a feature from the fixed image and a set of sampled features from the moving image using predicted correspondence candidates. Predicted offsets guide the sampling, and the resulting attention weights determine the final correspondence.

Figure 2: Overview of the Proposed Method. A shared encoder extracts multi-scale representations from both images, which are processed by ACorD blocks in a coarse-to-fine manner. At each scale, ACorD combines Correlation-guided Cross-scale Offset Prediction (CCOP) with Multi-Head Deformable Aggregation (MHDA) to refine the deformation field through attention-guided correspondence discovery and propagation.

Figure 3: Adaptive Correspondence Discovery (ACorD). ACorD consists of two modules: (1) CCOP predicts multi-head adaptive spatial offsets from projected features, local-window correlations, and cross-scale context, while generating features for the next scale. (2) MHDA uses these offsets to dynamically place attention keys, where queries come from fixed features, keys from warped moving features, and values from sampled positions. Attention weights modulate the contribution of each candidate and the multi-head design captures diverse deformation patterns.

Group	Method	MindBoggle (97 ROIs)			LPBA (54 ROIs)			OASIS (35 ROIs)			#params (M)
Group	Method	DSC $\uparrow$	ASSD $\downarrow$	NJD(%) $\downarrow$	DSC $\uparrow$	ASSD $\downarrow$	NJD(%) $\downarrow$	DSC $\uparrow$	ASSD $\downarrow$	NJD(%) $\downarrow$	#params (M)
Traditional	Affine	0.404 $\pm$ 0.021	2.137 $\pm$ 0.196	0	0.635 $\pm$ 0.020	2.120 $\pm$ 0.125	0	0.571 $\pm$ 0.052	1.618 $\pm$ 0.281	0	–
Traditional	SyN [ 2 ]	0.513 $\pm$ 0.031	1.692 $\pm$ 0.211	0	0.695 $\pm$ 0.017	1.780 $\pm$ 0.109	0	0.755 $\pm$ 0.029	0.837 $\pm$ 0.123	0	–
Single-Scale	VoxelMorph [ 4 ]	0.500 $\pm$ 0.108	2.147 $\pm$ 1.092	0.883 $\pm$ 0.289	0.638 $\pm$ 0.031	2.122 $\pm$ 0.242	0.589 $\pm$ 0.273	0.763 $\pm$ 0.031	0.823 $\pm$ 0.145	0.963 $\pm$ 0.179	0.27
	ViT-V-Net [ 6 ]	0.526 $\pm$ 0.112	2.086 $\pm$ 1.128	0.797 $\pm$ 0.332	0.648 $\pm$ 0.030	2.049 $\pm$ 0.225	0.735 $\pm$ 0.247	0.787 $\pm$ 0.031	0.733 $\pm$ 0.136	0.917 $\pm$ 0.151	31.56
	TransMorph [ 7 ]	0.551 $\pm$ 0.118	1.963 $\pm$ 1.083	0.691 $\pm$ 0.181	0.663 $\pm$ 0.030	1.982 $\pm$ 0.232	0.648 $\pm$ 0.199	0.806 $\pm$ 0.025	0.661 $\pm$ 0.103	0.827 $\pm$ 0.169	46.77

Table 1: Quantitative comparisons on MindBoggle, LPBA, and OASIS datasets. Red , Green and Blue denote the 1st , 2nd and 3rd best.

Group	Method	AbdomenCT (13 ROIs)
Group	Method	DSC $\uparrow$	ASSD $\downarrow$	NJD(%) $\downarrow$
Traditional	Affine	0.251 $\pm$ 0.070	6.629 $\pm$ 2.416	0
Traditional	SyN [ 2 ]	0.319 $\pm$ 0.084	5.780 $\pm$ 2.443	0
Single-Scale	VoxelMorph [ 4 ]	0.349 $\pm$ 0.051	4.793 $\pm$ 1.047	1.746 $\pm$ 0.094
	ViT-V-Net [ 6 ]	0.385 $\pm$ 0.054	4.547 $\pm$ 1.032	4.798 $\pm$ 1.400
	TransMorph [ 7 ]	0.409 $\pm$ 0.061	4.422 $\pm$ 1.118	5.554 $\pm$ 1.368

Table 2: Quantitative comparison on AbdomenCT dataset. Red , Green and Blue denote the 1st , 2nd and 3rd best resutls.

Figure 4: Visual comparisons on the Brain LPBA dataset (top two rows) and the Abdomen CT dataset (bottom two rows)

Figure 5: Dice score evaluation on AbdomenCT (13 ROIs)

	Components						AbdomenCT (13 ROIs)
Exp	CG	CS	DA	MH	Agg	Offset	DSC $\uparrow$	ASSD $\downarrow$	NJD(%) $\downarrow$
1							0.549 $\pm$ 0.054	2.83 $\pm$ 0.67	0.105 $\pm$ 0.08
2	✓						0.570 $\pm$ 0.051	2.731 $\pm$ 0.658	0.103 $\pm$ 0.066
3	✓	✓					0.577 $\pm$ 0.046	2.626 $\pm$ 0.646	0.151 $\pm$ 0.105
4			✓		Attn	Learned	0.564 $\pm$ 0.045	2.649 $\pm$ 0.598	0.098 $\pm$ 0.064
5			✓	✓	Attn	Learned	0.579 $\pm$ 0.049	2.585 $\pm$ 0.651	0.084 $\pm$ 0.048

Table 3: Ablation study on the AbdomenCT dataset. CG : Correlation Guidance; CS : Cross-Scale Feature Propagation; DA : Deformation Aggregation; MH : Multi-Head mechanism; Agg : aggregation rule ( Attn : attention-weighted, Avg : uniform averaging); Offset : correspondence offset source ( Learned : CCOP-predicted, Default : fixed neighbors).

Lvl.	MindBoggle		AbdomenCT
Lvl.	Unif.	Attn.	Unif.	Attn.
1	73.89	74.40	78.35	78.56
2	65.95	68.07	74.78	75.84
3	55.55	58.03	68.79	70.79
4	49.56	49.64	59.14	62.07
5	49.19	48.91	48.29	47.31

Table 4: Aggregated position label accuracy (%) using uniform and attention-based aggregation over ACorD candidates.

Transition	ACorD(%)	Exp.7(%)	ACorD(%)	Exp.8(%)
L4 $\rightarrow$ L3	40.4	38.3 ( $p=0.004$ )	40.2	37.0 ( $p<1\mathrm{e}{-4}$ )
L3 $\rightarrow$ L2	23.6	22.6 ( $p=0.032$ )	23.8	20.6 ( $p<1\mathrm{e}{-4}$ )
L2 $\rightarrow$ L1	13.2	12.2 ( $p=0.001$ )	12.6	10.4 ( $p<1\mathrm{e}{-4}$ )

Table 5: Common-set recovery on AbdomenCT. For voxels misregistered by both models at the coarser scale, we report the percentage corrected one level finer. Higher is better; p-values use paired Wilcoxon signed-rank tests.

Lvl.	AbdomenCT		LPBA		OASIS		MindBoggle
Lvl.	Ratio	Frac. $>1$	Ratio	Frac. $>1$	Ratio	Frac. $>1$	Ratio	Frac. $>1$
1	1.46 $\pm$ 0.02	0.73 $\pm$ 0.01	1.45 $\pm$ 0.07	0.75 $\pm$ 0.03	1.07 $\pm$ 0.01	0.55 $\pm$ 0.00	1.05 $\pm$ 0.02	0.51 $\pm$ 0.05
2	1.14 $\pm$ 0.02	0.55 $\pm$ 0.01	1.18 $\pm$ 0.03	0.68 $\pm$ 0.02	0.68 $\pm$ 0.00	0.11 $\pm$ 0.00	1.09 $\pm$ 0.02	0.58 $\pm$ 0.01
3	0.93 $\pm$ 0.02	0.35 $\pm$ 0.02	0.95 $\pm$ 0.01	0.32 $\pm$ 0.01	0.79 $\pm$ 0.01	0.20 $\pm$ 0.00	0.97 $\pm$ 0.01	0.36 $\pm$ 0.01
4	1.15 $\pm$ 0.03	0.50 $\pm$ 0.02	0.65 $\pm$ 0.01	0.09 $\pm$ 0.01	0.75 $\pm$ 0.00	0.20 $\pm$ 0.00	0.65 $\pm$ 0.03	0.13 $\pm$ 0.03
5	1.54 $\pm$ 0.04	0.73 $\pm$ 0.01	1.10 $\pm$ 0.01	0.49 $\pm$ 0.01	0.97 $\pm$ 0.01	0.29 $\pm$ 0.01	0.91 $\pm$ 0.00	0.19 $\pm$ 0.00

Table 6: Statistics of predicted candidate offsets. Ratio : the ratio of predicted offsets to default offset magnitude. Frac.

>1

: fraction of predicted candidate offsets with ratios greater than 1.

$K^{3}$	AbdomenCT			MindBoggle
$K^{3}$	DSC $\uparrow$	ASSD $\downarrow$	NJD(%) $\downarrow$	DSC $\uparrow$	ASSD $\downarrow$	NJD(%) $\downarrow$
$2^{3}=$ 8	0.595 $\pm$ 0.056	2.475 $\pm$ 0.614	0.135 $\pm$ 0.096	0.621 $\pm$ 0.054	1.512 $\pm$ 0.517	0.005 $\pm$ 0.006
$3^{3}=27$	0.597 $\pm$ 0.052	2.411 $\pm$ 0.603	0.136 $\pm$ 0.092	0.624 $\pm$ 0.049	1.468 $\pm$ 0.447	0.005 $\pm$ 0.004

Table S1 : Deformable Attention Size

Figure S1 : Network Structure for the Encoder layers and Offset Predictor.

$\lambda_{jac}$	ACorD
	AbdomenCT		MindBoggle
	DSC $\uparrow$	NJD(%) $\downarrow$	DSC $\uparrow$	NJD(%) $\downarrow$
0	0.590	8.381	0.625	0.191
20	0.592	0.317	0.623	0.011
50	0.597	0.136	0.624	0.005
80	0.595	0.086	0.623	0.003

Table S2 : Hyperparameter Tuning for Negative Jacobian Loss

Head Allocation	DSC $\uparrow$	ASSD $\downarrow$	NJD(%) $\downarrow$
$[2,2,4,4,1]$	$0.595\pm 0.057$	$2.428\pm 0.601$	$0.18\pm 0.14$
$[4,4,2,2,1]$	$0.594\pm 0.052$	$2.418\pm 0.603$	$0.094\pm 0.05$
default $[8,4,2,1,1]$	$0.597\pm 0.052$	$2.411\pm 0.603$	$0.136\pm 0.092$

Table S3 : Ablation Study on Head Allocation

Figure S2 : Dice score evaluation on LPBA brain image registration (54 ROIs).

Figure S3 : Artifact of ACorD registration. Left: In regions highlighted by the red boxes, the cortical surface is shown to have unnecessary folding. Right: While intensity similarity metrics may indicate good alignment, errors revealed through segmentation comparisons demonstrate the shortcomings of structure-unaware supervision. The mismatches show that the model might achieve high image similarity at the cost of structural integrity, leading to inaccuracies at the structure level.

Model	MACs (GMac)	Runtime (ms)	Memory (MB)
TransMorph	714.16	161.15	5353.16
TransMatch	717	257	4768
TransDCA	4766	593	9431
NICE-Trans	313	415	6205
RDPNet	3109	428	9306
Dual-PR++	1773	491	9461

Table S4 : Computational cost comparison on A6000.

Model	L5	L4	L3	L2	L1
Exp.8 (Default+Attn)	0.298	0.403	0.476	0.534	0.562
Exp.7 (Learned+Avg)	0.289	0.403	0.472	0.554	0.587
ACorD	0.304	0.408	0.500	0.563	0.597

Table S5 : Per-scale DSC on AbdomenCT

Medical Image Registration

Medical image registration aligns scans across time, patients, or imaging modalities to track anatomy and support diagnosis, treatment, and image-guided procedures.

See all 96 papers

FlexiMIND: Combined Pretrained Vision Features and Local Structure for Training-Free Deformable Medical Image Registration
Oct 25, 2026Mojtaba Safari, Mingzhe Hu, Shansong Wang +6Medical Image Registration Deformable Image Registration
Department of Radiation and Cellular Oncology, The University of Chicago, Chicago, IL, United States. · Department of Computer Science and Informatics, Emory University, Atlanta, GA, United States. · Department of Radiology, The University of Chicago, Chicago, IL, United States.
CoRe: Joint Optimization with Contrastive Learning for Medical Image Registration
Mar 24, 2026Eytan Kats, Christoph Grossbroehmer, Ziad Al-Haj Hemidi +3Medical Image Registration Deformable Image Registration
Insitute of Medical Informatics, University of Luebeck, Luebeck, Germany
UniReg: Conditional Unified Model for Medical Image Registration
Mar 17, 2025Zi Li, Jianpeng Zhang, Tai Ma +7Healthcare Medical Image Registration
Department of Electrical and Computer Engineering, The University of Hong Kong, Hong Kong SAR, China · DAMO Academy, Alibaba Group · The First Affiliated Hospital, Zhejiang University School of Medicine, Hangzhou, China +1

Oct 25, 2026J/K move · Enter open · S save

FlexiMIND: Combined Pretrained Vision Features and Local Structure for Training-Free Deformable Medical Image Registration

Mojtaba Safari, Mingzhe Hu, Shansong Wang, Suyang Xi+5

Department of Radiation and Cellular Oncology, The University of Chicago, Chicago, IL, United States. · Department of Computer Science and Informatics, Emory University, Atlanta, GA, United States. · Department of Radiology, The University of Chicago, Chicago, IL, United States.

Medical Image Registration Deformable Image Registration Cross-Modal Image Registration Cross-Modal Feature Matching

Open paper AskPDF arXiv

Objective. Deformable image registration remains challenging because corresponding anatomy may exhibit different intensity relationships. Hand-crafted modality-robust descriptors such as Modality-Independent Neighborhood Descriptor with Self-Similarity Context (MIND-SSC) and features from pretrained vision encoders provide distinct representations for correspondence estimation, but how they should be combined remains unclear. We investigate whether tighter integration improves registration over a sequential, scale-separated design. Approach. We introduce FlexiMIND, a training-free and label-free two-stage framework in which frozen FlexiCT-2D encoder features establish contextual correspondence and MIND-SSC refines the deformation within ConvexAdam. Five alternative feature-coupling strategies were evaluated under a protocol across CT-CT (DIR-Lab 4DCT 4DCT, 10 cases), MR-MR (ACDC, 50 patients), and MR-CT (Learn2Reg Abdomen MR-CT, 8 pairs) registration. Performance was assessed using target registration error (TRE), Dice similarity coefficient (Dice), SDlogJ, and local folding. Main results. FlexiMIND achieved a TRE of 1.27 mm on DIR-Lab 4DCT, a DSC of 0.807 on ACDC, and an organ-averaged DSC of 0.825 on Abdomen MR-CT. Across 15 coupling-strategy dataset comparisons, only four outperformed the reference method, with a maximum absolute improvement of 0.0043 in Dice, whereas degradation reached 20.6 % in TRE and 37.9 % in Dice. Changes in registration accuracy were closely associated with changes in deformation-field regularity, and strategies that better preserved regularity generally remained closest to the reference performance. Landmark-level analysis showed that the pretrained encoder stage primarily reduced large correspondence errors rather than uniformly improving local precision. Code is available upon acceptance at https://github.com/mosaf/FlexiMIND.

Abstract

Medical Image Registration

Explore similar work

FlexiMIND: Combined Pretrained Vision Features and Local Structure for Training-Free Deformable Medical Image Registration

CoRe: Joint Optimization with Contrastive Learning for Medical Image Registration

UniReg: Conditional Unified Model for Medical Image Registration