CoRe: Joint Optimization with Contrastive Learning for Medical Image Registration

Eytan Kats; Christoph Grossbroehmer; Ziad Al-Haj Hemidi; Fenja Falta; Wiebke Heyer; Mattias P. Heinrich

cs.CVMar 24, 2026

CoRe: Joint Optimization with Contrastive Learning for Medical Image Registration

Authors: Eytan Kats, Christoph Grossbroehmer, Ziad Al-Haj Hemidi, Fenja Falta, Wiebke Heyer, Mattias P. Heinrich

Organizations: Insitute of Medical Informatics, University of Luebeck, Luebeck, Germany

Abstract

Medical image registration is a fundamental task in medical image analysis, enabling the alignment of images from different modalities or time points. However, intensity inconsistencies and nonlinear tissue deformations pose significant challenges to the robustness of registration methods. Recent approaches leveraging self-supervised representation learning show promise by pre-training feature extractors to generate robust anatomical embeddings, that farther used for the registration. In this work, we propose a novel framework that integrates equivariant contrastive learning directly into the registration model. Our approach leverages the power of contrastive learning to learn robust feature representations that are invariant to tissue deformations. By jointly optimizing the contrastive and registration objectives, we ensure that the learned representations are not only informative but also suitable for the registration task. We evaluate our method on abdominal and thoracic image registration tasks, including both intra-patient and inter-patient scenarios. Experimental results demonstrate that the integration of contrastive learning directly into the registration framework significantly improves performance, surpassing strong baseline methods.

Figures & tables

Figure 1: Comparison of hybrid registration methods. From left to right: (1) Feature extractor pretrained separately and used without further optimization during registration (SAMConvex Li et al. (2023) ); (2) Feature extractor optimized exclusively with a registration loss during training (Bigalke et al. Bigalke et al. (2023) ); (3) Proposed CoRe method, where the feature extractor is jointly optimized under both registration and contrastive loss objectives to enhance feature robustness and registration accuracy.

Figure 2: Overview of the proposed CoRe framework: The feature extractor is jointly optimized using registration and equivariance-based contrastive objectives, enabling robust and spatially coherent feature representations for precise displacement field estimation.

Method	AbdomenCT			RadChestCT
Method	$\text{DSC}\uparrow$	$\text{SDLogJ}\downarrow$	$\text{T}_{inf}\downarrow$	$\text{DSC}\uparrow$	$\text{SDLogJ}\downarrow$	$\text{T}_{inf}\downarrow$
Initial	25.9 $\pm$ 7.1	–	–	34.1 $\pm$ 14.5	–	–
DEEDs Heinrich et al. (2013)	46.5 $\pm$ 8.1	0.058	45.2 s	88.5 $\pm$ 6.9	0.049	45.0 s
NiftyReg Modat et al. (2010)	34.9 $\pm$ 9.3	0.034	123.5 s	84.4 $\pm$ 7.4	0.021	85.5 s
VoxelMorph Balakrishnan et al. (2019)	35.4 $\pm$ 8.9	0.134	0.2 s	55.6 $\pm$ 10.3	0.101	0.3 s
LapIRN Mok and Chung (2020)	42.4 $\pm$ 8.2	0.089	0.7 s	83.4 $\pm$ 7.6	0.075	0.8 s

Table 1: Quantitative results for AbdomenCT and RadChestCT registration tasks. The proposed CoRe method outperforms all comparison methods in terms of Dice, with statistical significance confirmed by a Wilcoxon signed-rank test (p ¡ 0.0001), underscoring the advantages of the joint optimization strategy. The smoothness of the predicted displacement fields (SDlogJ) remains low and comparable to other methods.

Figure 3: Qualitative results of the proposed CoRe method. From left to right: fixed image, fixed image with its segmentation overlay, fixed image with the overlay of the moving image segmentation, and fixed image with the overlay of the warped segmentation. The top two rows show examples from the AbdomenCT dataset in the axial plane, while the bottom two rows present examples from the RadChestCT dataset in axial and coronal planes.

$\mathcal{L}_{reg}$	$\mathcal{L}_{cl}$	AbdomenCT		RadChestCT
$\mathcal{L}_{reg}$	$\mathcal{L}_{cl}$	$\text{DSC}\uparrow$	$\text{SDLogJ}\downarrow$	$\text{DSC}\uparrow$	$\text{SDLogJ}\downarrow$
✓		51.1 $\pm$ 7.3	0.146	87.3 $\pm$ 6.9	0.106
	✓	50.3 $\pm$ 7.5	0.151	86.7 $\pm$ 7.1	0.117
✓	✓	52.6 $\pm$ 7.5	0.148	89.4 $\pm$ 6.6	0.109

Table 2: Comparison of training the feature extractor with registration and contrastive objectives separately and jointly. The proposed joint optimization strategy achieves the best performance.

$T_{i}$	$T_{g}$	AbdomenCT		RadChestCT
$T_{i}$	$T_{g}$	$\text{DSC}\uparrow$	$\text{SDLogJ}\downarrow$	$\text{DSC}\uparrow$	$\text{SDLogJ}\downarrow$
		51.1 $\pm$ 7.3	0.146	87.3 $\pm$ 6.9	0.106
✓		49.4 $\pm$ 8.1	0.139	87.5 $\pm$ 7.0	0.115
	✓	52.6 $\pm$ 7.5	0.148	89.4 $\pm$ 6.6	0.109
✓	✓	50.2 $\pm$ 7.6	0.138	89.3 $\pm$ 6.5	0.102

Table 3: Comparison of the effect of intensity invariance and geometric equivariance constraints, applied separately and jointly. Training with the equivariance constraint achieves the best performance, highlighting the importance of feature robustness to non-linear tissue deformations.

Figure 4: (a) Influence of the contrastive loss weighting coefficient

\alpha

on registration performance. (b) Evolution of the Dice score over training iterations, comparing the proposed joint optimization strategy with a baseline trained using only the registration loss.

Explore similar work

Effective Feature Learning for 3D Medical Registration via Domain-Specialized DINO Pretraining
Mar 14, 2026Eytan Kats, Mattias P. HeinrichMedical Image Registration Deformable Image Registration
Institute of Medical Informatics, University of Lübeck, Germany
Search-MIND: Training-Free Multi-Modal Medical Image Registration
Apr 10, 2026Boya Wang, Ruizhe Li, Chao Chen +1Medical Image Registration Cross-Modal Image Registration
Intelligent Modelling & Analysis Group (IMA), School of Computer Science, University of Nottingham, UK · Lab for Uncertainty in Data and Decision Making (LUCID), Nottingham Biomedical Research Centre (BRC), School of Medicine, University of Nottingham, UK
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

Mar 14, 2026J/K move · Enter open · S save

Effective Feature Learning for 3D Medical Registration via Domain-Specialized DINO Pretraining

Eytan Kats, Mattias P. Heinrich

Institute of Medical Informatics, University of Lübeck, Germany

Medical Image Registration Deformable Image Registration Medical Image Representation Learning Self-Supervised Learning Cross-Modal Image Registration

Open paper AskPDF arXiv

Medical image registration is a critical component of clinical imaging workflows, enabling accurate longitudinal assessment, multi-modal data fusion, and image-guided interventions. Intensity-based approaches often struggle with interscanner variability and complex anatomical deformations, whereas feature-based methods offer improved robustness by leveraging semantically informed representations. In this work, we investigate DINO-style self-supervised pretraining directly on 3D medical imaging data, aiming to learn dense volumetric features well suited for deformable registration. We assess the resulting representations on challenging interpatient abdominal registration task across both MRI and CT modalities. Our domain-specialized pretraining outperforms the DINOv2 model trained on a large-scale collection of natural images, while requiring substantially lower computational resources at inference time. Moreover, it surpasses established registration models under out-of-domain evaluation, demonstrating the value of task-agnostic yet medical imaging-focused pretraining for robust and efficient 3D image registration.

Abstract

Explore similar work

Effective Feature Learning for 3D Medical Registration via Domain-Specialized DINO Pretraining

Search-MIND: Training-Free Multi-Modal Medical Image Registration

UniReg: Conditional Unified Model for Medical Image Registration