Active Diffusion Matching: Score-based Iterative Alignment of Cross-Modal Retinal Images

Kanggeon Lee; Su Jeong Song; Soochahn Lee; Kyoung Mu Lee

cs.CVApr 11, 2026

Active Diffusion Matching: Score-based Iterative Alignment of Cross-Modal Retinal Images

Authors: Kanggeon Lee, Su Jeong Song, Soochahn Lee, Kyoung Mu Lee

Organizations: ASRI, Dept. of ECE, Seoul National University, Korea · Dept. of Ophthalmology, Kangbuk Samsung Hospital, Sungkyunkwan University, Korea · School of Electrical Engineering, Kookmin University, Korea

Abstract

Objective: The study aims to address the challenge of aligning Standard Fundus Images (SFIs) and Ultra-Widefield Fundus Images (UWFIs), which is difficult due to their substantial differences in viewing range and the amorphous appearance of the retina. Currently, no specialized method exists for this task, and existing image alignment techniques lack accuracy. Methods: We propose Active Diffusion Matching (ADM), a novel cross-modal alignment method. ADM integrates two interdependent score-based diffusion models to jointly estimate global transformations and local deformations via an iterative Langevin Markov chain. This approach facilitates a stochastic, progressive search for optimal alignment. Additionally, custom sampling strategies are introduced to enhance the adaptability of ADM to given input image pairs. Results: Comparative experimental evaluations demonstrate that ADM achieves state-of-the-art alignment accuracy. This was validated on two datasets: a private dataset of SFI-UWFI pairs and a public dataset of SFI-SFI pairs, with mAUC improvements of 5.2 and 0.4 points on the private and public datasets, respectively, compared to existing state-of-the-art methods. Conclusion: ADM effectively bridges the gap in aligning SFIs and UWFIs, providing an innovative solution to a previously unaddressed challenge. The method's ability to jointly optimize global and local alignment makes it highly effective for cross-modal image alignment tasks. Significance: ADM has the potential to transform the integrated analysis of SFIs and UWFIs, enabling better clinical utility and supporting learning-based image enhancements. This advancement could significantly improve diagnostic accuracy and patient outcomes in ophthalmology.

Figures & tables

Figure 1: Alignment of standard fundus images (SFIs) and ultra-widefield images (UWFIs) using ADM . We present a method for the alignment of SFI-UWFI pairs. The FOV of the SFI is limited to the orange box region of the UWFI. The cropped and zoomed-in green and red boxes highlight the alignment results of SuperRetina [ 1 ] , GeoFormer [ 2 ] , and our proposed ADM . The image below shows the intersection area between the SFI vessel (red line) and the UWFI vessel (green line), which exhibits the maximum alignment error (MAE) [ 3 , 1 , 4 , 2 ] .

Figure 2: Overview of ADM . ADM aligns the source image $\mathcal{I}_{s}$ (SFI) to the destination image $\mathcal{I}_{d}$ (UWFI) using a dual diffusion model architecture. Two score networks are employed: $\mathbf{s}_{\bm{\theta}}$ estimates global homography $\mathcal{H}$ , while $\mathbf{s}_{\bm{\phi}}$ estimates local displacement field $\mathbf{v}$ . Both networks are conditioned on the input image pair $(\mathcal{I}_{s},\mathcal{I}_{d})$ via dedicated encoders $\mathcal{E}_{\mathcal{H}}$ and $\mathcal{E}_{\mathbf{v}}$ , which extract modality-adapted latent features. At each diffusion step $t$ , $\mathbf{s}_{\bm{\theta}}$ estimates $\mathcal{H}_{t}$ , and $\mathbf{s}_{\bm{\phi}}$ predicts $\mathbf{v}_{t}$ conditioned on both input images and the current estimate of $\mathcal{H}_{t}$ . This cyclic interaction allows $\mathcal{H}_{t}$ to guide the estimation of $\mathbf{v}_{t}$ , while the reverse influence is incorporated via a guidance term during the update of $\mathcal{H}_{t}$ . The final aligned image $\hat{\mathcal{I}}_{s}$ is obtained by sequentially applying the predicted global transformation $\mathcal{H}$ and local deformation $\mathbf{v}$ to $\mathcal{I}_{s}$ through Spatial Transformer Layers (STL), adapted from [ 20 ] .

Figure 3: Architectural details of the network components in the homography estimation path. $\mathcal{E}_{\mathcal{H}}$ and $\mathbf{s}_{\bm{\theta}}$ first estimate the homography parameters $\mathcal{H}_{t}$ .

Figure 4: Architectural details of the network components in the displacement field estimation path. $\mathcal{E}_{\mathbf{v}}$ and $\mathbf{s}_{\bm{\phi}}$ then estimate the displacement field parameters $\mathbf{v}_{t}$ , while STL generates the warped image $\hat{\mathcal{I}}_{s}$ .

Symbol	Description
$\mathcal{I}_{s}$ , $\mathcal{I}_{d}$	Source and destination images ( Input )
$\mathbf{x}_{s}$ , $\mathbf{x}_{d}$	Pixel grids of source and destination images
$\mathcal{H}$	Homography parameters
$\mathbf{v}$	Pixel-wise displacement field
$W(\cdot;\mathcal{H})$	Grid warping function using homography
$\mathcal{E}_{\mathcal{H}}$ , $\mathcal{E}_{\mathbf{v}}$	Encoders for homography and displacement

Table 1: Summary of Notations Used in ADM

Figure 5: Score-based Iterative Alignment. ADM progressively predicts the global transform and local deformations to align SFI-UWFI pairs.

Figure 6: Qualitative comparisons of direct homography estimation methods using sample images from the KBSMC dataset. We illustrate alignment results for SFI-UWFI pairs with GLAMPoints [ 4 ] , NCNet [ 37 ] , RigidIRNet [ 56 ] , ISTN [ 55 ] , SuperRetina [ 1 ] , GeoFormer [ 2 ] , and ADM (ours).

Figure 7: Qualitative comparisons of iterative homography estimation methods on sample images from the KBSMC dataset. Alignment results between SFI-UWFI pairs are illustrated using DLKFM [ 42 ] , MCNet [ 45 ] , and ADM (ours).

Methods	Success Rate (%)			mAUC
Methods	Failed	Acceptable	Inaccurate	mAUC
SIFT [ 28 ]	0	8.29	91.71	5.2
SuperPoint [ 32 ]	0	9.09	90.91	8.7
GLAMpoints [ 4 ]	0	9.89	90.11	8.4
NCNet [ 37 ]	0	12.30	87.70	9.6
RigidIRNet [ 56 ]	0	12.57	87.43	10.6

Table 2: Comparative evaluation of the KBSMC dataset.

Figure 8: Qualitative evaluation of ADM on the FIRE [ 67 ] dataset.

Methods	Success Rate (%)			mAUC
Methods	Failed	Acceptable	Inaccurate	mAUC
SIFT [ 28 ]	0	79.85	20.15	57.3
SuperPoint [ 32 ]	0	94.78	5.22	67.4
GLAMpoints [ 4 ]	0	92.54	7.46	61.1
NCNet [ 37 ]	0	85.82	14.18	61.2
SuperRetina [ 1 ]	0	98.51	1.49	75.5

Table 3: Comparative evaluation of the FIRE [ 67 ] dataset.

Methods			FIRE [ 67 ] $(n=134)$
Methods		KBSMC $(n=374)$
Methods		Failed	Acceptable	Inaccurate	mAUC
ADM : full		0	41.98	58.02	29.3
	without Iterative ADM	0	39.84	60.16	27.8
	without $\nabla_{\mathcal{H}_{t}}\mathcal{L}^{a}_{\mathbf{x}}$ guidance	0	31.82	68.18	17.1
Methods		Failed	Acceptable	Inaccurate	mAUC

Table 4: Ablative evaluation on inference strategy.

$\delta_{\mathbf{s}}$	$\delta_{\mathbf{x}}$	$\delta_{\text{R}}$	KBSMC $(n=374)$
$\delta_{\mathbf{s}}$	$\delta_{\mathbf{x}}$	$\delta_{\text{R}}$	Failed	Acceptable	Inaccurate	mAUC
✓	✓	✓	0	41.98	58.02	29.3
✓	✗	✗	0	37.43	62.57	26.6
✗	✓	✗	0	34.60	65.40	20.9
✗	✗	✓	0	36.63	63.37	21.1
✗	✗	✗	0	34.22	65.78	20.5

Table 5: Ablative evaluation on dynamic scheduling.

Methods	KBSMC $(n=374)$
Methods	Failed	Acceptable	Inaccurate	mAUC
Transformer $\mathbf{s}_{\bm{\theta}}$ + CNN $\mathbf{s}_{\bm{\phi}}$	0	41.98	58.02	29.3
Transformer $\mathbf{s}_{\bm{\theta}}$ + Transformer $\mathbf{s}_{\bm{\phi}}$	0	38.77	69.23	28.5
CNN $\mathbf{s}_{\bm{\theta}}$ + CNN $\mathbf{s}_{\bm{\phi}}$	0	34.49	65.61	21.4
CNN $\mathbf{s}_{\bm{\theta}}$ + Transformer $\mathbf{s}_{\bm{\phi}}$	0	31.55	68.45	18.5
Methods		FIRE [ 67 ] $(n=134)$

Table 6: Ablative evaluation on network architecture.

Percentage of train/test samples	90/10	80/20	70/30	60/40	50/50
mAUC	29.3	28.9	26.5	22.7	16.3

Table 7: Ablative evaluation of the test sample ratio on the KBSMC dataset.

Figure 9: Abalative evaluation of ADM per different sampling steps and hyperparameters.

Degradations		mAUC
Degradations		KBSMC	FIRE [ 67 ]
Gaussian noise	$\sigma=5$	28.9	75.7
	$\sigma=10$	26.1	72.6
	$\sigma=25$	22.5	68.2
Gaussian blur	$\sigma=1$	29.0	75.9
	$\sigma=2.5$	27.8	73.8

Table 8: Ablative evaluation on degradations.

Figure 10: Failure cases. The transformation estimation results of ADM and the baseline GeoFormer [ 2 ] are presented on two highly challenging registration samples from the KBSMC dataset.

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

Particle Diffusion Matching: Random Walk Correspondence Search for the Alignment of Standard and Ultra-Widefield Fundus Images
Apr 11, 2026Kanggeon Lee, Soochahn Lee, Kyoung Mu LeeMedical Image Registration Cross-Modal Image Registration
ASRI, Dept. of ECE, Seoul National University, Korea · School of Electrical Engineering, Kookmin University, Korea
Modality-Invariant Coarse-to-Fine Retinal Image Registration
Aug 14, 2026Bo Wen, Nehal Nailesh Mehta, Melanie Tran +3Medical Image Registration Cross-Modal Image Registration
Diff-RF: Mutually Reinforced Image Registration and Fusion via Degradation-Aware Learning
Sep 23, 2026Xunpeng Yi, Zaixi Du, Qinglong Yan +3Multi-Source Image Fusion Cross-Modal Image Registration
Electronic Information School, Wuhan University, Wuhan 430072, China

Apr 11, 2026J/K move · Enter open · S save

Particle Diffusion Matching: Random Walk Correspondence Search for the Alignment of Standard and Ultra-Widefield Fundus Images

Kanggeon Lee, Soochahn Lee, Kyoung Mu Lee

ASRI, Dept. of ECE, Seoul National University, Korea · School of Electrical Engineering, Kookmin University, Korea

Medical Image Registration Cross-Modal Image Registration Fundus Imaging Cross-Modal Feature Matching

Open paper AskPDF arXiv

We propose a robust alignment technique for Standard Fundus Images (SFIs) and Ultra-Widefield Fundus Images (UWFIs), which are challenging to align due to differences in scale, appearance, and the scarcity of distinctive features. Our method, termed Particle Diffusion Matching (PDM), performs alignment through an iterative Random Walk Correspondence Search (RWCS) guided by a diffusion model. At each iteration, the model estimates displacement vectors for particle points by considering local appearance, the structural distribution of particles, and an estimated global transformation, enabling progressive refinement of correspondences even under difficult conditions. PDM achieves state-of-the-art performance across multiple retinal image alignment benchmarks, showing substantial improvement on a primary dataset of SFI-UWFI pairs and demonstrating its effectiveness in real-world clinical scenarios. By providing accurate and scalable correspondence estimation, PDM overcomes the limitations of existing methods and facilitates the integration of complementary retinal image modalities. This diffusion-guided search strategy offers a new direction for improving downstream supervised learning, disease diagnosis, and multi-modal image analysis in ophthalmology.

Abstract

Medical Image Registration

Explore similar work

Particle Diffusion Matching: Random Walk Correspondence Search for the Alignment of Standard and Ultra-Widefield Fundus Images

Modality-Invariant Coarse-to-Fine Retinal Image Registration

Diff-RF: Mutually Reinforced Image Registration and Fusion via Degradation-Aware Learning