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

Kanggeon Lee; Soochahn Lee; Kyoung Mu Lee

cs.CVApr 11, 2026

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

Authors: Kanggeon Lee, Soochahn Lee, Kyoung Mu Lee

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

Abstract

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.

Figures & tables

Figure 1: Random Walk Correspondence Search (RWCS). Particles are defined as pairs of points from SFI and UWFI, respectively. Particle random walks at sample steps $t+1$ , $t$ , and $t-1$ are depicted. Particle with correct and incorrect correspondences are marked by green and red lines, respectively. The bottom row depicts particles at one-third intervals of the entire sampling process.

Figure 2: Particle Diffusion Matching ( PDM ). PDM can align Standard Fundus Images (SFIs) and Ultra-Widefield Fundus Images (UWFIs), which may have extreme differences in view and scale (The resolutions of UWFI and SFI here are approximately $2600\times 2000$ and $3600\times 3600$ , respectively.). In PDM , a particle is defined as a point pair comprising a keypoint from the SFI and a potential matching point from the UWFI. UWFI points are initialized as random Gaussian points at $t=T$ , (bottom left) and undergo random walk correspondence search (RWCS) to eventually reach the matching points (bottom middle) to the paired SFI points (bottom right). The random walk trajectories of four sample particles have been highlighted in various colors.

Figure 3: The architecture of PDM . Given a source image $I_{s}$ from SFI and a destination image $I_{d}$ from UWFI, the random walker $\mathbf{S}_{\theta}$ iteratively refines the correspondence point set $\mathbf{x}_{t}^{d}$ by computing the displacement $\Delta\mathbf{x}_{t}^{d}$ corresponding to the query point set $\mathbf{x}^{s}$ , which is generated by the key point detector $\mathcal{D}$ . This process starts from initial random Gaussian points $\mathbf{x}_{T}^{d}$ and progressively improves the alignment.

Figure 4: Detailed structure of $\mathbf{S}_{\theta}$ . Two RWTr s with the same structure form $\mathbf{S}_{\theta}$ , where the estimated displacement of the coarse level corresponding points $\mathbf{\tilde{x}}_{t}^{d}$ for the coarse-level features $C_{s}$ and $C_{d}$ contributes to the local feature selection of the fine-level features $F_{s}^{crop}$ and $F_{d}^{crop}$ extracted from $F_{s}$ and $F_{d}$ , ultimately estimating the displacement of the fine level corresponding points $\mathbf{x}_{t}^{d}$ .

Figure 5: Qualitative evaluation of PDM on the KBSMC, FIRE, and FLORI21 datasets. For input image pairs with significantly different styles, PDM robustly identify sufficient matches compared to SuperRetina [ 10 ] , GeoFormer [ 13 ] and RetinaRegNet [ 22 ] . The red box indicates the zoomed-in overlay region between SFI and UWFI for the KBSMC dataset. For the FIRE and FLORI21 datasets, correspondence matches are visualized with green lines (accurate matches) and red lines (inaccurate matches), while the blue boxes show magnified regions for verifying vessel alignment.

Methods	KBSMC		FIRE		FLORI21
Methods	Acceptable (%, $\uparrow$ )	mAUC ( $\uparrow$ )	Acceptable (%, $\uparrow$ )	mAUC ( $\uparrow$ )	Acceptable (%, $\uparrow$ )	mAUC ( $\uparrow$ )
SuperPoint [ 7 ]	9.09 ∗	8.7 ∗	94.78 ∗	67.3 ∗	40 ∗	39.1 ∗
GLAMpoints [ 8 ]	9.89 ∗	8.4 ∗	93.28 ∗	61.9 ∗	33.33 ∗	34.4 ∗
ISTN [ 21 ]	20.86 ∗	12.1 ∗	86.57 ∗	60.9 ∗	53.33 ∗	52.5 ∗
NCNet [ 39 ]	12.30 ∗	9.6 ∗	86.57 ∗	61.4 ∗	53.33 ∗	50.8 ∗
SuperGlue [ 9 ]	24.06 ∗	15.3 ∗	95.52 ∗	68.7 ∗	80 ∗∗∗	59.8 ∗

Table 1: Quantitative evaluation on KBSMC, FIRE, and FLORI21 dataset.

Methods	Inference time	Memory (model / inference)
MCNet [ 14 ]	35.76s	1692MB / 2280MB
GeoFormer [ 13 ]	1.15s	242MB / 918MB
RetinaRegNet [ 22 ]	0.09s	4826MB / 5415MB
PDM (ours)	0.45s	1132MB / 1721MB

Table 2: Computational cost comparison

Particle Count	mAUC ( $\uparrow$ )
Particle Count	KBSMC	FIRE	FLORI21
4	24.2±1.8	65.8±1.5	71.3±1.9
50	34.5±1.2	76.2±1.1	85.2±1.4
100	34.8±1.3	78.2±0.9	92.3±1.2
200	34.8±1.2	78.1±0.8	92.4±1.1
500	34.8±1.1	78.2±0.9	92.2±1.0

Table 3: Ablation study of different particle configurations.

RWCS Network Configurations	mAUC ( $\uparrow$ )
RWCS Network Configurations	KBSMC	FIRE	FLORI21
w/ Coarse-Level RWTr	32.1±1.4	74.2±1.2	88.6±1.3
w/ Fine-Level RWTr	32.9±1.5	75.0±1.1	90.8±1.2
w/ Coarse-to-Fine RWTr s	34.8±1.3	76.2±0.9	92.3±1.2

Table 4: Ablation study of RWCS Network Variants.

Query Generators	mAUC ( $\uparrow$ )
Query Generators	KBSMC	FIRE	FLORI21
SIFT [ 5 ] + Blob [ 72 ]	34.8±1.3	78.2±0.9	92.3±1.2
SURF [ 6 ]	34.3±1.4	77.9±1.0	91.9±1.3
FAST [ 81 ]	32.0±1.5	75.1±1.2	87.3±1.4
Haris Corner Detector [ 82 ]	30.9±1.6	72.3±1.3	85.8±1.5
KAZE [ 83 ]	34.7±1.3	78.1±0.9	92.4±1.1

Table 5: Ablation study of different query generator.

Particle Range	mAUC ( $\uparrow$ )
Particle Range	KBSMC	FIRE	FLORI21
$[-0.1,0.1]$	30.0±1.5	73.1±1.3	84.9±1.4
$[-1,1]$	34.8±1.3	78.2±0.9	92.3±1.2
$[-10,10]$	33.2±1.4	76.0±1.1	88.8±1.3
$[-100,100]$	25.4±1.7	65.1±1.5	70.9±1.6
$[-384,384]$ (full resolution)	23.1±1.9	61.8±1.7	64.6±1.8

Table 6: Ablation study of different particle initialization.

Sampling Steps	mAUC ( $\uparrow$ )
Sampling Steps	KBSMC	FIRE	FLORI21
50	34.2±1.4	77.8±1.0	91.5±1.3
100	34.8±1.3	78.2±0.9	92.3±1.2
250	34.7±1.3	78.2±0.9	92.2±1.2
500	33.2±1.5	76.9±1.1	90.8±1.4
1000	30.8±1.6	75.9±1.3	89.4±1.5

Table 7: Ablation study of different sampling steps.

Methods	mAUC ( $\uparrow$ )
Methods	KBSMC	FIRE	FLORI21
GeoFormer [ 13 ]	24.1	70.3	65.9
ROMA [ 44 ]	23.1	73.3	69.42
MINIMA [ 31 , 44 ]	27.6	75.1	87.5
PDM (ours)	34.8±1.3	75.8±1.2	87.9±1.3

Table 8: Ablation study of zero-shot evaluation.

Figure 6: Extreme cases. The transformation estimation results of PDM are presented on synthesized less-overlapping pair (left) and rotated pair (right).

Figure 7: Partial correspondence. The transformation estimation results of PDM are presented on two intentionally selected textureless regions to demonstrate robustness.

Figure 8: Failure cases. The transformation estimation results of PDM and the baseline GeoFormer [ 13 ] are presented on two highly challenging registration samples from the KBSMC dataset.

Figure 9: Qualitative evaluation of PDM on the HPatches, MegaDepth, and ScanNet datasets. Correspondence matches are visualized with green lines (accurate matches) and red lines (inaccurate matches).

Methods	mAUC $\uparrow$
Methods	HPatches	MegaDepth	ScanNet
LoFTR [ 32 ]	75.4	67.7	40.7
LightGlue [ 37 ]	77.5	72.3	48.2
RoMa [ 44 ]	78.4	75.2	52.0
PDM (ours)	77.0	72.4	46.5

Table 9: Quantitative Evaluation on Scene Datasets.

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

Apr 11, 2026

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

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.

Kanggeon Lee, Su Jeong Song, Soochahn Lee +1

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

Medical Image Registration Cross-Modal Image Registration Diffusion Model Alignment Deformable Image Registration Fundus Imaging

Ask PDF

Feb 26, 2026

DMAligner: Enhancing Image Alignment via Diffusion Model Based View Synthesis

Image alignment is a fundamental task in computer vision with broad applications. Existing methods predominantly employ optical flow-based image warping. However, this technique is susceptible to common challenges such as occlusions and illumination variations, leading to degraded alignment visual quality and compromised accuracy in downstream tasks. In this paper, we present DMAligner, a diffusion-based framework for image alignment through alignment-oriented view synthesis. DMAligner is crafted to tackle the challenges in image alignment from a new perspective, employing a generation-based solution that showcases strong capabilities and avoids the problems associated with flow-based image warping. Specifically, we propose a Dynamics-aware Diffusion Training approach for learning conditional image generation, synthesizing a novel view for image alignment. This incorporates a Dynamics-aware Mask Producing (DMP) module to adaptively distinguish dynamic foreground regions from static backgrounds, enabling the diffusion model to more effectively handle challenges that classical methods struggle to solve. Furthermore, we develop the Dynamic Scene Image Alignment (DSIA) dataset using Blender, which includes 1,033 indoor and outdoor scenes with over 30K image pairs tailored for image alignment. Extensive experimental results demonstrate the superiority of the proposed approach on DSIA benchmarks, as well as on a series of widely-used video datasets for qualitative comparisons. Our code is available at https://github.com/boomluo02/DMAligner.

Xinglong Luo, Ao Luo, Zhengning Wang +5

University of Electronic Science and Technology of China · Qianyuan Laboratory, Hangzhou, China · Southwest Jiaotong University +1

Topic pending

Ask PDF

Aug 14, 2026

Modality-Invariant Coarse-to-Fine Retinal Image Registration

Retinal image registration is essential for ophthalmic diagnosis, longitudinal disease monitoring, and multimodal retinal image analysis. Existing retinal registration methods are typically modality-dependent: they are designed or optimized either for a single imaging modality in mono-modal registration or for a fixed pair of modalities in cross-modal registration. This limits their flexibility and applicability in practical scenarios involving diverse retinal imaging modalities and different combinations of them. In this work, we propose a generalizable two-stage, modality-invariant framework for retinal image registration. First, we introduce a sparse feature-matching model driven by a universal retinal vessel segmentation to achieve robust coarse global alignment across modalities. Second, we develop a modality-invariant optical flow estimation network, termed MI-RAFT, to refine the alignment through dense local registration. Extensive experiments demonstrate that the proposed method can handle diverse combinations of commonly used retinal imaging modalities, exhibiting strong modality invariance while outperforming state-of-the-art modality-dependent registration methods.

Bo Wen, Nehal Nailesh Mehta, Melanie Tran +3

Medical Image Registration Cross-Modal Image Registration Deformable Image Registration Optical Flow

Ask PDF

Abstract

Medical Image Registration

Explore similar work

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

DMAligner: Enhancing Image Alignment via Diffusion Model Based View Synthesis

Modality-Invariant Coarse-to-Fine Retinal Image Registration