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

Kanggeon Lee, Soochahn Lee, Kyoung Mu LeearXiv 2026 · arXiv

TL;DR

Particle Diffusion Matching (PDM) is a novel image registration framework that aligns standard and ultra-widefield fundus retinal images by framing keypoint matching as a physics-inspired reverse diffusion process. By treating corresponding point pairs as coupled entities and optimizing them through an iterative "random walk," PDM achieves a massive 17.65% absolute improvement in alignment accuracy over previous state-of-the-art methods while maintaining a low inference time of 0.45 seconds.

What Problem Does This Paper Address?

In ophthalmology, doctors rely on two main types of retinal photographs: Standard Fundus Images (SFIs), which offer high-resolution but narrow views (30° to 60°), and Ultra-Widefield Fundus Images (UWFIs), which capture up to 200° of the retina but often compromise on peripheral clarity. Combining these modalities is crucial for accurate disease diagnosis, but aligning them automatically is a nightmare for computer vision models.

The challenge stems from extreme differences in scale, field of view, and local distortions. Traditional point-matching methods fail because distinctive textures are sparse, leading to high outlier rates. Conversely, local deformation models fail to handle the massive global perspective shifts between the two imaging modalities. Existing two-step cascade methods (global alignment followed by local refinement) cannot recover if the initial global estimation step fails.

Methods

The authors introduce Particle Diffusion Matching (PDM), which unifies global and local alignment into a single, stochastic, iterative search process.

Instead of making discrete matching decisions, PDM treats keypoint matching as a Denoising Diffusion Probabilistic Model (DDPM) problem:

  1. Query Generation: Keypoints are detected only on the clean, texture-rich source image (SFI).
  2. Particle Initialization: Corresponding points on the destination image (UWFI) are initialized as random Gaussian noise, termed "particles."
  3. Random Walk Correspondence Search (RWCS): The model executes a reverse diffusion process. At each timestep, a trained neural network ($S_\theta$) predicts displacement vectors, guiding the particles to "walk" toward their true matching positions on the destination image.
Overview of the iterative Random Walk Correspondence Search (RWCS) process.

To make this computationally efficient, PDM uses a Coarse-to-Fine Random Walker Transformer (RWTr) architecture. The coarse level handles large global misalignments, while the fine level crops localized feature patches around the coarse estimates to resolve fine-grained, non-rigid deformations.

The network is trained end-to-end using a combination of three losses: a standard DDPM denoising loss, a pixel matching loss (maximizing normalized cross-correlation), and a homography regularization loss computed via a differentiable RANSAC algorithm.

The detailed architecture of PDM, showcasing the coarse-to-fine refinement strategy.

Datasets

The model was trained and evaluated across several diverse datasets:

  • KBSMC Dataset: The primary dataset, consisting of 3,744 clinical SFI-UWFI pairs from Kangbuk Samsung Medical Center.
  • FIRE & FLORI21: Public retinal datasets used to evaluate intra-modality generalization (SFI-SFI and UWFI-UWFI alignment, respectively).
  • HPatches, MegaDepth, & ScanNet: Standard computer vision benchmarks featuring planar, outdoor, and indoor scenes, used to test the model's generalizability outside of medical imaging.

Key Insights and Ideas

  • Continuous Refinement over Discrete Matching: Traditional matchers try to find the single best match from a discrete set of candidates, which easily fails when features are sparse. PDM treats matching as a continuous trajectory optimization, allowing points to gradually correct their paths over multiple diffusion steps.
  • Coupled Optimization: Each particle maintains an explicit source-to-destination bond throughout the search. Because the transformer updates all particles simultaneously, the model inherently considers both individual match quality and global geometric coherence, drastically reducing the impact of outliers.
  • Robustness to Detector Choice: Ablation studies revealed that PDM is highly robust to how the initial source keypoints are found (e.g., SIFT, SuperPoint, or KAZE). What actually matters is that the keypoints are well-distributed across the image structure rather than clustered in one region.

Key Results

PDM achieves state-of-the-art performance across all evaluated retinal benchmarks:

  • SFI-UWFI Alignment (KBSMC): PDM achieved a 58.56% Acceptable alignment rate, outperforming the previous best method, MINIMA (40.91%), by an absolute margin of 17.65%, and doubling the performance of traditional transformer-based matchers like LoFTR (26.20%).
  • Efficiency: Despite its iterative nature, PDM boasts an inference time of just 0.45 seconds per image pair. This efficiency is achieved because the diffusion process operates on a sparse set of 100 keypoint particles rather than a dense pixel grid.
  • Zero-Shot Generalization: When trained on KBSMC and evaluated directly on the unseen FIRE and FLORI21 datasets, PDM maintained top-tier performance without any fine-tuning.
Qualitative comparison showing PDM's superior ability to align retinal blood vessels compared to baseline methods.

Limitations and Future Work

While powerful, PDM has a few notable limitations:

  • Source Keypoint Dependency: Because particles are conditioned on source keypoints, the method can fail if the source image is so severely degraded or blurred that reliable initial keypoints cannot be extracted.
  • Fixed Sampling Steps: The current implementation uses a fixed 100-step sampling process for every image pair, which can lead to over-computation for simple alignments and insufficient refinement for highly complex ones.
  • 3D Scene Limitations: When tested on natural 3D scene datasets (like MegaDepth), PDM's performance dropped slightly. This is likely because its regularization loss assumes a planar homography, which fails to capture complex 3D geometric relationships.

The authors note that future work will focus on implementing adaptive sampling strategies and integrating learnable feature extractors to handle highly degraded images.

Summary updated

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.
BibTeX
@misc{lee2026particle,
  author = {Kanggeon Lee and Soochahn Lee and Kyoung Mu Lee},
  title = {{Particle Diffusion Matching: Random Walk Correspondence Search for the Alignment of Standard and Ultra-Widefield Fundus Images}},
  year = {2026},
  eprint = {2604.10085},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2604.10085},
}

Loading related content...

Figure