💻 computer science

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

This paper proposes Active Diffusion Matching (ADM), a novel score-based iterative method that effectively aligns Standard Fundus and Ultra-Widefield Retinal Images by jointly estimating global and local transformations, thereby achieving state-of-the-art accuracy and enabling improved clinical analysis.

Original authors: Kanggeon Lee, Su Jeong Song, Soochahn Lee, Kyoung Mu Lee

Published 2026-04-15

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Kanggeon Lee, Su Jeong Song, Soochahn Lee, Kyoung Mu Lee

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

The Big Problem: The "Wide Shot" vs. The "Zoom Shot"

Imagine you are looking at a map of a city.

The Ultra-Widefield Image (UWFI) is like a satellite photo taken from space. You can see the whole city, all the neighborhoods, and the rivers. However, because it's so wide, the buildings look tiny and blurry. You can't read the street signs.
The Standard Fundus Image (SFI) is like a high-resolution photo taken with a zoom lens right in front of a specific building. The details are crystal clear, but you can only see that one building. You have no idea where it is in the city.

The Goal: Doctors need to combine these two. They want to take the clear details from the "zoom shot" and paste them onto the "satellite photo" so they can see exactly where the problem is in the whole eye.

The Challenge: Doing this manually is a nightmare. The two photos are different sizes, different shapes, and the "roads" (blood vessels) look different in each. Existing computer programs try to match them but often get confused, like trying to fit a square peg into a round hole.

The Solution: "Active Diffusion Matching" (ADM)

The authors created a new AI method called Active Diffusion Matching (ADM). To understand how it works, let's use a few analogies.

1. The "Blind Sculptor" Analogy (How Diffusion Works)

Imagine you have a block of clay (the image) and you want to sculpt it to match a specific shape (the target).

Old AI methods try to guess the final shape in one giant leap. If they guess wrong, the whole thing fails.
ADM works like a blind sculptor who starts with a messy lump of clay.
- The sculptor doesn't know the final shape perfectly.
- Instead, they make a tiny, small adjustment.
- Then they step back, feel the clay again, and make another tiny adjustment.
- They repeat this hundreds of times. With every small step, the clay gets closer to the perfect shape.
- In the paper, this "step-by-step" process is called a diffusion process. It starts with "noise" (confusion) and slowly refines it into a clear alignment.

2. The "Two-Handed Dance" (Global vs. Local)

Aligning these eye images requires two types of movement:

Global Movement: Moving the whole image (rotating it, zooming it in/out, shifting it left or right).
Local Movement: Warping specific parts (bending a blood vessel slightly because the eye isn't perfectly flat).

Most old methods try to do these separately, like trying to dance with one hand tied behind your back.

ADM uses two dancers (two AI models) working together.
- Dancer A handles the big moves (Global).
- Dancer B handles the tiny, detailed moves (Local).
- They hold hands. Dancer A tells Dancer B, "We moved left, so you need to adjust your feet." Then Dancer B says, "My feet are adjusted, so you need to tilt your body a bit more."
- They keep talking to each other in a loop until they are perfectly in sync.

3. The "GPS with a Compass" (Active Guidance)

Sometimes, the AI gets stuck or makes a wrong turn.

Old AI: Just keeps walking in the direction it thinks is right, even if it's leading off a cliff.
ADM: Has a compass (called "Active Guidance"). Every time it makes a move, it checks: "Does this look like the target image?" If the answer is "No," the compass nudges the AI back on the right path immediately. This ensures the AI doesn't get lost in the middle of the process.

Why is this a Big Deal?

It's the First of Its Kind: Before this, no computer program could automatically and accurately stitch these two specific types of eye photos together. Doctors had to do it by hand, which was slow and prone to error.
It's Smarter: The paper tested ADM against other top AI methods. ADM won every time, especially on the hardest, most blurry images.
- The Result: It improved the accuracy score by a significant margin (about 5 points higher than the next best method).
It Saves Lives: By perfectly aligning these images, doctors can use AI to enhance the blurry wide-angle photos. This means they can detect diseases like diabetic retinopathy earlier and more accurately, potentially saving a patient's vision.

The Trade-off (The "Price" of Perfection)

There is one downside: Speed.
Because ADM takes hundreds of tiny steps to get the alignment perfect (like the sculptor slowly chipping away at the clay), it takes about 47 seconds to process one image. Other faster methods take only 1 or 2 seconds but make more mistakes.

The authors argue that in medicine, accuracy is more important than speed. It's better to wait an extra minute to get a diagnosis that is 100% correct than to get a fast diagnosis that is wrong.

Summary

Active Diffusion Matching is a new, super-smart AI that acts like a patient, step-by-step sculptor. It takes a blurry, wide view of an eye and a sharp, zoomed-in view, and slowly, carefully, and cooperatively warps them until they fit together perfectly. This allows doctors to see the whole picture with crystal-clear detail, leading to better eye care for patients.

1. Problem Statement

The paper addresses the critical challenge of aligning Standard Fundus Images (SFIs) and Ultra-Widefield Fundus Images (UWFIs).

The Gap: SFIs capture a central field of view (30°–60°, <20% of the retina) with high resolution, while UWFIs capture up to 200° (~82% of the retina) but with lower resolution and clarity.
The Need: To enhance UWFIs using machine learning (e.g., super-resolution) or to integrate them for comprehensive diagnosis, accurate pixel-level alignment between SFI-UWFI pairs is required.
The Difficulty: Existing alignment methods (designed for SFI-SFI pairs) fail on SFI-UWFI pairs due to:
- Extreme Scale and FOV Differences: The massive disparity in field of view.
- Domain Shift: Significant differences in color characteristics and image quality.
- Lack of Distinctive Features: Retinal textures are often amorphous, making keypoint-based matching unreliable.
- Absence of Specialized Methods: No prior automated method exists specifically for this cross-modal task; previous attempts relied on manual intervention.

2. Methodology: Active Diffusion Matching (ADM)

The authors propose Active Diffusion Matching (ADM), a novel framework that integrates two interdependent score-based diffusion models to perform iterative alignment via a Langevin Markov chain.

Core Architecture

ADM treats alignment as a stochastic, progressive search for optimal parameters, consisting of two coupled paths:

Global Transformation Path ( $s_\theta$ ): Estimates the Homography ( $H$ ) to handle global scale, rotation, and perspective differences.
- Architecture: Uses a Vision Transformer (initialized with DINO) as an encoder ( $E_H$ ) and a Transformer-based network for the score function.
- Role: Captures global context, which is crucial given the massive FOV differences.
Local Deformation Path ( $s_\phi$ ): Estimates the Pixel-wise Displacement Field ( $v$ ) to handle non-linear local deformations.
- Architecture: Uses a U-Net based structure with a vessel enhancement filter as the encoder ( $E_v$ ).
- Role: Refines local details, leveraging the fact that CNNs excel at extracting local features.

The Iterative Process

The method operates in a reverse diffusion process where the models iteratively refine estimates from noisy initial states ( $H_T, v_T$ ) to clean parameters ( $H_0, v_0$ ):

Conditional Interaction: The global estimate $H_t$ guides the local estimation $v_t$ . Conversely, the local estimation path provides feedback to correct the global estimate.
Spatial Transformer Layers (STL): The estimated $H$ and $v$ are applied to the source image ( $I_s$ ) to generate a warped image ( $\hat{I}_s$ ) for comparison with the destination ( $I_d$ ).

Key Technical Innovations

Input-Adaptive Guided Sampling: During inference, the global homography estimate is adjusted using the gradient of a pixel-matching loss (Normalized Cross-Correlation) computed between the warped image and the destination. This allows the model to adapt dynamically to specific input pairs rather than relying solely on training distributions.
Dynamic Scheduling of Losses: The training loss function ( $L = L_s + \lambda_x L_x + \lambda_R L_R$ $L = L_{s} + λ_{x} L_{x} + λ_{R} L_{R}$ ) employs time-dependent weights:
- $\delta_s$ : Suppresses the influence of the local loss early in training to prevent unstable homography estimation.
- $\delta_x$ (Pixel Loss): Increases over time to enforce appearance consistency as alignment improves.
- $\delta_R$ (Regularization): Decreases over time, penalizing large deviations from identity mapping early on to ensure stability.
Iterative Inference: The output of one alignment pass is fed back as the input for the next, allowing for progressive refinement of severely misaligned pairs.

3. Key Contributions

First Automated SFI-UWFI Aligner: ADM is the first fully automatic, accurate method designed specifically for aligning standard and ultra-widefield retinal images.
Joint Global-Local Diffusion: Unlike previous diffusion methods that estimate local or global variations separately, ADM uses a dual-score model to jointly optimize both via a coupled Langevin dynamics process.
Adaptive Sampling Strategy: Introduces a guidance mechanism that uses the gradient of the pixel-matching loss to correct global homography estimates during inference, significantly improving robustness to domain shifts.
Architecture Specialization: Demonstrates that a Transformer-based architecture is superior for global homography estimation, while a CNN-based architecture is superior for local displacement fields in this specific medical imaging context.

4. Experimental Results

The method was evaluated on two datasets:

Private Dataset (KBSMC): 374 pairs of SFI-UWFI (challenging cross-modal alignment).
Public Dataset (FIRE): 134 pairs of SFI-SFI (standard alignment benchmark).

Performance Metrics:

KBSMC (SFI-UWFI): ADM achieved a Mean AUC (mAUC) of 29.3, outperforming the second-best method (GeoFormer, 24.1) by 5.2 points. The "Acceptable Rate" (MAE < 50, MEE < 20) was 41.98%, a significant improvement over baselines.
FIRE (SFI-SFI): ADM achieved an mAUC of 76.0, matching or slightly exceeding state-of-the-art methods like SuperRetina and GeoFormer, proving its versatility even on easier, single-modal tasks.
Qualitative Results: Visual comparisons showed ADM successfully aligned vessel structures in regions where other methods (SIFT, SuperRetina, GeoFormer) failed completely or produced high Maximum Alignment Errors (MAE).

Ablation Studies:

Removing the iterative refinement or the guided sampling caused significant performance drops, validating the necessity of the dual-diffusion and feedback loop.
Sampling Steps: Increasing steps for the global estimator ( $s_\theta$ ) improved performance significantly, while excessive steps for the local estimator ( $s_\phi$ ) led to diminishing returns due to over-smoothing.

5. Significance and Future Impact

Clinical Utility: By enabling precise alignment, ADM facilitates the use of machine learning to enhance the resolution of UWFIs using high-quality SFIs as ground truth. This could allow UWFIs to eventually replace SFIs in clinical workflows, offering both wide-field coverage and high-resolution detail.
Diagnostic Improvement: Improved alignment supports better detection of peripheral pathologies (e.g., diabetic retinopathy, vascular occlusions) and central diseases (e.g., macular degeneration) simultaneously.
Generalizability: The framework's reliance on anatomical structures (vessels) suggests potential applicability to other cross-modal medical imaging tasks beyond retinal fundus images.
Limitations & Future Work: The current inference time (~47 seconds) is slower than feedforward models. Future work aims to accelerate inference via knowledge distillation or one-step diffusion models and to improve robustness against severe image degradation (blur, low illumination).

In conclusion, Active Diffusion Matching represents a significant leap forward in medical image registration, successfully bridging the gap between high-resolution and wide-field retinal imaging through a sophisticated, iterative, score-based diffusion framework.