🔬 optics

Improving Richardson--Lucy Deconvolution with Diffusion Priors for Fluorescence Microscopy

This paper proposes a novel fluorescence microscopy deconvolution framework that integrates a score-based diffusion prior into the Richardson--Lucy optimization process to effectively suppress noise amplification and preserve fine biological structures under low-photon conditions, overcoming the limitations of traditional regularizers.

Original authors: Hao Chen, Scott S. Howard

Published 2026-06-25

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Hao Chen, Scott S. Howard

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

Imagine you are trying to look at a tiny, intricate city inside a living cell using a microscope. But there's a catch: the "camera" you are using is very old and blurry (due to the laws of physics called diffraction), and it's also taking pictures in the dark, so the images are full of grainy static (noise from random photon counts).

The goal of this paper is to fix these blurry, grainy pictures to reveal the city's streets and buildings clearly.

The Old Way: Guessing and Checking

Traditionally, scientists use a method called Richardson–Lucy (RL) deconvolution. Think of this as a very smart guess-and-check game.

How it works: You have a blurry photo. You know exactly how your camera blurs things (the "Point Spread Function" or PSF). The RL algorithm tries to reverse that blur mathematically.
The Problem: If the photo is too dark (low light), the algorithm starts to "hallucinate." It tries to sharpen the image so much that it turns the grainy static into fake, jagged lines. It's like trying to sharpen a blurry photo of a foggy night until the fog looks like a detailed forest that isn't actually there.
The Old Fix: Scientists used to add a "smoothness" rule (like Total Variation) to stop the noise. But this is like smoothing out a photo with a heavy hand: it gets rid of the noise, but it also smears out the tiny, delicate details you actually wanted to see, like thin cell threads or small dots.

The New Way: The "AI Art Guide" + The "Physics Check"

The authors, Hao Chen and Scott S. Howard, created a new system called Diffusion-RL. They combined two powerful tools to solve the problem:

The "AI Art Guide" (The Diffusion Prior):
Imagine an artist who has studied millions of high-quality, perfect photos of cells. This artist has memorized what a healthy cell should look like—how the threads connect, how the dots are spaced, and how the shapes fit together. This is the Diffusion Model.
- When the system starts, this "artist" suggests a clean, plausible version of the cell based on what it has learned. It says, "I think the cell probably looks this."
The "Physics Check" (The RL Step):
Now, bring in the old RL algorithm. It looks at the "artist's suggestion" and compares it to the actual blurry, noisy photo you took.
- It asks: "Does this suggestion match the actual light particles (photons) we detected?"
- If the artist guessed a structure that doesn't match the light data, the RL step says, "No, that's not right based on our measurements," and adjusts the image.
- If the artist guessed something that matches the data but is too blurry, the RL step sharpens it.

The Magic Loop:
They don't just do this once. They run a loop:

The AI proposes a clean structure.
The Physics Check (RL) tweaks it to match the real, noisy data.
The AI looks at that tweaked version and refines it again.
They repeat this back and forth many times.

Why This is Better

No More Fake Noise: Because the AI knows what real cells look like, it doesn't get tricked by the grainy static. It knows that a jagged line in the noise isn't a real cell thread.
Saving the Tiny Details: Unlike the old "smoothness" rule that smears everything out, the AI knows that cells have fine, delicate threads. It preserves these details because it has seen them before in its training.
Working in the Dark: This method shines brightest when the photo is very dark (low photon counts). In these conditions, the raw data is too weak to tell the truth on its own. The AI fills in the gaps with "educated guesses" based on biology, while the Physics Check ensures those guesses don't drift too far from reality.

The "Uncertainty" Feature

The paper also notes something cool: because the AI starts with a bit of randomness, if you run the process multiple times on the same blurry photo, you get slightly different results.

Stable parts: The big, obvious structures (like the main cell body) stay the same in every run.
Uncertain parts: In areas where the photo is extremely noisy and the data is weak, the results might vary slightly.
The Benefit: This isn't a bug; it's a feature. It tells the scientist, "I am very confident about this part of the cell, but this tiny, fuzzy area is hard to see, so here are a few possibilities of what it might be."

Summary

The paper introduces a method that uses a trained AI to teach the deconvolution algorithm what a real cell looks like, while keeping a physics-based check to ensure the result matches the actual light data. This allows scientists to see clear, detailed images of cells even when the microscope is struggling with low light and blur, without inventing fake details or smoothing away the truth.

Technical Summary: Improving Richardson–Lucy Deconvolution with Diffusion Priors for Fluorescence Microscopy

Problem Statement

Fluorescence microscopy is essential for visualizing labeled molecules in biological systems, yet image formation is fundamentally limited by diffraction and Poisson photon shot noise. Classical Richardson–Lucy (RL) deconvolution attempts to recover the original fluorescence signal by maximizing the Poisson log-likelihood under a physical forward model. However, this inverse problem is ill-posed, particularly in low-photon regimes. Without additional constraints, iterative RL updates amplify noise and produce high-frequency artifacts, while fine biological structures (such as weak filaments or puncta) may be lost due to optical degradation.

Traditional regularization methods, such as Total Variation (TV), stabilize the solution by enforcing spatial smoothness. While effective at reducing noise, TV regularization often introduces oversmoothing, suppressing weak subcellular details and requiring careful hyperparameter tuning across different imaging conditions. Furthermore, existing deep learning approaches often rely on supervised training with paired data specific to a single microscope configuration or specimen type, limiting their generalizability. There remains a need for a method that provides robust structural guidance for RL deconvolution without sacrificing the physical consistency of the Poisson measurement model or the ability to generalize across diverse biological samples.

Methodology

The authors propose Diffusion-RL, a decoupled inverse-problem framework that integrates a score-based diffusion prior into the RL optimization loop. The core innovation lies in separating the learning of biological morphology from the enforcement of measurement consistency.

1. Decoupled Framework

The framework alternates between two distinct components:

Diffusion Prior (Structural Guidance): An unconditional Elucidating Diffusion Model (EDM) is trained on a diverse corpus of fluorescence microscopy images (including confocal, widefield, and two-photon modalities). This model learns the manifold of recurring biological structures (e.g., cellular morphology, filament networks) without being conditioned on specific microscope parameters. During inference, the diffusion model proposes a structural estimate $\hat{x}_0$ from the learned distribution.
RL Data Consistency (Physical Constraints): The proposed structural estimate is refined using the classical RL update rule. This step enforces consistency with the measured photon counts ( $y$ ) and the known Point Spread Function (PSF, $H$ ) under a Poisson noise model. The RL update acts as a modular data-consistency operator, correcting the diffusion estimate to align with the physical measurement.

2. Algorithmic Flow

The reconstruction process follows a reverse diffusion trajectory with an inner refinement loop:

Initialization: At each diffusion step $t$ , the perturbed state $x_t$ is mapped to a clean structural estimate $\bar{x}^0_t$ using the Tweedie denoising formula.
Inner RL Loop: This clean estimate serves as the initialization for $K$ iterations of the RL update (Eq. 2), which refines the image to satisfy the Poisson likelihood $p(y|x)$ .
Re-noising: The RL-refined estimate $\hat{x}^0_{t,RL}$ is re-noised along the noise schedule to produce $\hat{x}_{t,RL}$ .
Propagation: The re-noised, data-consistent state is passed to the next diffusion step. When the noise level is sufficiently low, an ODE solver is used to recover fine-scale details.

3. Training and Implementation

The foundational diffusion model was trained unconditionally on 125,464 patches from four public repositories (SR-Caco-2, FMD, Neuronal Cells, RxRx1) using a U-Net architecture. For specimen-specific validation, the model was fine-tuned on a BPAE confocal dataset. Crucially, the same pretrained prior is deployed across different PSFs and photon regimes without retraining, as the microscope-specific physics are introduced only during the RL inference step.

Key Contributions

Integration of Learned Priors with Physical Models: The paper introduces a method to combine the distribution-fitting capabilities of score-based diffusion models with the rigorous physical constraints of Poisson deconvolution.
Decoupled Inference Strategy: By separating prior learning from measurement consistency, the framework allows a single, generalizable biological prior to be applied across diverse optical configurations (different PSFs, photon counts, and modalities) without retraining.
Preservation of Weak Structures: The method specifically addresses the trade-off between noise suppression and detail preservation, aiming to recover weak, fine-scale structures that are typically lost in TV-regularized or unregularized RL approaches.
Uncertainty Quantification: The stochastic nature of the diffusion sampler allows for the generation of multiple plausible reconstructions from the same measurement, providing a visual measure of reconstruction confidence in low-photon regimes.

Results

The framework was validated on diverse biological samples, including F-actin, mitochondria, and nuclei in BPAE and HUVEC cells, under varying photon budgets ( $\alpha = 5$ to $255$) and PSF widths.

Quantitative Performance: In low-photon regimes ( $\alpha = 5, 10$ ), Diffusion-RL significantly outperformed both standard RL and TV-RL in terms of Peak Signal-to-Noise Ratio (PSNR) and Structural Similarity Index (SSIM). For instance, at $\alpha=5$ with a Gaussian PSF ( $\sigma=3$ ), Diffusion-RL achieved a PSNR of $19.94 \pm 1.96$ dB compared to $14.86 \pm 2.14$ dB for RL and $15.72 \pm 2.10$ dB for TV-RL.
Qualitative Improvements: Visual comparisons demonstrated that Diffusion-RL better preserved filamentous structures and punctate details. Unlike TV-RL, which tended to smooth out weak features, Diffusion-RL maintained coherent cellular organization and local texture while avoiding the noise amplification typical of iterative RL.
Robustness to PSF Variations: The method maintained superior performance even under wider PSFs ( $\sigma=5$ ) and simulated realistic confocal PSFs, indicating that the learned prior effectively compensates for increased optical degradation.
Diversity and Uncertainty: Sampling multiple reconstructions from different stochastic initializations revealed that while major cellular structures remained stable (anchored by RL), variations occurred in low-contrast regions. This behavior reflects the inherent uncertainty of the inverse problem rather than uncontrolled hallucination.

Significance and Claims

The paper claims that Diffusion-RL offers a significant advancement for fluorescence microscopy, particularly in measurement-limited regimes where photon budgets are constrained by photobleaching, phototoxicity, or acquisition speed.

Modularity: Unlike supervised deep learning methods (e.g., CARE, 3D-RCAN) that require paired training data for specific microscope configurations, Diffusion-RL uses an unconditional prior that can be combined with any known PSF during inference. This modularity reduces the data burden and enhances generalizability.
Stability: The framework provides a stable alternative to classical deconvolution, preventing the noise amplification that occurs when RL iterations are pushed to recover fine details in noisy data.
Physical Fidelity: By retaining the explicit Poisson likelihood and PSF constraints within the RL loop, the method ensures that reconstructions remain physically consistent with the measured data, avoiding the potential for "hallucinated" structures that might occur in purely generative approaches.

The authors conclude that Diffusion-RL effectively strengthens classical Poisson deconvolution by leveraging learned fluorescence morphology, offering a robust solution for recovering fine biological structures under challenging imaging conditions.