🔬 physics

X-ray dark-field imaging from intensity flow: A Fokker-Planck approach to grating interferometry

This paper introduces a novel Fokker-Planck-based algorithm for retrieving transmission and dark-field images in grating interferometry, which demonstrates comparable performance to conventional methods while significantly reducing artifacts and improving robustness under conditions of low flux, high noise, and grating perturbations.

Original authors: Samantha J. Alloo, Florian Schaff, Regine Gradl, Benedikt Gunther, Franz Pfeiffer, Kaye S. Morgan

Published 2026-05-05

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Samantha J. Alloo, Florian Schaff, Regine Gradl, Benedikt Gunther, Franz Pfeiffer, Kaye S. Morgan

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 take a perfect photograph of a delicate, intricate object using X-rays. In the world of medical imaging, there are three main ways X-rays interact with your body:

Transmission: Some X-rays pass straight through (like light through a window). This gives us the standard "shadow" X-ray.
Phase: Some X-rays get slightly delayed or bent as they pass through soft tissues. This helps us see things that are usually invisible, like soft lung tissue.
Dark-Field: Some X-rays bounce off tiny, invisible structures (like the tiny air sacs in lungs) and scatter. This creates a "foggy" blur that tells us about the microscopic texture of the material.

The Problem: The "One-Pixel" Detective
For decades, scientists have used a technique called Grating Interferometry to capture all three of these images at once. Think of this technique as a sophisticated camera that takes a series of photos while shifting a special filter (a grating) back and forth.

To get the final picture, the computer has to analyze the data from each tiny square on the detector (a pixel) individually. It's like having a team of detectives, where each detective is locked in a tiny room with only one piece of evidence. They have to guess the whole story based only on that single piece of data.

If that single piece of data is noisy, or if the filter has a tiny scratch, that detective gets confused. They might shout out a false alarm, creating "artifacts" (weird spots or noise) in the final image. This is especially bad when you try to take a picture quickly with low light (short exposure), because the data becomes "grainy" and hard to read.

The Solution: The "Neighborhood Watch"
The authors of this paper, Samantha Alloo and her team, developed a brand-new way to process this data. Instead of letting each pixel work alone, they used a mathematical concept called the Fokker–Planck equation.

Here is the analogy:

The Old Way: Imagine trying to understand the weather in a city by asking one person on one street corner what the wind feels like. If that person is wrong or if a gust of wind hits them randomly, your weather map is ruined.
The New Way (Fokker–Planck): Imagine asking the whole neighborhood. You look at how the wind flows from one street to the next. You understand that if the wind is blowing hard on one block, it's likely affecting the next block too. By sharing information with neighbors, you get a much smoother, more accurate picture of the weather, even if one person's report is a bit shaky.

How They Did It
The team didn't just change the math; they changed how they looked at the data.

The Interlacing Trick: They took the series of photos taken while the filter moved and "interlaced" them. Imagine taking a deck of cards (the photos) and weaving them together into one giant, wide image. Suddenly, the tiny, invisible patterns created by the filter became visible, like a barcode.
The Flow Model: They treated the X-rays not as static dots, but as a flowing fluid. Just as water flows around a rock in a river, X-rays flow and diffuse around the tiny structures in a body. The Fokker–Planck equation is a map that predicts exactly how this "X-ray fluid" should flow.
The Result: By solving this flow equation, they could reconstruct the transmission and "dark-field" images. Because the math looks at the whole image at once (the whole neighborhood), it ignores the random noise that confuses the single-pixel detectives.

What They Found
They tested this new method on two things: a simple test object (wood and tissue paper) and a mouse's chest.

Handling Scratches: When the filter had a tiny scratch, the old method produced a chaotic, noisy mess in that spot. The new method simply smoothed it out, realizing, "Oh, that pixel is broken, but its neighbors tell us what's really there."
Low Light/Speed: When they simulated taking a picture very quickly (with very little light, creating a lot of grainy noise), the old method produced a very noisy, fuzzy image. The new method produced a much cleaner, smoother image. It was better at ignoring the "static" and keeping the true picture clear.
Consistency: The new method didn't invent new details; it just removed the errors. The images looked very similar to the old method where the data was good, but much better where the data was bad.

The Bottom Line
This paper introduces a smarter way to process X-ray images. By treating the image data as a connected flow rather than a collection of isolated dots, the new method creates clearer pictures, especially when the conditions aren't perfect (like when the equipment has scratches or when you need to take a picture very fast). It's like upgrading from a team of isolated detectives to a coordinated neighborhood watch that knows how to ignore the noise and find the truth.

Technical Summary: X-ray Dark-Field Imaging from Intensity Flow: A Fokker–Planck Approach to Grating Interferometry

Problem Statement
Grating interferometry (GI) is a mature diagnostic technique capable of simultaneously acquiring transmission, differential phase, and dark-field X-ray images. While GI offers significant advantages for human-scale imaging—such as compatibility with large-pixel detectors and low-coherence sources—the standard image retrieval process has remained fundamentally unchanged for nearly two decades. The conventional "gold standard" relies on pixel-wise sinusoidal fitting of phase-stepping curves to separate physical mechanisms. This local, single-pixel approach is susceptible to artifacts arising from grating perturbations (e.g., structural defects like scratches), reduced flux, or high Poisson noise. In high-noise scenarios, such as fast imaging with short exposure times, the conventional method can amplify high-frequency noise and produce spurious dark-field signals, particularly in regions of strong attenuation where signal counts are low.

Methodology
The authors propose a fundamentally new retrieval algorithm derived from the X-ray Fokker–Planck equation (XFPE). Unlike the local sinusoidal fitting method, the XFPE approach models image contrast formation as a combination of coherent (phase) and diffusive (dark-field) energy flows. The methodology involves transforming standard GI phase-stepping data into a format compatible with single-exposure XFPE algorithms originally developed for speckle-based imaging:

Data Interlacing: The phase-stepping scans (acquired at $n$ analyzer grating positions) are reorganized. For each pixel column in the original image, the intensity values across the $n$ steps are laid out sequentially across $n$ adjacent columns in a new, wider image. This creates a "resolved" 1D periodic grating pattern (reference) and a corresponding sample-modulated pattern.
Resizing: The interlaced images are resized vertically (via row duplication) to ensure square pixels, a requirement for the Fourier-space analysis inherent in the XFPE solution.
XFPE Solution: The algorithm applies two forms of the XFPE—the "evolving" equation (modeling the transition from reference to sample) and the "devolving" equation (modeling the reverse). These equations solve for the transmitted intensity ( $t$ $t$ ), phase shift ( $\phi$ $ϕ$ ), and the effective diffusion coefficient ( $D$ $D$ ), which characterizes the dark-field signal.
- Transmission is initially approximated using a Transport-of-Intensity Equation (TIE) derived approach, neglecting dark-field effects.
- The dark-field images are then retrieved using the solved transmission and the specific XFPE differential operators.
Ensemble Averaging: To maximize signal-to-noise ratio (SNR) and robustness, the process is repeated with "reverse" interlacing (steps ordered right-to-left). The final output is an average of two transmission images and four dark-field images (two from evolving and two from devolving solutions), effectively treating the retrieval as an ensemble method.
Regularization: A specific, data-driven expression for the Tikhonov regularization parameter ( $\epsilon$ ) is introduced, calculated based on the measured signal-to-noise ratio and pixel size, removing the need for manual tuning.

Key Contributions

Novel Algorithm: The development of the first XFPE-based image retrieval method specifically adapted for grating interferometry data.
Data Transformation: The introduction of an interlacing technique that converts phase-stepping curves into a spatially resolved grating pattern, enabling the application of whole-image XFPE solvers.
Noise Resilience: The demonstration that the XFPE framework, by incorporating information from neighboring pixels (global processing), is inherently more robust to noise and perturbations than local, single-pixel fitting.
Artifact Suppression: The method effectively suppresses artifacts caused by grating defects (scratches) and low-flux conditions where conventional methods fail or produce random values.

Results
The authors validated the algorithm using two datasets: a test sample containing wood and tissue paper, and a dynamic mouse chest dataset acquired at varying exposure times (including simulated high-noise conditions).

Test Sample: The XFPE-retrieved transmission images were qualitatively and quantitatively consistent with conventional results. However, in the dark-field channel, the XFPE method successfully suppressed the random, high-contrast artifacts generated by a scratch on the phase grating, whereas the conventional method produced a rapidly varying, noisy signal in the same region.
Mouse Chest: Across all exposure times, the XFPE dark-field images were visually smoother and exhibited less high-frequency noise than the conventional method. In regions of high attenuation (e.g., the spine/stomach area), where photon counts were low and noise high, the conventional method produced spurious dark-field contrast due to biases in the sinusoidal fit. The XFPE method suppressed these artifacts, yielding a more physically consistent image.
Quantitative Analysis: Fourier power spectra analysis confirmed that the conventional method amplifies high-spatial-frequency noise more strongly than the XFPE method, particularly in short-exposure (high-noise) datasets. The XFPE approach demonstrated a 17.3% lower standard deviation in retrieved dark-field values compared to 10.7% for the conventional method under simulated high-noise conditions. Spatial resolution measurements also indicated an improvement (up to a factor of 1.77) when using XFPE in high-noise scenarios.

Significance
The paper claims that the presented XFPE-based method offers a superior alternative to conventional sinusoidal fitting for grating interferometry, particularly in scenarios involving fast imaging, short exposure times, or imperfect experimental conditions (e.g., grating defects). By shifting from a local, single-pixel analysis to a global, whole-image approach that leverages the physics of intensity flow, the method yields smoother, more reproducible dark-field images with reduced artifacts. The authors emphasize that the algorithm is designed for easy integration into existing pipelines and has been released as open-source code to facilitate adoption in the grating-interferometry community. The work suggests that this approach could be extended to other stepping-based imaging techniques, such as analyzer-based imaging and edge-illumination, to provide a cohesive framework for phase and dark-field retrieval.

More like this