Far-field phase retrieval with unpaired AI approaches

Abstract

Coherent Diffraction Imaging (CDI) is an imaging technique that allows getting a high contrast image from phase interactions within a sample. However, the field of X-ray CDI has been dealing with a common issue known as the phase problem. Upon interaction with the sample, the incoming photons get absorbed or scatter and therefore, the wavefield gets altered. It continues to propagate towards the detector. The latter is unable to record the phase of the propagated waves directly, but instead can only pick up on their intensities. Therefore, a technique is needed to reconstruct the missing phase information from the recorded intensity reading on the detector.
Classical phase retrieval involves iterative computer algorithms that try to approximate the missing phase data stepwise, which is a long process and does not always guarantee a solution. When trying to take a vast amount of images in a short period of time, this long winding process can be limiting. It is, therefore, of great advantage to use an algorithm that can reconstruct the missing phase in a short period of time.
This is where machine learning algorithms come in to determine the missing phase data. So far, this has been successfully implemented for experiments in the near field [1]. The latter is based on CycleGAN [2], an image generation algorithm using Generative Adversarial Networks (GANs) that are known for their image generation capabilities. In this thesis, the focus is on using this technique to reconstruct phase images for a far-field technique such as CDI. The GAN model was trained with simulated detector intensity images and objects based on the MNIST database. The GAN model was trained using unpaired sets of images from each domain (objects and detector images). By implementing the far-field Fraunhofer propagator into the model and adjusting its parameters, phase objects from never-before-seen detector images were successfully reconstructed.
Reconstructions were performed on 1000 images not used during training and the results were statistically analysed: The mean squared error (MSE) as well as the dissimilarity index (DSSIM) was computed for each reconstruction and a corresponding reference image. Both metrics present a Gaussian distribution, with a relatively low mean value and spread. This indicates that the model produces data predictably with good accuracy.
As only simulated and simplified data was used, future research would include adjusting the model for the usage of real experimental data. If this succeeds, the model can be used to obtain real-time reconstructions and will greatly simplify image acquisition in many fields where CDI is performed.


1. Zhang, Y. et al. PhaseGAN: A deep-learning phase-retrieval approach for unpaired
datasets. Optics Express 29 (June 2021).
2. Zhu, J. - Y., Park, T., Isola, P. & Efros, A. A. Unpaired Image-to-Image Translation using Cycle-Consistent Adversarial Networks 2020.

Popular Abstract

The usage of X-rays is very common in the area of medical examinations, where most of us have encountered them before. These rays, however, hold the potential to analyse many more structures than just the human body and are able to probe samples down to the nanometre scale. This is why they are a very powerful tool to analyse materials in a variety of different areas within science and industry.
Just like taking a picture with a camera by using visible light, we can take a picture of objects using X-rays, which are light waves not visible to the human eye. A common way to examine materials by shining X-rays at them is though a process called scattering: the atoms within the sample cause the incoming waves to change direction, each one differently. This is the same phenomenon that occurs when trying to see through thick fog. One such technique making use of light rays that have been scattered by a sample is Coherent Diffraction Imaging (CDI). We start by shining parallel rays at the sample. These rays consist of light waves that are in sync (or, more scientifically, “in phase”). After scattering from the sample, the rays are no longer parallel, nor in phase. The light waves interfere with each other because the scattering caused them to change directions. The differently scattered waves now interact with each other by overlapping, similar to the way ripples overlap when throwing stones into a pond. This overlap is determined by how much the waves get thrown out of sync - or, more scientifically, their phase difference. The combination of these interactions leads to a so-called diffraction pattern, which is nothing more than a new set of waves with different peaks and valleys at various points in space. However, there is a problem: The detectors can only measure how strong these waves are at each point in space but cannot make out where the peaks and valleys are located, meaning it cannot make out their phase differences. This is known as the phase problem. Current methods of restoring the lost phase information involve computer algorithms that take a long time to run and are not always successful. Once the phase is restored, it is possible to see an image of what the sample actually looks like.
This thesis explored the possibilities of machine learning (ML) as an alternative to these classical methods, with the hopes of providing a faster algorithm that produces higher quality reconstructions. The code used for this has been adapted from an already existing algorithm that is capable of transforming images into a different domain: it can for example transform a photograph into an image that looks like a Van Gogh painting. By adding the physics of image formation into the code, this algorithm was now able to transform detector images into pictures of the corresponding samples, signifying a success for this thesis.
For this project, only simulated and simplified data was used. Further research will therefore involve the usage of real experimental data. If implemented successfully, this allows us to produce higher quality images in a shorter time frame and reduce the dosage of radiation that the sample is exposed to. This would be valuable in the field of medicine, for example, where the samples are sensitive to the dosage of X-rays applied to them. The fast image acquisition would also allow us to capture images in fast succession, allowing us to create video footage of processes that happen within microseconds.

• author: Brumat, Claire
• supervisor: Pablo Villanueva Perez; Zisheng Yao
• organization: Synchrotron Radiation Research; Department of Physics
• course: FYSK04 20242
• year: 2025
• type: M2 - Bachelor Degree
• subject: Physics and Astronomy
• keywords: Coherent diffraction imaging, CDI, X-ray imaging, Synchrotron radiation, Phase problem, Phase retrieval, Phase reconstruction, Far-field imaging, Fraunhofer propagator, Machine learning, Deep learning, Generative Adversarial Network, GAN, CycleGAN
• language: English
• id: 9209840
• date added to LUP: 2025-08-19 09:35:11
• date last changed: 2025-08-19 09:35:11