Quality enhancement of time-resolved computed tomography scans with cycleGAN

Abstract

Time-resolved x-ray tomography enables us to dynamically and non-destructively study the interior of a specimen. The obtainable temporal resolution is limited by the x-ray flux and the desired spatial resolution. To allow faster acquisition speeds, we explore a deep-learning approach that applies super-resolution and image denoising to fast time-resolved tomograms. The domain translation algorithm, cycleGAN, can apply the high image quality of slow-acquisition tomograms to low-quality fast-acquisition tomograms. It can be trained with unpaired datasets, enabling different samples and detector setups for recording the fast-acquisition and slow-acquisition datasets. In this thesis, we use time-resolved tomograms of carbon microfibers to evaluate the cycleGAN algorithm.
The aim is to retrieve ultra-fast, high-quality tomograms that permit detailed studies of carbon fibers.

Popular Abstract

Popular Abstract in Q/A style:

What did you do for your thesis?
In my thesis, I modified a machine learning algorithm, called cycleGAN, to enhance the image quality of 3D videos, which are called time-resolved tomograms.

So, what are time-resolved tomograms?
In tomograms, we combine x-ray images from different angles into one 3D image that shows the interior of our sample. For time-resolved tomograms, we collect x-ray images while performing experiments on a continuously rotating sample. Like this, we can record a 3D x-ray video.

And what types of samples did you look at?
In our experiments, we studied how carbon fibers break. Carbon fibers are a lightweight and super-strong material that we use in airplanes, cars, or bikes. Basically whenever low weight and high strength are important. Engineers who work with carbon fibers need to know under which circumstances and at which force the carbon fibers break so that they can design a strong product with as little material as possible. We investigate these parameters that lead to fiber breaks with our measurements.

Alright, and how does image enhancement with machine learning work?
In most of these machine learning algorithms, we need the same image in 1) a low-quality version and 2) a high-quality version. Then we tell the computer "find a conversion from the low-quality to the high-quality image". We call it a paired dataset when we have a paired low and high quality version of every image. We cannot use paired datasets, because we intentionally break our samples and we cannot repeat the exact same experiment with the exact same sample. Instead, we use a machine learning algorithm that works with unpaired datasets. CycleGAN first 'looks' at a set of high-quality images and then 'looks' at a set of low-quality images that contain similar but different samples. After that, it can transfer high-quality features, such as contrast and shape, to a low-quality image, while retaining important information, such as fiber positions. This is illustrated in the figure on the right. The image information from the low-quality image on the top left is combined with the high image quality from the image on the top right, resulting in an enhanced image on the bottom.

You modified this algorithm?
Yes! I made two major changes to the code.
1) I changed it so that it can increase the pixel density and apply super-resolution.
2) I made it possible to use the algorithm with 3D data, rather than stacked 2D images. The context from the 3rd dimension significantly improves the image enhancements.

You have high-quality tomograms already. Why do you want to improve low-quality tomograms?
That has to do with the acquisition speed. We can either take slow-acquisition, high-quality images, or fast-acquisition, low-quality images. In our x-ray videos we want to increase the temporal resolution. That means we want to precisely tell which fiber breaks when. We can only record low-quality tomograms at high acquisition speeds, but we want to combine high-quality images with high temporal resolution.

Do you have results?
Yes! On the right, we can see a slice from a fast acquisition tomogram and the corresponding enhancement. The fast-acquisition image has 128 x 128 pixels, and these numbers were doubled in the enhancement. The signal-to-noise ratio is also bad in the fast acquisition image, where we have difficulty see in the fiber outlines. In the enhancement, we can clearly see the round fiber borders.

• author: Stubbe, Johannes
• supervisor: Yuhe Zhang; Pablo Villanueva Perez
• organization: Synchrotron Radiation Research; Department of Physics
• course: FYSM60 20221
• year: 2023
• type: H2 - Master's Degree (Two Years)
• subject: Physics and Astronomy
• keywords: carbon fibers, carbon fibres, microfibers, tomography, deep learning, cycleGAN, time-resolved tomography
• language: English
• id: 9130196
• date added to LUP: 2023-07-04 16:03:36
• date last changed: 2023-07-04 16:03:36