Simulating X-ray Computed Tomography: Developing a Virtual Representation and Generating Synthetic Training Data for Deep Learning
(2026) FHLM01 20261Solid Mechanics
Department of Construction Sciences
- Abstract
- This thesis investigates simulating X-ray computed tomography with two primary objectives: developing a virtual representation of an X-ray computed tomography scanner and generating labelled synthetic training data for deep learning-based fibre segmentation. A simulation pipeline to generate X-ray projections based on the Beer-Lambert law was developed using the open-source framework gVirtualXray.
The first part explores developing a virtual representation of a laboratory-based X-ray computed tomography scanner at Tetra Pak. This was achieved by optimising simulation parameters such as filtration and detector response in gVirtualXray to match experimentally acquired projections. This method proved successful in improving the accuracy of... (More) - This thesis investigates simulating X-ray computed tomography with two primary objectives: developing a virtual representation of an X-ray computed tomography scanner and generating labelled synthetic training data for deep learning-based fibre segmentation. A simulation pipeline to generate X-ray projections based on the Beer-Lambert law was developed using the open-source framework gVirtualXray.
The first part explores developing a virtual representation of a laboratory-based X-ray computed tomography scanner at Tetra Pak. This was achieved by optimising simulation parameters such as filtration and detector response in gVirtualXray to match experimentally acquired projections. This method proved successful in improving the accuracy of the virtual representation, which produced simulated projections with a root mean square error of 0.0123 when compared to the experimental projections. Despite promising results, this method is limited in its generalisability and a more advanced optimisation would be required to take multiple X-ray computed tomography settings into account.
The second objective explores generating labelled synthetic X-ray computed tomography datasets for training convolutional neural networks to segment fibres. For this purpose, a pipeline was designed to generate multiple diverse datasets of synthetic fibre scans by varying simulation parameters. A U-NET model architecture was trained exclusively on this synthetic data and then evaluated using an experimental dataset. The models showed good performance where the best model managed to generate an F1-score of 0.956, success fully demonstrating the use of synthetic data for training convolutional neural networks in segmentation tasks. However, the work is limited to binary 2D image segmentation, and future advancements should focus on multi-class and 3D volumetric fibre segmentation. (Less) - Popular Abstract
- What if time-consuming X-ray scans could be replaced with realistic simulations? This thesis shows how virtual imaging can replicate such scans and generate synthetic data to train machine learning models to segment complex materials accurately.
X-ray computed tomography (XCT) is an imaging technique that allows one to see inside objects without cutting them open. It is widely used in medicine, but also plays a large role in industry. It is, for example, used as a tool by Tetra Pak to analyse the internal structure of paper packaging materials. This is however a very time consuming process and analysing the large volumes that are produced can be very challenging. This thesis explores how to overcome these limitations in two separate... (More) - What if time-consuming X-ray scans could be replaced with realistic simulations? This thesis shows how virtual imaging can replicate such scans and generate synthetic data to train machine learning models to segment complex materials accurately.
X-ray computed tomography (XCT) is an imaging technique that allows one to see inside objects without cutting them open. It is widely used in medicine, but also plays a large role in industry. It is, for example, used as a tool by Tetra Pak to analyse the internal structure of paper packaging materials. This is however a very time consuming process and analysing the large volumes that are produced can be very challenging. This thesis explores how to overcome these limitations in two separate objectives.
First, a virtual XCT scanner was developed with the goal of being capable of mimicking the behaviour of a real scanner. By adjusting simulation parameters based on real settings and sample information, the virtual scanner was able to produce data which closely matched real experimental data. This showed that simulation can be used to test scanning settings, saving time by reducing the need for repeated physical experiments. The current model is however limited to specific scanning conditions and further work is needed to make it more generally applicable.
The second part focused on applying the virtual scanner to generate artificial XCT images of fibres. The main advantage of this is that the artificial data includes the exact location of the fibres, which is very difficult and time consuming to find manually in real data. Using this data, machine learning models were trained to automatically identify fibres in real XCT images with great success. This showed that even though the models had been trained on artificial data, they could perform well when applied to real data. (Less)
Please use this url to cite or link to this publication:
https://lup.lub.lu.se/student-papers/record/9243853
- author
- Fahlgren, Kajsa LU and Almqvist, Elias
- supervisor
-
- Eoin Walsh LU
- Eskil Andreasson LU
- Erik Bergvall LU
- organization
- course
- FHLM01 20261
- year
- 2026
- type
- H3 - Professional qualifications (4 Years - )
- subject
- report number
- TFHF-5274
- language
- English
- id
- 9243853
- date added to LUP
- 2026-06-25 14:54:23
- date last changed
- 2026-06-25 14:54:23
@misc{9243853,
abstract = {{This thesis investigates simulating X-ray computed tomography with two primary objectives: developing a virtual representation of an X-ray computed tomography scanner and generating labelled synthetic training data for deep learning-based fibre segmentation. A simulation pipeline to generate X-ray projections based on the Beer-Lambert law was developed using the open-source framework gVirtualXray.
The first part explores developing a virtual representation of a laboratory-based X-ray computed tomography scanner at Tetra Pak. This was achieved by optimising simulation parameters such as filtration and detector response in gVirtualXray to match experimentally acquired projections. This method proved successful in improving the accuracy of the virtual representation, which produced simulated projections with a root mean square error of 0.0123 when compared to the experimental projections. Despite promising results, this method is limited in its generalisability and a more advanced optimisation would be required to take multiple X-ray computed tomography settings into account.
The second objective explores generating labelled synthetic X-ray computed tomography datasets for training convolutional neural networks to segment fibres. For this purpose, a pipeline was designed to generate multiple diverse datasets of synthetic fibre scans by varying simulation parameters. A U-NET model architecture was trained exclusively on this synthetic data and then evaluated using an experimental dataset. The models showed good performance where the best model managed to generate an F1-score of 0.956, success fully demonstrating the use of synthetic data for training convolutional neural networks in segmentation tasks. However, the work is limited to binary 2D image segmentation, and future advancements should focus on multi-class and 3D volumetric fibre segmentation.}},
author = {{Fahlgren, Kajsa and Almqvist, Elias}},
language = {{eng}},
note = {{Student Paper}},
title = {{Simulating X-ray Computed Tomography: Developing a Virtual Representation and Generating Synthetic Training Data for Deep Learning}},
year = {{2026}},
}