Publication detail

Self-supervised learning for high quality tiled X-ray computed tomography imaging: a simulation study

MATULA, J. PELT, D. VAN LEEUWEN, T. ZIKMUND, T. KAISER, J.

Original Title

Self-supervised learning for high quality tiled X-ray computed tomography imaging: a simulation study

Type

abstract

Language

English

Original Abstract

Tiling, also referred to as stitching, is a simple computed tomography (CT) measurement technique to scan objects larger than the field of view of the CT system. It is utilised in both lab-based CT systems and synchrotron beamlines. Briefly, independent CT measurements are performed in different parts of the sample and then stitched together to obtain the tomographic reconstruction of the whole object. This scanning strategy is, however, extremely time demanding due to the need to perform multiple scans per sample. A compromise is often made between the image quality and acquisition time. This compromise may lead to noisy tomographic reconstructions, where the low data quality complicates further image analysis. Tiled CT measurements are typically performed with an overlap that allows precise alignment and stitching of the individual scan tiles into a single volume. Our work shows that Noise2Noise-like denoising convolutional neural network (CNN) can be trained in this overlapping region and applied to the entire reconstructed CT volume. This approach offers a way to utilise otherwise redundant image data to improve the overall quality of the CT scan. We perform simulation experiments to evaluate the noise reduction performance of this approach in different settings (different levels of overlap, CT scanning geometry, noise level). These experiments demonstrate that by training a CNN-based denoising model only on the noisy data in the limited region of the overlap between the individual scan tiles, we can achieve denoising performance nearing a CNN trained in a fully-supervised manner with clean targets. We also demonstrate that the proposed methodology not only improves the signal-to-noise ratio but can also reduce the effect of some of the most common tomographic artefacts.

Keywords

noise reduction; deep learning; X-ray computed tomography

Authors

MATULA, J.; PELT, D.; VAN LEEUWEN, T.; ZIKMUND, T.; KAISER, J.

Released

19. 11. 2022

Location

Rome

Pages count

1

URL

http://icmv.org/

BibTex

@misc{BUT180117,
  author="Jan {Matula} and Daniel M. {Pelt} and Tristan {van Leeuwen} and Tomáš {Zikmund} and Jozef {Kaiser}",
  title="Self-supervised learning for high quality tiled X-ray computed tomography imaging: a simulation study",
  booktitle="Fifteenth International Conference on Machine Vision (ICMV 2022)",
  year="2022",
  pages="1",
  address="Rome",
  url="http://icmv.org/",
  note="abstract"
}