Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

SR4ZCT: Self-supervised Through-plane Resolution Enhancement for CT Images with Arbitrary Resolution and Overlap

Jiayang Shi, Daniel M. Pelt, K. Joost Batenburg

TL;DR

The paper tackles the challenge of anisotropic through-plane resolution and slice overlap in CT imaging by introducing SR4ZCT, a self-supervised method that supports arbitrary resolution and overlap via explicit modeling of voxel spacings across orientations and off-axis training. It synthesizes training pairs through vertical and horizontal down/up-sampling using linear interpolation, trains a 2D MS-D network to map degraded inputs back to HR axial content, and then applies the learned mapping to coronal and sagittal views after upscaling. The approach avoids dependence on HR references and demonstrates competitive or superior performance to supervised baselines and convolution-based methods, with strong evidence for the necessity of accurate training-data modeling. The work suggests practical clinical utility in improving through-plane CT resolution across diverse acquisition settings, with potential for broader application where resolution mismatch across orientations limits 3D CT quality.

Abstract

Computed tomography (CT) is a widely used non-invasive medical imaging technique for disease diagnosis. The diagnostic accuracy is often affected by image resolution, which can be insufficient in practice. For medical CT images, the through-plane resolution is often worse than the in-plane resolution and there can be overlap between slices, causing difficulties in diagnoses. Self-supervised methods for through-plane resolution enhancement, which train on in-plane images and infer on through-plane images, have shown promise for both CT and MRI imaging. However, existing self-supervised methods either neglect overlap or can only handle specific cases with fixed combinations of resolution and overlap. To address these limitations, we propose a self-supervised method called SR4ZCT. It employs the same off-axis training approach while being capable of handling arbitrary combinations of resolution and overlap. Our method explicitly models the relationship between resolutions and voxel spacings of different planes to accurately simulate training images that match the original through-plane images. We highlight the significance of accurate modeling in self-supervised off-axis training and demonstrate the effectiveness of SR4ZCT using a real-world dataset.

SR4ZCT: Self-supervised Through-plane Resolution Enhancement for CT Images with Arbitrary Resolution and Overlap

TL;DR

The paper tackles the challenge of anisotropic through-plane resolution and slice overlap in CT imaging by introducing SR4ZCT, a self-supervised method that supports arbitrary resolution and overlap via explicit modeling of voxel spacings across orientations and off-axis training. It synthesizes training pairs through vertical and horizontal down/up-sampling using linear interpolation, trains a 2D MS-D network to map degraded inputs back to HR axial content, and then applies the learned mapping to coronal and sagittal views after upscaling. The approach avoids dependence on HR references and demonstrates competitive or superior performance to supervised baselines and convolution-based methods, with strong evidence for the necessity of accurate training-data modeling. The work suggests practical clinical utility in improving through-plane CT resolution across diverse acquisition settings, with potential for broader application where resolution mismatch across orientations limits 3D CT quality.

Abstract

Computed tomography (CT) is a widely used non-invasive medical imaging technique for disease diagnosis. The diagnostic accuracy is often affected by image resolution, which can be insufficient in practice. For medical CT images, the through-plane resolution is often worse than the in-plane resolution and there can be overlap between slices, causing difficulties in diagnoses. Self-supervised methods for through-plane resolution enhancement, which train on in-plane images and infer on through-plane images, have shown promise for both CT and MRI imaging. However, existing self-supervised methods either neglect overlap or can only handle specific cases with fixed combinations of resolution and overlap. To address these limitations, we propose a self-supervised method called SR4ZCT. It employs the same off-axis training approach while being capable of handling arbitrary combinations of resolution and overlap. Our method explicitly models the relationship between resolutions and voxel spacings of different planes to accurately simulate training images that match the original through-plane images. We highlight the significance of accurate modeling in self-supervised off-axis training and demonstrate the effectiveness of SR4ZCT using a real-world dataset.
Paper Structure (4 sections, 1 equation, 5 figures, 2 tables)

This paper contains 4 sections, 1 equation, 5 figures, 2 tables.

Table of Contents

  1. Introduction
  2. Method
  3. Experiments and Results
  4. Conclusion

Figures (5)

  • Figure 1: Overview of our SR4ZCT method.
  • Figure 3: The comparison of supervised learning and SR4ZCT on $6.25mm$ resolution and $3.125mm$ overlap dataset. Patches are shown for visualization. The PSNR and SSIM of each patch are shown in the top- and bottom-left corners.
  • Figure 4: Results of SR4ZCT applied on L291 from Low Dose CT Grand Challenge mccollough2017low. We show two image patches of coronal and sagittal images. The first row contains patches of original coronal and sagittal images, and the second row shows the output of SR4ZCT. Extra results are shown in the supplementary file.
  • Figure 5: The PSNR of cases where training images were inaccurately modeled. Each block refers to a case where the corresponding resolution/overlap are used to model the training images. The actual resolution/overlap are $5mm$/$2.5mm$.
  • Figure 7: Visual comparison of state-of-the-art CT images supervised super-resolution methods SAINT peng2020saint, RPLHR yu2022rplhr, self-supervised method SMORE zhao2020smore and our method SR4ZCT applied to a real-world sagittal image.