Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

SCOUT: Fast Spectral CT Imaging in Ultra LOw-data Regimes via PseUdo-label GeneraTion

Guoquan Wei, Liu Shi, Shaoyu Wang, Mohan Li, Cunfeng Wei, Qiegen Liu

TL;DR

This work proposes a reconstruction method under ultra-low raw data conditions, requiring no external data and avoiding lengthy pre-training processes, and provides a new paradigm for research using unlabeled raw projection data.

Abstract

Noise and artifacts during computed tomography (CT) scans are a fundamental challenge affecting disease diagnosis. However, current methods either involve excessively long reconstruction times or rely on data-driven models for optimization, failing to adequately consider the valuable information inherent in the data itself, especially medical 3D data. This work proposes a reconstruction method under ultra-low raw data conditions, requiring no external data and avoiding lengthy pre-training processes. By leveraging spatial nonlocal similarity and the conjugate properties of the projection domain to generate pseudo-3D data for self-supervised training, high-fidelity results can be achieved in a very short time. Extensive experiments demonstrate that this method not only mitigates detector-induced ring artifacts but also exhibits unprecedented capabilities in detail recovery. This method provides a new paradigm for research using unlabeled raw projection data. Code is available at https://github.com/yqx7150/SCOUT.

SCOUT: Fast Spectral CT Imaging in Ultra LOw-data Regimes via PseUdo-label GeneraTion

TL;DR

This work proposes a reconstruction method under ultra-low raw data conditions, requiring no external data and avoiding lengthy pre-training processes, and provides a new paradigm for research using unlabeled raw projection data.

Abstract

Noise and artifacts during computed tomography (CT) scans are a fundamental challenge affecting disease diagnosis. However, current methods either involve excessively long reconstruction times or rely on data-driven models for optimization, failing to adequately consider the valuable information inherent in the data itself, especially medical 3D data. This work proposes a reconstruction method under ultra-low raw data conditions, requiring no external data and avoiding lengthy pre-training processes. By leveraging spatial nonlocal similarity and the conjugate properties of the projection domain to generate pseudo-3D data for self-supervised training, high-fidelity results can be achieved in a very short time. Extensive experiments demonstrate that this method not only mitigates detector-induced ring artifacts but also exhibits unprecedented capabilities in detail recovery. This method provides a new paradigm for research using unlabeled raw projection data. Code is available at https://github.com/yqx7150/SCOUT.
Paper Structure (14 sections, 10 equations, 4 figures)

This paper contains 14 sections, 10 equations, 4 figures.

Table of Contents

  1. Introduction
  2. Results
  3. Overview of SCOUT
  4. Applying SCOUT to Address Intrinsic Noise in Data Acquisition Equipment
  5. The Remarkable Performance in Walnut Dual-Energy CT Data
  6. Universality of Data from Traditional Energy Integral Detectors
  7. Discussion
  8. Methods
  9. Experimental Equipment
  10. Network Process and Objective Function
  11. Explanation of Similarity of Projection Data
  12. Data Operations and Simulation Instructions
  13. Comparison of Advanced Self-Supervised Methods
  14. Evaluation Metrics

Figures (4)

  • Figure 1: Schematic diagram of the overall workflow and principles of SCOUT.(a) (top half) Introduces the detailed process of creating a voxel library starting from low-dose CT projection data, using spatial nonlocal similarity and the conjugate theorem. (a) (bottom half) After obtaining the voxel library, two original volume projection data are randomly selected for self-supervised training, followed by testing to obtain denoised volume data. (b) Introduces the similar anatomical structures that exist in large numbers in human data, laying the groundwork for explaining the existence of the same similarity in the projection domain. (c) Intuitively demonstrates the conjugate symmetry caused by the scanning process, which also supports the creation of pseudo-label data. Then, it presents the manifestation of conjugate properties in single slices, providing theoretical support for extending to domain slices.
  • Figure 2: Mouse and acquisition equipment model and results description.(a) The coronal and sagittal plots of the self-supervised method with good results. (b) Axial slice results showing that the second and third rows are magnified views of the selected regions of interest. (c) The record shows the total time spent on training and testing. This demonstrates that our algorithm remains extremely fast even when processing volumetric data. (d) (top half) The scanning geometry of the mouse acquisition system $\mu$Color SA. Right half, the mice used in the experiment. (d) (bottom half) The mice used in the experiment.
  • Figure 3: Relevant results from the walnut data.(a) Box plots of PSNR, SSIM, and RMSE for all comparative experiments and reference reconstructed images. (b) Coronal and sagittal plots. (c) Axial plane dual-energy spectral data results and magnified views of the region of interest.
  • Figure 4: Results related to traditional medical human data.(a) Quantitative evaluation results on Mayo 2016 data, with the first and second rows showing axial slices and magnified views of the region of interest. The third and fourth rows show the coronal and sagittal results of the volume data. (b) Results demonstrating effectiveness at ultra-low doses. (c) Experimental results addressing ring artifacts involving bright streaks, dark streaks, and broad streaks.