Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Towards reconstructing experimental sparse-view X-ray CT data with diffusion models

Nelas J. Thomsen, Xinyuan Wang, Felix Lucka, Ezgi Demircan-Tureyen

TL;DR

While severe mismatch causes model collapse and hallucinations, diverse priors outperform well-matched but narrow priors, and future development must validate against real-world benchmarks.

Abstract

Diffusion-based image generators are promising priors for ill-posed inverse problems like sparse-view X-ray Computed Tomography (CT). As most studies consider synthetic data, it is not clear whether training data mismatch (``domain shift'') or forward model mismatch complicate their successful application to experimental data. We measured CT data from a physical phantom resembling the synthetic Shepp-Logan phantom and trained diffusion priors on synthetic image data sets with different degrees of domain shift towards it. Then, we employed the priors in a Decomposed Diffusion Sampling scheme on sparse-view CT data sets with increasing difficulty leading to the experimental data. Our results reveal that domain shift plays a nuanced role: while severe mismatch causes model collapse and hallucinations, diverse priors outperform well-matched but narrow priors. Forward model mismatch pulls the image samples away from the prior manifold, which causes artifacts but can be mitigated with annealed likelihood schedules that also increase computational efficiency. Overall, we demonstrate that performance gains do not immediately translate from synthetic to experimental data, and future development must validate against real-world benchmarks.

Towards reconstructing experimental sparse-view X-ray CT data with diffusion models

TL;DR

While severe mismatch causes model collapse and hallucinations, diverse priors outperform well-matched but narrow priors, and future development must validate against real-world benchmarks.

Abstract

Diffusion-based image generators are promising priors for ill-posed inverse problems like sparse-view X-ray Computed Tomography (CT). As most studies consider synthetic data, it is not clear whether training data mismatch (``domain shift'') or forward model mismatch complicate their successful application to experimental data. We measured CT data from a physical phantom resembling the synthetic Shepp-Logan phantom and trained diffusion priors on synthetic image data sets with different degrees of domain shift towards it. Then, we employed the priors in a Decomposed Diffusion Sampling scheme on sparse-view CT data sets with increasing difficulty leading to the experimental data. Our results reveal that domain shift plays a nuanced role: while severe mismatch causes model collapse and hallucinations, diverse priors outperform well-matched but narrow priors. Forward model mismatch pulls the image samples away from the prior manifold, which causes artifacts but can be mitigated with annealed likelihood schedules that also increase computational efficiency. Overall, we demonstrate that performance gains do not immediately translate from synthetic to experimental data, and future development must validate against real-world benchmarks.
Paper Structure (14 sections, 6 equations, 4 figures, 2 tables)

This paper contains 14 sections, 6 equations, 4 figures, 2 tables.

Table of Contents

  1. Introduction
  2. Related Works
  3. Preliminaries on Diffusion Models
  4. Diffusion-based Priors for Tomography
  5. Methodology
  6. Training Data
  7. Generative Priors
  8. Reconstruction Pipeline
  9. Test domains
  10. Experimental Results
  11. Impact of Training Distribution
  12. Impact of Forward Model Mismatch
  13. Impact of Inference Design
  14. Conclusion

Figures (4)

  • Figure 1: PSNR (dB) as a function of the number of projections across four different test domains for the reconstructions obtained using CGLS and DDS with three different diffusion priors. Solid lines show PSNR, while diamond markers indicate sampled SSIM values at selected projection counts (5, 9, 12). Shaded areas denote $\pm1$ standard deviation across 10 random seeds.
  • Figure 2: Example reconstructions using 12 sparse projections across four test domains (rows). From left to right, columns represent: the gold standard, the baseline reconstruction without imposing any prior, reconstructions using standard SL ($\mathbf{f}_\text{std}$), mixed ($\mathbf{f}_\text{mix}$), and experimental SL ($\mathbf{f}_\text{exp}$) diffusion priors.
  • Figure 3: Line profiles from reconstructions shown in Fig. \ref{['fig:recons']} through the three small holes in the laser-cut phantom. The results involving $\mathbf{f}_\text{std}$ and/or $\mathbf{y}_\text{sim(cad)}$ are excluded. Gold standard is represented by red dashed line. Shaded areas show $\pm1$ standard deviation across 10 random seeds.
  • Figure 4: PSNR (dB) vs number of projections for three resolutions (128, 256, 512) using $\mathbf{f}_\text{exp}$ on experimental data $\mathbf{y}_\text{exp}$ and simulated data $\mathbf{y}_\text{sim(recon)}$, revealing the performance gap caused by forward model mismatch. Arrows indicate SSIM values at selected projection counts (9, 25). Shaded bands denote $\pm 1$ standard deviation.