Enhancing industrial X-ray tomography by data-centric statistical methods

Jarkko Suuronen, Muhammad F. Emzir, Sari Lasanen, Simo Särkkä, Lassi Roininen

In the high-stakes world of industrial inspection—whether you're scanning a log for metallic shrapnel before it hits a saw blade or analyzing the microscopic pores of a drill core—the quality of your X-ray reconstruction is everything. A new paper titled "Enhancing industrial X-ray tomography by data-centric statistical methods" explores how moving beyond traditional "smooth" mathematical models to more sophisticated, "edge-preserving" Bayesian priors can revolutionize how we see inside objects.

TL;DR

The researchers demonstrate that using non-Gaussian statistical priors (specifically Cauchy and Besov distributions) significantly improves X-ray tomography reconstructions for industrial targets like logs and rock samples. These methods preserve sharp edges and high-contrast features even with sparse, noisy data, while advanced MCMC algorithms allow for robust uncertainty quantification in high-dimensional spaces.

What Problem Does This Paper Address?

Traditional X-ray reconstruction often relies on Filtered Back-Projection (FBP) or Gaussian-based Bayesian methods. While these are computationally efficient, they have a major flaw: they assume the world is "smooth." In industrial settings, this isn't true. A knot in a tree or a pore in a rock has a sharp boundary.

When X-ray data is limited (sparse-angle tomography) or noisy, Gaussian models tend to blur these critical edges or create distracting artifacts. The authors tackle the "inverse problem" of tomographic reconstruction—essentially working backward from 2D X-ray shadows to a 3D object—by testing mathematical frameworks that actually "expect" sharp transitions and high contrast.

Methods

The authors framed the reconstruction as a Bayesian statistical inverse problem. They compared four different "priors" (mathematical assumptions about the object's internal structure):

  1. Gaussian: Favors smoothness (the industry standard).
  2. Total Variation (TV): Promotes sharp edges but isn't "mesh-invariant" (the results change if you change the pixel size).
  3. Besov: A wavelet-based prior that is mesh-invariant but can be "blocky."
  4. Cauchy: A heavy-tailed prior that is excellent for sharp transitions and high-contrast features.

To solve these problems, which involve between 100,000 and 1,000,000 unknowns, they used:

  • L-BFGS Optimization: For fast "Maximum A Posteriori" (MAP) point estimates.
  • Hamiltonian Monte Carlo (HMC-NUTS): A sophisticated algorithm used for Uncertainty Quantification (UQ). Unlike a single "best guess" image, UQ provides "error bars" for every pixel, telling engineers how much they can trust the reconstruction.

Datasets

The study utilized two primary synthetic benchmarks designed to simulate real-world industrial challenges:

  • Sawmill Log Tomography: A high-contrast 3D model of a log containing knots, rot, and a metallic inclusion (a "worst-case" scenario for traditional smoothing).
  • Drill-core Rock Sample: A lower-contrast target focused on identifying complex pore structures and mineralized soil within a rock core.
  • Sparsity Testing: Both datasets were tested using 90, 30, and 10 measurement angles to see how the methods held up under extreme data scarcity.

Key Insights and Ideas

  • Cauchy is the "Contrast King": The Cauchy prior was particularly effective at capturing the metallic piece in the log. Even when the number of X-ray angles was slashed to 10, the Cauchy prior "found" the metal, whereas Gaussian models failed entirely.
  • The Mesh Invariance Problem: The authors highlight that Total Variation (TV) priors—while popular—fail at systematic uncertainty quantification because they don't behave consistently when you refine the computational grid. Besov and Cauchy priors solve this, making them better candidates for "data-centric" sensor optimization.
  • HMC for High Dimensions: The paper proves that Hamiltonian Monte Carlo (specifically the NUTS variant) is viable for full uncertainty quantification in industrial imaging, even with nearly 300,000 variables, provided the distribution isn't too "pathologically" complex.

Key Results

  • Quantitative Gains: In the log tomography test (90 angles), the TV and Cauchy priors achieved Peak Signal-to-Noise Ratios (PSNR) of 36.6 dB and 36.0 dB respectively, significantly outperforming the Gaussian prior's 33.7 dB.
  • Qualitative Superiority: Visually, the Cauchy reconstructions suppressed the "streaking" artifacts common in sparse-angle CT.
  • Sparsity Resilience: At just 10 angles, the non-Gaussian priors maintained a PSNR above 31 dB, while the Gaussian model dropped to 26.6 dB, resulting in an image too blurry for industrial use.

Limitations and Future Work

Despite the impressive results, two main hurdles remain:

  1. Computational Speed: While MAP estimates are fast, full MCMC uncertainty quantification is still too slow for a high-speed sawmill production line. It is currently better suited for "non-time-critical" applications like lab-based rock analysis.
  2. Sampling Difficulties: Cauchy priors create "heavy-tailed" and sometimes multimodal probability distributions. This makes it difficult for MCMC samplers to explore the space efficiently, leading to potential convergence issues.

The authors suggest that the next frontier lies in Deep Bayesian Inversion—combining these robust statistical priors with machine learning to get the "best of both worlds": the speed of AI and the reliability of Bayesian physics.

Summary updated

Abstract

X-ray tomography has applications in various industrial fields such as sawmill industry, oil and gas industry, as well as chemical, biomedical, and geotechnical engineering. In this article, we study Bayesian methods for the X-ray tomography reconstruction. In Bayesian methods, the inverse problem of tomographic reconstruction is solved with the help of a statistical prior distribution which encodes the possible internal structures by assigning probabilities for smoothness and edge distribution of the object. We compare Gaussian random field priors, that favor smoothness, to non-Gaussian total variation (TV), Besov, and Cauchy priors which promote sharp edges and high- and low-contrast areas in the object. We also present computational schemes for solving the resulting high-dimensional Bayesian inverse problem with 100,000–1,000,000 unknowns. We study the applicability of a no-U-turn variant of Hamiltonian Monte Carlo (HMC) methods and of a more classical adaptive Metropolis-within-Gibbs (MwG) algorithm to enable full uncertainty quantification of the reconstructions. We use maximum a posteriori (MAP) estimates with limited-memory BFGS (Broyden–Fletcher–Goldfarb–Shanno) optimization algorithm. As the first industrial application, we consider sawmill industry X-ray log tomography. The logs have knots, rotten parts, and even possibly metallic pieces, making them good examples for non-Gaussian priors. Secondly, we study drill-core rock sample tomography, an example from oil and gas industry. In that case, we compare the priors without uncertainty quantification. We show that Cauchy priors produce smaller number of artefacts than other choices, especially with sparse high-noise measurements, and choosing HMC enables systematic uncertainty quantification, provided that the posterior is not pathologically multimodal or heavy-tailed.
BibTeX
@inproceedings{suuronen2020enhancing,
  author = {Jarkko Suuronen and Muhammad F. Emzir and Sari Lasanen and Simo Särkkä and Lassi Roininen},
  title = {{Enhancing industrial X-ray tomography by data-centric statistical methods}},
  booktitle = {University},
  year = {2020},
}
Figure