🔢 mathematics

Microlocal analysis of non-linear artifacts in cone beam CT

This paper presents a novel microlocal analysis demonstrating that beam hardening artifacts in cone-beam CT arise from double-tangent X-ray rays, creating weaker conormal singularities on a specific 2-D surface that are rigorously characterized and validated through both theoretical proofs and simulations.

Original authors: James W. Webber, Alexander Katsevich

Published 2026-08-21

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: James W. Webber, Alexander Katsevich

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Medical imaging often feels like looking at the world through a slightly distorted window. In a standard X-ray, a beam of photons passes through the body, and the amount of light that gets through tells doctors what lies beneath. However, real-world X-ray machines do not produce a single, pure color of light. Instead, they emit a broad spectrum of energies, ranging from weak to very strong. As this mixed beam travels through dense materials like bone or metal, the weaker, softer rays are absorbed more easily than the powerful ones. This causes the beam to become "harder"—composed of higher-energy rays—as it moves through the object. This phenomenon, known as beam hardening, creates streaks and shadows in the final image that do not correspond to any actual physical structure, potentially confusing a diagnosis.

For decades, scientists have understood that these streaks appear when an X-ray beam grazes the edge of a dense object at two points simultaneously, effectively skimming the surface. While previous research had identified these double-tangent paths as the source of the problem, a new study by James Webber and Alexander Katsevich offers a much deeper, more precise explanation of exactly how these errors form and how strong they are compared to the real image. By treating the X-ray data not just as a simple measurement but as a complex mathematical object, the researchers have mapped the exact shape and location of these artifacts in three-dimensional space. Their work reveals that while these streaks are unavoidable in current technology, they are mathematically distinct from the real edges of the body, appearing as fainter, smoother features that occupy a specific, predictable surface within the reconstructed image.

The researchers focused on a common scanning setup where the X-ray source moves along a smooth, curved path, such as a circle or a helix, around the patient. They modeled the physics of the scan using the standard laws that describe how light is absorbed, but they paid special attention to the non-linear effects caused by the changing energy of the beam. In a simplified world where the beam's energy did not change, the math would be linear and straightforward. But because the beam hardens, the relationship between the object and the data becomes non-linear, creating new types of errors that simple linear models miss. The team used a branch of mathematics called microlocal analysis, which studies how sharp features and their directions propagate through a system, to track these errors from the raw data all the way to the final picture.

They discovered that the artifacts caused by beam hardening are not random noise. Instead, they form a very specific geometric structure. When an X-ray beam touches a dense object at two points at the same time, it generates a singularity, or a sharp mathematical feature, in the data. As the machine reconstructs the image, these singularities do not disappear; they travel into the final picture and settle on a two-dimensional surface. This surface is essentially the collection of all possible lines that can touch the object at two points while also passing through the path of the X-ray source. In a simulation involving two metal balls, the researchers visualized this surface as a distinct, curved sheet that cuts through the space between the objects. The streaks seen in the final image are simply cross-sections of this invisible sheet.

A crucial finding of the study is the difference in strength between the real image and these artifacts. The researchers proved that the sharp edges of the actual objects, such as the boundary of a metal ball, appear in the image with a certain intensity. The beam-hardening artifacts, however, are mathematically "smoother" and weaker. In technical terms, they are one degree smoother on a scale used to measure the sharpness of functions. This means that while the streaks are visible, they are inherently less intense than the true edges of the anatomy being scanned. This distinction is vital because it suggests that while we cannot easily remove these artifacts without advanced processing, they are fundamentally different from the real structures, making them theoretically easier to separate from the truth if one knows exactly where to look.

The study also addressed what happens when the objects being scanned are not perfectly smooth, such as the sharp edges of a cube or a box. In these cases, the geometry of the artifact changes slightly, but the underlying principle remains the same. The researchers extended their theory to cover these "ridge" singularities, showing that the artifacts still form on a predictable surface, though the mathematical description of their strength shifts slightly depending on whether the beam grazes a smooth curve or a sharp corner. This generalization ensures that the theory applies to a wide variety of real-world objects, not just idealized spheres.

To confirm their mathematical predictions, the team ran computer simulations of a circular scan around two metal spheres, one made of gold and the other of silver. They generated data using a realistic X-ray spectrum and reconstructed the images using standard filtering techniques. The results matched their theory perfectly. In the images reconstructed from the non-linear data, faint streaks appeared between the two balls, tracing the path of the predicted artifact surface. These streaks were noticeably fainter than the sharp boundaries of the balls themselves. The team also compared these results with images reconstructed using a different, iterative method, observing that the artifacts behaved consistently across different reconstruction algorithms, further validating their mathematical model.

This work provides a rigorous map of where and why beam-hardening artifacts appear in cone-beam CT scans. By proving that these errors are conormal distributions—meaning they have a specific, smooth structure concentrated on a particular surface—the authors have moved beyond simply observing the problem to defining its exact nature. They have shown that these artifacts are not chaotic noise but are instead highly organized features that arise from the physics of the scan. While the study does not offer a new machine to eliminate these streaks, it provides the precise mathematical blueprint needed to understand them, laying the groundwork for future methods that might suppress them more effectively by targeting their specific geometric signature.

Technical Summary: Microlocal Analysis of Non-linear Artifacts in Cone Beam CT

Problem Statement
This paper addresses the microlocal analysis of beam hardening artifacts arising in cone-beam X-ray computed tomography (CT). Unlike standard linear CT models, the physical data acquisition in CT is inherently non-linear due to the polychromatic nature of X-ray sources and the energy-dependent attenuation of materials. The forward model is governed by the Beer-Lambert law, where the measured data involves an integral over the energy spectrum of the source multiplied by an exponential term containing the line integral of the attenuation coefficient $\mu(x, E)$ .

The authors focus on the specific scenario where the set of X-ray sources is restricted to a 1D smooth curve $\gamma \subset \mathbb{R}^3$ , which is the standard model for cone-beam CT. A critical distinction of this work is that it does not rely on the restrictive assumption that the source spectrum is confined to a narrow energy bin (a common simplification in prior literature). Instead, it assumes only that the source spectrum $P(E)$ is a non-negative integrable function with positive total mass. The reconstruction target, the attenuation coefficient $\mu$ , is modeled as having jump discontinuities across a surface $S$ (representing object boundaries) while being smooth elsewhere. The primary problem is to characterize how the non-linear exponential term in the data model generates singularities that are absent in the linear X-ray transform, and how these singularities propagate to the reconstructed image via Filtered Back Projection (FBP).

Methodology
The authors employ the theory of conormal distributions and microlocal analysis to rigorously describe the nature and strength of these artifacts. The methodology proceeds through the following steps:

Physical Modeling: The data $g(L)$ is modeled as $-\log \int P(E) \exp(-\int_L \mu(x, E) dx) dE$ . The authors analyze the singularities of this non-linear operator $R\mu$ near points in the sinogram corresponding to rays tangent to the surface $S$ .
Microlocal Decomposition:
- Single Tangent Rays: For rays tangent to $S$ at a single point, the authors show that the non-linear data behaves locally as a conormal distribution. They prove that the non-linearity introduces a square-root type singularity, resulting in a distribution of order $\dot{I}^{-3/2}(\Gamma)$ , where $\Gamma$ is the manifold of tangent rays.
- Double Tangent Rays: The core of the analysis focuses on rays tangent to $S$ at two distinct points simultaneously. In this case, the non-linear term creates a product of two conormal distributions (one for each tangency point). Using the theory of paired Lagrangian distributions and the product of distributions (Lemma 2.17), the authors demonstrate that this product generates new singularities in the data that are not present in the linear transform. These new singularities lie on the intersection of the manifolds of the two tangent rays.
Propagation to Reconstruction: The authors apply the adjoint operator $R^*$ (part of the FBP reconstruction $R^*P$ ) to these singularities. By analyzing the composition of the canonical relations of the forward and adjoint operators, they determine the location and strength of the resulting artifacts in the image domain.
Geometric Conditions: The analysis requires specific geometric conditions to ensure the validity of the microlocal calculus, particularly regarding the Bolker condition. The authors introduce Assumption 4.1 and Assumption 5.1, which essentially require that the tangent planes to the source trajectory $\gamma$ do not coincide with the tangent planes of the object surface $S$ in a way that creates degeneracies. They argue these conditions are generically satisfied.
Extension to Non-Smooth Surfaces: The theory is extended to cases where $S$ contains "ridges" (edges where two smooth surface patches meet transversely, e.g., the edge of a cube). The authors define "generalized tangency" and show that the analysis holds, though the order of the resulting singularities changes depending on whether the tangency is conventional (smooth surface) or generalized (ridge).

Key Contributions and Results

Characterization of Artifact Type: The paper proves that beam hardening artifacts in cone-beam CT are locally of conormal type. Specifically, the artifacts due to non-linearities lie on a 2-D surface $L_a$ , which is the union of all double tangent rays that intersect the source trajectory $\gamma$ .
Quantification of Artifact Strength: A significant contribution is the quantification of the strength of these artifacts using the Sobolev order of the conormal distributions.
- The desired singularities (reconstructed jumps of $\mu$ ) correspond to the linear transform and are of a certain Sobolev regularity.
- The non-linear artifacts are shown to be one degree smoother (in terms of Sobolev scale) than the desired singularities. For instance, if the desired singularities are in $H^\alpha$ for $\alpha < 1/2$ , the non-linear artifacts are in $H^\beta$ for $\beta < 3/2$ . This implies that while the artifacts are present, they are theoretically weaker than the edges of the object.
Relaxation of Spectrum Assumptions: Unlike previous works (e.g., [25, 3]) that required the source spectrum to be supported on a narrow interval to perform microlocal analysis, this paper demonstrates that the analysis holds for general integrable spectra $P(E)$ .
Ridge Singularities: The authors generalize the theory to non-smooth surfaces with ridges, showing that the non-linear artifacts still form a conormal distribution on a specific surface, with the order of the distribution adjusted based on the type of tangency (smooth vs. ridge).
Validation: The theoretical predictions are validated through simulated reconstructions of metal objects (gold and silver spheres) in a circular cone-beam geometry. The simulations confirm the presence of streak artifacts along the predicted surface $L_a$ (the union of double tangents) in non-linear reconstructions, which are absent in linear reconstructions. The simulations also demonstrate that these artifacts are fainter than the primary object edges and distinct from artifacts caused by the violation of the Bolker condition (which arise from the restricted ray transform geometry itself).

Significance
The paper claims significance in providing a rigorous, non-heuristic microlocal description of beam hardening artifacts in a practically relevant setting (restricted ray transforms with general spectra). By establishing that these artifacts are conormal distributions of a specific, lower order than the true object edges, the work provides a theoretical foundation for understanding why these artifacts appear as "faint" streaks rather than sharp edges. This distinction is crucial for developing correction algorithms, as it suggests that the artifacts occupy a different regularity class than the true signal. The extension to non-smooth boundaries (ridges) further broadens the applicability of the theory to real-world objects with sharp edges. The authors explicitly note that while their theory applies to smooth regions and generic ridges, specific pathological cases where their geometric assumptions fail can be constructed, though these are unlikely in practical applications.

More like this