arxiv: 2605.11637 · v1 · submitted 2026-05-12 · ⚛️ physics.med-ph · cond-mat.dis-nn· physics.data-an

Recognition: no theorem link

Computed Tomography Reconstruction Algorithm Using Markov Random Field Model

Akihisa Takeuchi, Hayaru Shouno, Masato Okada, Masayuki Uesugi, Taichi Kusumi, Taiga Shimomiya, Yuichi Yokoyama, Yuki Sada

Pith reviewed 2026-05-13 01:29 UTC · model grok-4.3

classification ⚛️ physics.med-ph cond-mat.dis-nnphysics.data-an

keywords computed tomographyMarkov random fieldBayesian inferenceimage reconstructionfiltered back projectionlow-dose imagingsparse-view CThyperparameter estimation

0 comments

The pith

A Markov random field Bayesian algorithm reconstructs CT images more accurately than filtered back projection when X-ray dose or view count is limited.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper proposes using a Markov random field as a prior in Bayesian reconstruction of computed tomography images. It shows through simulations that this method outperforms the standard filtered back projection algorithm in conditions with low radiation dose or limited projection views. Hyperparameters are chosen automatically by minimizing the Bayesian free energy, which lets the reconstruction adapt to the specific noise in the data. This approach could extend CT use to cases where minimizing radiation or scan time is critical, such as medical imaging of sensitive patients or industrial inspections with time constraints.

Core claim

The authors develop a Bayesian CT reconstruction method that incorporates a Markov random field model to represent the statistical structure of the images. By estimating hyperparameters through minimization of the Bayesian free energy, the algorithm adapts to the noise characteristics of the projection data. Simulations demonstrate superior performance over filtered back projection under both low-dose and sparse-view conditions.

What carries the argument

Markov random field prior in a Bayesian framework, with hyperparameters optimized by Bayesian free energy minimization, to model and reconstruct CT images from projections.

If this is right

The method produces higher quality reconstructions from noisy or incomplete projection data.
It enables adaptive tuning without manual hyperparameter selection.
CT becomes viable for dose-sensitive applications.
Time-constrained measurements with limited views are supported.
Overall, it broadens the practical range of CT imaging.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

The approach may extend to other inverse problems in imaging where similar statistical priors apply.
Real-world validation on experimental data rather than simulations could confirm the gains.
Integration with machine learning could further enhance the prior model.

Load-bearing premise

The Markov random field model accurately captures the statistical structure of actual CT images, and minimizing the Bayesian free energy reliably estimates hyperparameters that match the real noise in the projections.

What would settle it

Performing the reconstruction on actual experimental CT projection data with known ground truth images and comparing quantitative metrics like mean squared error or structural similarity index against filtered back projection results under matched low-dose and sparse-view settings.

Figures

Figures reproduced from arXiv: 2605.11637 by Akihisa Takeuchi, Hayaru Shouno, Masato Okada, Masayuki Uesugi, Taichi Kusumi, Taiga Shimomiya, Yuichi Yokoyama, Yuki Sada.

**Figure 1.** Figure 1: Schematic diagram of the Radon transform used in this study. The object ξ(x, y) is integrated along lines parallel to t to obtain the projection τ(s, θ) on the detector axis s, which forms an angle θ with the x axis. For each θ, the noiseless Radon transform18) is defined as the line integral of ξ(x, y) along the direction t at fixed s. Assuming additive white Gaussian noise, the observed projection data τ… view at source ↗

**Figure 2.** Figure 2: Ground-truth images used in the numerical experiments. (a) Image generated from the MRF model corresponding to the first term of Eq. (3). (b) Shepp–Logan phantom.1, 7) 3.1 Reconstruction of the synthetic image generated by MRF model 3.1.1 Reconstruction performance under various noise intensity We first evaluate the reconstruction performance by varying the noise intensity with a fixed number of projection… view at source ↗

**Figure 3.** Figure 3: Reconstruction procedures in the numerical experiments. (a) FBP: from the observed projection data τ(s, θ), frequency-domain filters (Ramp, Shepp–Logan, and Hamming) are applied, and the image is reconstructed by back projection. (b) Proposed algorithm: the hyperparameters (γ, β, h) are estimated by minimizing the Bayesian free energy F(γ, β, h | τ˜) via grid search; the MAP estimate in Eq. (8) is then ev… view at source ↗

**Figure 4.** Figure 4: Reconstruction performance obtained by fixing the number of projections at Nθ = 1800 and varying the noise intensity. (a) Reconstruction results of the proposed algorithm and FBP for the representative noise conditions σnoise = 1.0 and 4.0, and magnified views of the regions indicated by the boxes. (b) Error images, defined as the reconstructed image minus the ground truth. The color bars indicate the sign… view at source ↗

**Figure 5.** Figure 5: Reconstruction performance obtained by fixing the noise intensity and varying the number of projections. (a) Reconstruction results of the proposed algorithm and FBP for the representative conditions Nθ = 1800 and 450, and magnified views of the regions indicated by the boxes. (b) Error images, defined as the reconstructed image minus the ground truth. The color bars indicate the signed error scale. (c) … view at source ↗

**Figure 6.** Figure 6: Comparison of reconstruction results obtained by fixing the number of projections at Nθ = 1800 and varying the noise intensity σnoise as 1.0, 2.0, and 4.0. (a) Reconstruction results of the proposed algorithm and FBP with different filters under each noise condition, together with magnified views of the regions indicated by the boxes. (b) Horizontal center line profiles of the magnified regions. The red li… view at source ↗

**Figure 7.** Figure 7: Comparison of reconstruction results obtained by fixing the noise intensity at σnoise = 2.0 and varying the number of projections as Nθ = 1800, 900, and 450. (a) Reconstruction results of the proposed algorithm and FBP with different filters for each projection number, together with magnified views of the regions indicated by the boxes. (b) Horizontal center line profiles of the magnified regions. The red … view at source ↗

read the original abstract

X-ray computed tomography (CT) reveals the materials' internal structures non-destructively from a tilt series of projected images. Filtered back projection (FBP) is a widely-adopted reconstruction algorithm in CT owing to its small computational cost. Under low-dose or sparse-view conditions, however, FBP often amplifies noise, severely degrading the reconstructed images. In this study, we evaluated the performance of a Bayesian CT reconstruction algorithm based on the Markov random field model under such adverse conditions. Through simulations, we demonstrated that the proposed algorithm shows higher reconstruction performance than FBP under both low-dose and sparse-view conditions. The hyperparameters are estimated by minimizing the Bayesian free energy, enabling adaptive reconstruction that reflects the noise characteristics of the observed projection data. These results suggest that the proposed algorithm can broaden the applicability of CT to dose-sensitive applications and time-constrained measurements, where only limited observed projection data are available.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

This paper applies a standard MRF Bayesian prior to CT reconstruction with free-energy hyperparameter tuning and shows simulation gains over FBP in low-dose and sparse-view cases, but the advance is mostly engineering rather than conceptual.

read the letter

The main takeaway is that this is a competent implementation of an existing Bayesian approach rather than a new framework. They model the image with a Markov random field prior, reconstruct via what looks like MAP estimation, and tune the hyperparameters by minimizing the Bayesian free energy on the observed projections. In simulations this beats filtered back projection when dose drops or views are thinned out, which matches the abstract claim and is the practical point they emphasize.

Referee Report

2 major / 1 minor

Summary. The paper proposes a Bayesian CT reconstruction algorithm that employs a Markov random field (MRF) prior to improve image quality over filtered back-projection (FBP) under low-dose and sparse-view conditions. Hyperparameters are estimated adaptively by minimizing the Bayesian free energy derived from the observed projection data, and the method is evaluated via simulations demonstrating superior performance.

Significance. If the simulation results prove robust with quantitative validation, the adaptive hyperparameter estimation via free-energy minimization could offer a practical way to tailor reconstructions to noise characteristics, extending CT utility in dose-limited or time-constrained settings such as medical imaging or industrial inspection.

major comments (2)

[Abstract] Abstract: the central claim of 'higher reconstruction performance than FBP' is unsupported by any quantitative metrics (e.g., RMSE, PSNR, SSIM), error bars, specific simulation parameters (photon flux for low-dose, projection count for sparse-view), or baseline details, rendering the performance assertion unverifiable.
[Hyperparameter Estimation] Hyperparameter estimation section: minimizing Bayesian free energy is defined in terms of the same MRF prior and data likelihood; without explicit equations demonstrating that the procedure is independent of the fitted quantities rather than self-consistent by construction, the adaptivity claim risks circularity.

minor comments (1)

[Abstract] The abstract would benefit from a brief statement of the specific MRF neighborhood and potential functions used.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for their thoughtful review and constructive comments on our manuscript. We have carefully considered each point and provide our responses below. We believe the revisions will strengthen the paper.

read point-by-point responses

Referee: [Abstract] Abstract: the central claim of 'higher reconstruction performance than FBP' is unsupported by any quantitative metrics (e.g., RMSE, PSNR, SSIM), error bars, specific simulation parameters (photon flux for low-dose, projection count for sparse-view), or baseline details, rendering the performance assertion unverifiable.

Authors: We agree with the referee that the abstract should provide more specific quantitative support for our claims to make them verifiable. Although the manuscript includes simulation results with performance comparisons in the main text, we will revise the abstract to include key quantitative metrics such as RMSE, PSNR, and SSIM values, along with specific simulation parameters (e.g., photon flux for low-dose cases and number of projections for sparse-view). Error bars from multiple simulations will also be mentioned where applicable. This will be incorporated in the revised version. revision: yes
Referee: [Hyperparameter Estimation] Hyperparameter estimation section: minimizing Bayesian free energy is defined in terms of the same MRF prior and data likelihood; without explicit equations demonstrating that the procedure is independent of the fitted quantities rather than self-consistent by construction, the adaptivity claim risks circularity.

Authors: We appreciate this important clarification. The minimization of the Bayesian free energy (which approximates the negative log evidence) is performed with respect to the hyperparameters, while marginalizing or integrating over the image variables using the MRF prior and likelihood. This is not circular because the free energy is a function of the hyperparameters only after the integration. To address the concern, we will add explicit mathematical derivations and equations in the revised manuscript to demonstrate the independence and show the optimization procedure step by step, clarifying that it is not self-consistent by construction but follows standard variational Bayesian inference principles. revision: yes

Circularity Check

0 steps flagged

No significant circularity detected

full rationale

The paper describes a Bayesian reconstruction algorithm using an MRF prior, with hyperparameters estimated via minimization of the Bayesian free energy to adapt to observed noise. This is a standard, non-circular procedure in statistical image processing: the free energy is computed from the joint model and data to optimize hyperparameters, without redefining inputs or forcing predictions by construction. Performance is demonstrated via independent simulations against FBP under low-dose and sparse-view conditions. No equations, self-citations, or steps in the abstract or description reduce the central claims to tautologies, fitted renamings, or load-bearing self-references. The derivation chain is self-contained and externally benchmarked.

Axiom & Free-Parameter Ledger

1 free parameters · 2 axioms · 0 invented entities

The central claim rests on the assumption that an MRF prior plus Bayesian free-energy minimization yields superior reconstructions; the paper introduces no new entities but relies on standard statistical assumptions.

free parameters (1)

MRF hyperparameters
Estimated by minimizing Bayesian free energy from the projection data; their specific values are not reported in the abstract.

axioms (2)

domain assumption Markov random field model accurately captures local pixel dependencies in CT images
Invoked to define the image prior; appears in the description of the Bayesian reconstruction algorithm.
domain assumption Bayesian free energy minimization recovers noise-appropriate hyperparameters
Used to enable adaptive reconstruction; stated in the abstract as the mechanism for hyperparameter estimation.

pith-pipeline@v0.9.0 · 5486 in / 1539 out tokens · 37969 ms · 2026-05-13T01:29:50.571803+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

20 extracted references · 20 canonical work pages

[1]

A. C. Kak and M. Slaney,Principles of Computerized Tomographic Imaging(SIAM, Philadelphia, 2001)

work page 2001
[2]

G. T. Herman,Fundamentals of Computerized Tomography(Springer, London, 2009)

work page 2009
[3]

M. J. Willemink and P. B. No¨el, Eur. Radiol.29, 2185–2195 (2019)

work page 2019
[4]

E. Y . Sidky and X. Pan, Phys. Med. Biol.53, 4777–4807 (2008)

work page 2008
[5]

Natterer,The Mathematics of Computerized Tomography(SIAM, Philadelphia, 2001)

F. Natterer,The Mathematics of Computerized Tomography(SIAM, Philadelphia, 2001)

work page 2001
[6]

G. N. Ramachandran and A. V . Lakshminarayanan, Proc. Natl. Acad. Sci. USA68, 2236 (1971)

work page 1971
[7]

L. A. Shepp and B. F. Logan, IEEE Trans. Nucl. Sci.21, 21 (1974)

work page 1974
[8]

F. J. Harris, Proc. IEEE66, 51 (1978)

work page 1978
[9]

Lange and J

K. Lange and J. A. Fessler, IEEE Trans. Image Process.4, 1430–1438 (1995)

work page 1995
[10]

Geman and D

S. Geman and D. Geman, IEEE Trans. Pattern Anal. Mach. Intell.6, 721–741 (1984)

work page 1984
[11]

Shouno and M

H. Shouno and M. Okada, J. Phys. Soc. Jpn.79, 074004 (2010)

work page 2010
[12]

Nakanishi-Ohno, K

Y . Nakanishi-Ohno, K. Nagata, H. Shouno, and M. Okada, J. Phys. A: Math. Theor.47, 045001 (2014)

work page 2014
[13]

Sakamoto, Y

H. Sakamoto, Y . Nakanishi-Ohno, and M. Okada, J. Phys. Soc. Jpn.85, 063801 (2016)

work page 2016
[14]

Katakami, H

S. Katakami, H. Sakamoto, S. Murata, and M. Okada, J. Phys. Soc. Jpn.86, 064801 (2017)

work page 2017
[15]

Kusumi, S

T. Kusumi, S. Katakami, R. Ishikawa, K. Kawahara, N. Shibata, and M. Okada, Ultra- microscopy253, 113811 (2023)

work page 2023
[16]

Kusumi, S

T. Kusumi, S. Katakami, R. Ishikawa, K. Kawahara, T. Mullarkey, J. M. Bekkevold, J. J. P. Peters, L. Jones, N. Shibata, and M. Okada, Ultramicroscopy264, 113996 (2024)

work page 2024
[17]

C. M. Bishop,Pattern Recognition and Machine Learning(Springer, New York, 2006)

work page 2006
[18]

Radon, IEEE Trans

J. Radon, IEEE Trans. Med. Imaging5, 170–176 (1986)

work page 1986
[19]

A. N. Tikhonov and V . Y . Arsenin,Solutions of Ill-Posed Problems(Winston, Washing- ton, 1977)

work page 1977
[20]

S. Z. Li,Markov Random Field Modeling in Image Analysis, 3rd ed. (Springer, London, 2009). 17/17

work page 2009