Pith. sign in

REVIEW 2 major objections 5 minor 34 references

FORCE-Interior shows that a Poisson-flow generative prior, anchored to the measured sinogram at every sampling step and started from a full-FOV reconstruction, improves structural and perceptual quality in interior tomography under severe R

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · deepseek-v4-flash

2026-08-02 02:26 UTC pith:7YVMDPYI

load-bearing objection A competent and mostly honest extension of FORCE to interior CT, but the headline p<0.001 claims are overstated because hyperparameters were tuned on test-patient slices and the sample is small and correlated. the 2 major comments →

arxiv 2607.14320 v1 pith:7YVMDPYI submitted 2026-07-15 eess.IV

FORCE-Interior: A Poisson Flow Generative Prior for Interior Tomography Reconstruction

Kang Chen , Wenjun Xia , Jianxu Wang , Mahmud Wasif Nafee , Ge Wang This is my paper
classification eess.IV
keywords interior tomographyregion-of-interest reconstructionPoisson flow generative modeldiffusion modelsCT reconstructiondata consistencygenerative priortruncated projections
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

Interior tomography reconstructs a small region of interest (ROI) from truncated CT projections, an ill-posed problem with non-unique solutions. This paper claims that a pre-trained Poisson-flow generative prior can be adapted to this setting by combining a full-FOV OS-SART warm start with per-step data consistency against the truncated sinogram. On a public low-dose CT benchmark with Poisson noise, the method beats conventional iterative reconstructions, a CNN baseline, and a diffusion baseline on PSNR, SSIM, and LPIPS at ROI radii of 96 and 128 pixels, with Holm-adjusted p<0.001, and remains competitive at 160 pixels. The central claim is that a generative prior alone is unreliable, but when repeatedly anchored to measurements it resolves the non-uniqueness and prevents drift to measurement-inconsistent structures.

Core claim

FORCE-Interior solves the ROI-truncated inverse problem y=MAx+n by sampling from a Poisson-flow generative prior (PFGM++) while enforcing data consistency. Two design choices carry the argument: (1) initializing the sampler with a full-FOV OS-SART reconstruction rather than an ROI-restricted one, so that out-of-ROI attenuation is accounted for and cupping/bias artifacts are reduced; (2) at each denoising step, applying OS-SART updates conditioned on the truncated sinogram before the ODE step, keeping the trajectory anchored to the measurements. Experiments report that at ROI radii 96 and 128 pixels, FORCE-Interior achieves the best ROI PSNR, SSIM, and LPIPS among OS-SART, SART-TV, ROI-CT-CNN

What carries the argument

The load-bearing mechanism is the combination of a full-FOV OS-SART warm-start and per-step data-consistency conditioning on the ROI-truncated sinogram, inside a Poisson-flow generative sampler (an EDM-based PFGM++ that learns a denoiser/score for full-dose CT images). The full-FOV initialization accounts for out-of-ROI attenuation, while per-step OS-SART updates enforce agreement with the measured truncated projections at every denoising iteration, preventing the sampling trajectory from drifting to measurement-inconsistent anatomy. A lightweight total-variation proximal step moderates noise amplification from repeated OS-SART corrections.

Load-bearing premise

The reported gains rest on the assumption that the hyperparameters (TV weight lambda=0.003 and starting noise level t_start=0.6) chosen by ablations on a small random subset are not effectively tuned to the structure-rich test slices, so the improvements are not an artifact of selection on the evaluation set.

What would settle it

A concrete check: retune the TV weight and starting noise level on a truly separate validation set disjoint from the evaluation slices, then re-run the comparison; if the PSNR/SSIM/LPIPS margins at radii 96 and 128 shrink to non-significance, the claim of out-of-the-box improvement is weakened. Additionally, applying the method to a different public CT dataset with varied anatomy and no structure-rich filtering would test whether the gains generalize beyond the specific test pool.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

Share X Bluesky LinkedIn Reddit HN

If this is right

  • If the central claim holds, generative-model-based CT reconstruction can be extended to interior tomography by explicitly modeling the truncation mask in the data-consistency update.
  • The full-FOV warm start implies that out-of-ROI attenuation must be accounted for even when only the ROI is of interest, challenging ROI-restricted initialization practices.
  • Per-step data consistency keeps projection-domain residuals near the noise floor; removing it raises residuals by roughly 24x at radius 128, indicating that the anchoring is essential.
  • The method retains small inserted lesions (false-negative rate 0) while a generative prior without data consistency misses all lesions, suggesting the anchoring preserves clinically relevant structures.
  • The advantage grows as truncation becomes more severe, so the approach is most relevant for tightly targeted ROIs where the interior problem is hardest.

Where Pith is reading between the lines

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

  • The same warm-start-plus-per-step-DC recipe could likely be applied to other generative priors (e.g., score-based diffusion) to test whether the gains stem from the PFGM++ prior specifically or from the anchoring scheme generally.
  • The noise-level sweep shows no per-dose retuning is needed, suggesting a path toward protocol-agnostic reconstruction, but the fixed TV weight chosen at the main dose may slightly oversmooth high-dose data, as the paper itself notes for LPIPS.
  • A testable extension would evaluate the method on volumetric or cone-beam geometries and on anatomy beyond the structure-rich subset, since the reported statistics are slice-level and from two patients, which may not capture full clinical variability.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

2 major / 5 minor

Summary. The paper proposes FORCE-Interior, a generative reconstruction framework for interior tomography that combines a pre-trained PFGM++ (Poisson-flow generative) prior with a full-FOV OS-SART warm start and per-step data-consistency conditioning against truncated ROI measurements. The method is evaluated on the 2016 NIH-AAPM-Mayo Low-Dose CT dataset at three ROI radii (96, 128, 160 px) under Poisson noise, against OS-SART, SART-TV, ROI-CT-CNN, and DDS. The central claim is that FORCE-Interior achieves improved structural and perceptual quality (ROI PSNR, SSIM, LPIPS) at the two smaller ROI sizes with Holm-adjusted p<0.001, and competitive results at the largest, while preserving projection-domain consistency. Ablations and additional analyses (warm-start variants, TV-weight sweep, noise-level sweep, projection-residual consistency, small-lesion retention) are provided.

Significance. If the central claim holds, the paper makes a useful methodological contribution: it demonstrates how a foundation-style generative prior can be adapted to the interior-tomography setting through a measurement-anchored warm start and per-step data consistency, rather than treating generative sampling as a plug-in post-processor. The experimental design is careful in several respects: paired Wilcoxon tests with Holm correction, multiple ROI radii, a dose sweep without per-dose retuning, projection-domain consistency metrics, and a lesion-retention analysis. These features make the evaluation substantially more informative than a simple PSNR table. However, the headline significance claims are weakened by the unresolved provenance of the hyperparameter-selection subset and by slice-level statistics over two patients, as detailed below.

major comments (2)
  1. [Section IV-D3, Fig. 5, Table V, Appendix A] The two key hyperparameters (TV weight λ=0.003 and starting noise level t_start=0.6) are selected via ablations on a 'random 20-slice subset.' The manuscript never states that this subset is disjoint from the structure-rich test set (N=34 at r=128) on which the headline comparisons in Table I are made. Appendix A mentions a separate pool of 20 held-out slices used for the radius sweep, but does not state whether the hyperparameter ablations use that pool or a random subset of the test partition. If the latter, the test data have influenced the model configuration, making the subsequent paired Wilcoxon p<0.001 results in Table VIII partially circular for those hyperparameters. Please clarify the provenance of the ablation subset, and ideally re-select hyperparameters on a truly held-out validation set and re-run the benchmark.
  2. [Appendix A, Statistical analysis; Section IV-A] The significance claims rest on slice-level paired Wilcoxon tests with N=18, 34, and 16 slices drawn from only two patients. The appendix concedes that 'adjacent slices from the same patient may not be fully independent,' but no correction is made for within-patient correlation. Treating correlated slices as independent inflates the effective sample size; with two patients, the reported Holm-adjusted p<0.001 are not credible as evidence about patient-level generalization. Please provide patient-level summaries (e.g., per-patient mean differences), a mixed-effects analysis, or a conservative cluster-based correction, and temper the abstract's significance claim accordingly.
minor comments (5)
  1. [Abstract/Introduction] Typo: 'proposeFORCE-Interior' should be 'propose FORCE-Interior'.
  2. [Table I caption] The heading 'POISSON-NOISE RECONSTRUCTION ATr∈{96,128,160}PX' appears to have a formatting error ('ATr' should be 'at r'). Also, the ROI-CT-CNN caveat (trained for r=128 only, evaluated under a shifted distribution) should appear in the table caption since the table includes it at r=128.
  3. [Fig. 5 caption / Section IV-D3] Please clarify which 'random 20-slice subset' is used for the TV-weight and t_start ablations and whether it is the same as the 'separate pool of 20 held-out slices' described in Appendix A. This is important for the reader to assess potential leakage.
  4. [Algorithm 3 / Eq. (14)] The notation M_g A_{S_g} and the update equation could be defined more explicitly; currently it is not immediately clear how the subset mask M_g interacts with the projection rows A_{S_g}. A brief sentence connecting Eq. (14) to the truncation mask M in Eq. (10) would help.
  5. [Section IV-C2 / Appendix A] The lesion-retention analysis uses '20 structure-rich test slices.' It would be useful to confirm whether these are the same 20 slices used for hyperparameter tuning or the radius-sweep pool, and to report the overlap explicitly.

Circularity Check

0 steps flagged

No significant circularity: central claims rest on held-out empirical comparisons, not on a fitted parameter or self-citation chain; noted caveats are statistical, not constructional.

full rationale

The paper does not derive its headline result from its own inputs by construction. FORCE-Interior is defined by explicit equations (Algorithm 3; Eqs. (13)-(14)) combining a pretrained PFGM++ prior, full-FOV OS-SART warm start, per-step OS-SART data consistency, and TV proximal steps, and it is evaluated against OS-SART, SART-TV, ROI-CT-CNN, and DDS on two held-out Mayo patients. The central metric improvements are empirical outcomes, not identities. The reliance on the authors' own FORCE/PFGM++ prior is a normal self-citation: the prior's training objective (Eq. (7)) does not include the interior-tomography benchmark, and the measured gains come from the experimental protocol rather than from the prior's objective. Likewise, the projection-consistency results are not circular: they are presented as a check of the enforced constraint, with an ablation (w/o per-step DC) showing a 24x residual increase, so the analysis has independent content. The appendix explicitly notes that adjacent slices from the same two patients may not be fully independent, and the TV weight/t_start were selected on a random 20-slice subset (Section IV-D3, Table V); these are statistical validity caveats (potential dependence/selection on evaluation data) rather than a reduction of prediction to input. No equation in the paper equates a fitted parameter with the reported outcome by construction, and no load-bearing claim rests solely on an unverified self-citation. Hence no significant circularity; score 1 reflecting only minor self-citation and mild statistical caveats.

Axiom & Free-Parameter Ledger

4 free parameters · 4 axioms · 0 invented entities

The method does not postulate new physical entities. It relies on standard generative-model theory, standard CT reconstruction theory, and several task-specific design choices that are validated empirically. The main free parameters are the TV weight and starting noise level, both selected by ablations on subsets of the evaluation data.

free parameters (4)
  • TV weight lambda = 0.003
    Chosen by sweeping lambda on a random 20-slice subset; SSIM peaks at 0.003 (Fig. 5, Section IV-D3).
  • Starting noise level t_start = 0.6
    Chosen from ablations showing best PSNR/SSIM/LPIPS (Table V, Section IV-D1); no separate validation set is stated.
  • Per-step OS-SART settings = G=16 subsets, 3 iterations, relaxation omega=1
    Fixed implementation choices in Appendix A; they control data-consistency strength but are not swept.
  • Warm-start OS-SART iterations = 30 iterations, 16 subsets
    Fixed in Appendix A; full-FOV variant is justified by ablation vs ROI-restricted warm start.
axioms (4)
  • standard math OS-SART converges as a data-consistency operator for truncated projections
    Used in Eq. (14) and in the warm start; standard iterative reconstruction theory from [8].
  • domain assumption The PFGM++/EDM denoiser approximates the score of the normal-dose CT image distribution
    The pretrained prior is the core generative mechanism, based on [26-28]; it is assumed to transfer to the held-out patients and to the interior-truncation task.
  • ad hoc to paper A full-FOV OS-SART warm start from truncated projections captures out-of-ROI attenuation
    This design principle is validated only by the warm-start ablation (Table IV), not by an independent derivation.
  • domain assumption Slice-level paired statistics on adjacent slices approximate independent samples
    Acknowledged in Appendix A: test slices come from two patients and adjacent slices may not be fully independent.

pith-pipeline@v1.3.0-alltime-deepseek · 15493 in / 10372 out tokens · 104319 ms · 2026-08-02T02:26:23.709204+00:00 · methodology

0 comments
Cite this review

Pith. "Pith review of FORCE-Interior: A Poisson Flow Generative Prior for Interior Tomography Reconstruction." pith.science (2026). https://pith.science/paper/7YVMDPYI

@misc{pith2026260714320,
  author       = {Pith},
  title        = {Pith review of: FORCE-Interior: A Poisson Flow Generative Prior for Interior Tomography Reconstruction},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/7YVMDPYI}},
  note         = {Machine review of arXiv:2607.14320}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Interior tomography reconstructs a region of interest (ROI) from truncated projection measurements. However, projection truncation makes the inverse problem severely ill-posed, leading to non-unique solutions, oversmoothing, and truncation-induced artifacts when conventional reconstruction methods are directly applied to interior tomography. Existing learning-based interior CT methods have shown promising performance, but their generalization across different truncation patterns, ROI sizes, and noise levels remains an important challenge. Meanwhile, current generative model-based reconstruction methods are primarily designed for non-interior tomography settings and do not directly address ROI-based projection truncation. Moreover, without sufficient data-consistency constraints, generative sampling may yield anatomically plausible but measurement-inconsistent structures. To address these challenges, we propose FORCE-Interior, a Poisson-flow generative reconstruction framework for interior tomography. FORCE-Interior combines a full-FOV measurement-constrained initialization with per-step data consistency for ROI-truncated measurements, anchoring generative sampling to the acquired measurements throughout reconstruction. Experiments show that FORCE-Interior achieves improved structural and perceptual reconstruction quality at the two more severely truncated ROI sizes, with competitive results at the largest, while maintaining projection-domain consistency.

Figures

Figures reproduced from arXiv: 2607.14320 by Ge Wang, Jianxu Wang, Kang Chen, Mahmud Wasif Nafee, Wenjun Xia.

Figure 1
Figure 1. Figure 1: Eccentric-ROI reconstruction at r = 128 px, window [−160, 240] HU. perceptual image patch similarity (LPIPS) [30], a deep-feature￾based perceptual distance for which a lower value indicates better perceptual fidelity to the reference. All quantitative metrics are computed within the ROI in the [0, 2000] HU window, which measures overall reconstruction fidelity over the full dynamic range and is consistent … view at source ↗
Figure 2
Figure 2. Figure 2: Eccentric-ROI reconstructions at r = 96/128/160 px, window [−160, 240] HU. Interior and each baseline on the slice-wise PSNR, SSIM, and LPIPS values (N=18, 34, and 16 pairs at ROI radii 96, 128, and 160 px, respectively). Paired differences are defined so that positive values favor FORCE-Interior (Ours−Baseline for PSNR and SSIM; Baseline−Ours for LPIPS). To account for multiple comparisons, Holm correctio… view at source ↗
Figure 3
Figure 3. Figure 3: Small-lesion retention comparison, window [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figure 3
Figure 3. Figure 3: Without per-step data consistency, FORCE-Interior can [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figure 5
Figure 5. Figure 5: Effect of the TV weight λ under Poisson noise: ROI PSNR, SSIM, and LPIPS. SSIM peaks at λ=0.003 while LPIPS is minimized near λ=0.002. The shaded bands mark the under-regularized, optimal, and over-smoothed regimes. quisition conditions before it can be considered for clinical deployment. APPENDIX A IMPLEMENTATION DETAILS Geometry and measurement. The forward model is parallel-beam with 720 views uniformly… view at source ↗
Figure 6
Figure 6. Figure 6: Noise-level sweep with every method fixed at its main-paper configuration. ROI PSNR, SSIM, and LPIPS versus photon count [PITH_FULL_IMAGE:figures/full_fig_p011_6.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

34 extracted references · 1 canonical work pages

  1. [1]

    The meaning of interior tomography,

    G. Wang and H. Yu, “The meaning of interior tomography,”Phys. Med. Biol., vol. 58, no. 16, pp. R161–R186, 2013

    2013

  2. [2]

    Compressed sensing based interior tomography,

    H. Yu and G. Wang, “Compressed sensing based interior tomography,” Phys. Med. Biol., vol. 54, no. 9, pp. 2791–2805, 2009

    2009

  3. [3]

    Image reconstruction for sparse-view CT and interior CT: Introduction to compressed sensing and differentiated backprojection,

    H. Kudo, T. Suzuki, and E. A. Rashed, “Image reconstruction for sparse-view CT and interior CT: Introduction to compressed sensing and differentiated backprojection,”Quant. Imaging Med. Surg., vol. 3, no. 3, pp. 147–161, 2013

    2013

  4. [4]

    A practical local tomography reconstruction algorithm based on a known sub-region,

    P. Paleo, M. Desvignes, and A. Mirone, “A practical local tomography reconstruction algorithm based on a known sub-region,”Journal of Synchrotron Radiation, vol. 24, no. 1, pp. 257–268, 2017

    2017

  5. [5]

    A. C. Kak and M. Slaney,Principles of Computerized Tomographic Imaging. New York, NY , USA: IEEE Press, 1988

    1988

  6. [6]

    FBP and the interior problem in 2D tomography,

    A. Bilgot, L. Desbat, and V . Perrier, “FBP and the interior problem in 2D tomography,” in2011 IEEE Nuclear Science Symposium Conference Record, 2011, pp. 4080–4085

    2011

  7. [7]

    Simultaneous algebraic reconstruction technique (SART): A superior implementation of the ART algorithm,

    A. H. Andersen and A. C. Kak, “Simultaneous algebraic reconstruction technique (SART): A superior implementation of the ART algorithm,” Ultrason. Imaging, vol. 6, no. 1, pp. 81–94, 1984

    1984

  8. [8]

    Ordered-subset simultaneous algebraic recon- struction techniques (OS-SART),

    G. Wang and M. Jiang, “Ordered-subset simultaneous algebraic recon- struction techniques (OS-SART),”J. X-Ray Sci. Technol., vol. 12, no. 3, pp. 169–177, 2004

    2004

  9. [9]

    Bone-induced streak artifact suppression in sparse-view CT image reconstruction,

    S. O. Jin, J. G. Kim, S. Y . Lee, and O.-K. Kwon, “Bone-induced streak artifact suppression in sparse-view CT image reconstruction,”BioMed. Eng. OnLine, vol. 11, p. 44, 2012

    2012

  10. [10]

    Artifact reduction methods for truncated projections in iterative breast tomosynthesis reconstruction,

    Y . Zhang, H.-P. Chan, B. Sahiner, J. Wei, C. Zhou, and L. M. Hadjiiski, “Artifact reduction methods for truncated projections in iterative breast tomosynthesis reconstruction,”J. Comput. Assist. Tomogr ., vol. 33, no. 3, pp. 426–435, 2009

    2009

  11. [11]

    A diffusion-based truncated projection artifact reduction method for iterative digital breast tomosynthesis reconstruction,

    Y . Lu, H.-P. Chan, J. Wei, and L. M. Hadjiiski, “A diffusion-based truncated projection artifact reduction method for iterative digital breast tomosynthesis reconstruction,”Phys. Med. Biol., vol. 58, no. 3, pp. 569– 587, 2013

    2013

  12. [12]

    Image reconstruction in circular cone-beam computed tomography by constrained, total-variation minimization,

    E. Y . Sidky and X. Pan, “Image reconstruction in circular cone-beam computed tomography by constrained, total-variation minimization,” Phys. Med. Biol., vol. 53, no. 17, pp. 4777–4807, 2008

    2008

  13. [13]

    Prior image constrained compressed sensing (PICCS): A method to accurately reconstruct dynamic CT images from highly undersampled projection data sets,

    G.-H. Chen, J. Tang, and S. Leng, “Prior image constrained compressed sensing (PICCS): A method to accurately reconstruct dynamic CT images from highly undersampled projection data sets,”Med. Phys., vol. 35, no. 2, pp. 660–663, 2008

    2008

  14. [14]

    Solving inverse problems using data-driven models,

    S. Arridge, P. Maass, O. ¨Oktem, and C.-B. Sch ¨onlieb, “Solving inverse problems using data-driven models,”Acta Numer ., vol. 28, pp. 1–174, 2019

    2019

  15. [15]

    Low-dose ct reconstruction via edge-preserving total variation regularization,

    Z. Tian, X. Jia, K. Yuan, T. Pan, and S. B. Jiang, “Low-dose ct reconstruction via edge-preserving total variation regularization,” Physics in Medicine and Biology, vol. 56, no. 18, p. 5949–5967, Aug. 2011. [Online]. Available: http://dx.doi.org/10.1088/0031- 9155/56/18/011

    doi:10.1088/0031- 2011

  16. [16]

    Deep learning interior tomography for region-of-interest reconstruction,

    Y . Han, J. Gu, and J. C. Ye, “Deep learning interior tomography for region-of-interest reconstruction,” 2018. [Online]. Available: https://arxiv.org/abs/1712.10248

    Pith/arXiv arXiv 2018

  17. [17]

    One network to solve all rois: Deep learning ct for any roi using differentiated backprojection,

    Y . Han and J. C. Ye, “One network to solve all rois: Deep learning ct for any roi using differentiated backprojection,” 2019. [Online]. Available: https://arxiv.org/abs/1810.00500

    Pith/arXiv arXiv 2019

  18. [18]

    End-to-end deep learning for interior tomography with low-dose x-ray ct,

    Y . Han, D. Wu, K. Kim, and Q. Li, “End-to-end deep learning for interior tomography with low-dose x-ray ct,” 2025. [Online]. Available: https://arxiv.org/abs/2501.05085

    Pith/arXiv arXiv 2025

  19. [19]

    Solving inverse problems in medical imaging with score-based generative models,

    Y . Song, L. Shen, L. Xing, and S. Ermon, “Solving inverse problems in medical imaging with score-based generative models,” inProc. Int. Conf. Learn. Represent. (ICLR), 2022

    2022

  20. [20]

    Dolce: A model-based probabilistic diffusion framework for limited-angle ct reconstruction,

    J. Liu, R. Anirudh, J. J. Thiagarajan, S. He, K. A. Mohan, U. S. Kamilov, and H. Kim, “Dolce: A model-based probabilistic diffusion framework for limited-angle ct reconstruction,” 2022. [Online]. Available: https://arxiv.org/abs/2211.12340

    Pith/arXiv arXiv 2022

  21. [21]

    Generative modeling in sinogram domain for sparse-view ct reconstruction,

    B. Guan, C. Yang, L. Zhang, S. Niu, M. Zhang, Y . Wang, W. Wu, and Q. Liu, “Generative modeling in sinogram domain for sparse-view ct reconstruction,” 2022. [Online]. Available: https://arxiv.org/abs/2211.13926

    Pith/arXiv arXiv 2022

  22. [22]

    Tomographic foundation model – force: Flow-oriented reconstruction conditioning engine,

    W. Xia, C. Niu, and G. Wang, “Tomographic foundation model – force: Flow-oriented reconstruction conditioning engine,” 2025. [Online]. Available: https://arxiv.org/abs/2506.02149

    Pith/arXiv arXiv 2025

  23. [23]

    Decomposed diffusion sampler for accelerating large-scale inverse problems,

    H. Chung, S. Lee, and J. C. Ye, “Decomposed diffusion sampler for accelerating large-scale inverse problems,” 2024. [Online]. Available: https://arxiv.org/abs/2303.05754

    Pith/arXiv arXiv 2024

  24. [24]

    Denoising diffusion probabilistic models,

    J. Ho, A. Jain, and P. Abbeel, “Denoising diffusion probabilistic models,” 2020. [Online]. Available: https://arxiv.org/abs/2006.11239

    Pith/arXiv arXiv 2020

  25. [25]

    Score-based generative modeling through stochastic differ- ential equations,

    Y . Song, J. Sohl-Dickstein, D. P. Kingma, A. Kumar, S. Ermon, and B. Poole, “Score-based generative modeling through stochastic differ- ential equations,” inProc. Int. Conf. Learn. Represent. (ICLR), 2021

    2021

  26. [26]

    Elucidating the design space of diffusion-based generative models,

    T. Karras, M. Aittala, T. Aila, and S. Laine, “Elucidating the design space of diffusion-based generative models,” inProc. Adv. Neural Inf. Process. Syst., 2022

    2022

  27. [27]

    Poisson flow generative models,

    Y . Xu, Z. Liu, M. Tegmark, and T. Jaakkola, “Poisson flow generative models,” inProc. Adv. Neural Inf. Process. Syst., 2022

    2022

  28. [28]

    PFGM++: Unlocking the potential of physics-inspired generative mod- els,

    Y . Xu, Z. Liu, Y . Tian, S. Tong, M. Tegmark, and T. Jaakkola, “PFGM++: Unlocking the potential of physics-inspired generative mod- els,” inProc. Int. Conf. Mach. Learn. (ICML), 2023, pp. 38 566–38 591

    2023

  29. [29]

    Image quality assessment: From error visibility to structural similarity,

    Z. Wang, A. C. Bovik, H. R. Sheikh, and E. P. Simoncelli, “Image quality assessment: From error visibility to structural similarity,”IEEE Trans. Image Process., vol. 13, no. 4, pp. 600–612, 2004

    2004

  30. [30]

    The unreasonable effectiveness of deep features as a perceptual metric,

    R. Zhang, P. Isola, A. A. Efros, E. Shechtman, and O. Wang, “The unreasonable effectiveness of deep features as a perceptual metric,” in Proc. IEEE Conf. Comput. Vis. Pattern Recognit. (CVPR), 2018, pp. 586–595

    2018

  31. [31]

    A first-order primal-dual algorithm for convex problems with applications to imaging,

    A. Chambolle and T. Pock, “A first-order primal-dual algorithm for convex problems with applications to imaging,”J. Math. Imaging Vis., vol. 40, no. 1, pp. 120–145, 2011

    2011

  32. [32]

    Tigre v3: Efficient and easy to use iterative computed tomographic reconstruction toolbox for real datasets,

    A. Biguri, T. Sadakane, R. Lindroos, Y . Liu, M. Sabat´e Landman, Y . Du, M. Lohvithee, S. Kaser, S. Hatamikia, R. Bryll, E. Valat, S. Wonglee, T. Blumensath, and C.-B. Sch ¨onlieb, “Tigre v3: Efficient and easy to use iterative computed tomographic reconstruction toolbox for real datasets,”Engineering Research Express, vol. 7, no. 1, p. 015011, mar

  33. [33]

    DM4CT: Benchmarking dif- fusion models for computed tomography reconstruction,

    J. Shi, D. M. Pelt, and K. J. Batenburg, “DM4CT: Benchmarking dif- fusion models for computed tomography reconstruction,”arXiv preprint arXiv:2602.18589, 2026

    arXiv 2026

  34. [2025]

    Available: https://doi.org/10.1088/2631-8695/adbb3a

    [Online]. Available: https://doi.org/10.1088/2631-8695/adbb3a

    work page doi:10.1088/2631-8695/adbb3a