Pith. sign in

REVIEW 6 major objections 4 minor 45 references

ResPF: Residual Poisson Flow for Efficient and Physically Consistent Sparse-View CT Reconstruction

T0 review · 6 major / 4 minor · reviewed 2026-08-07 · deepseek-v4-flash

Pith's one-line read This paper claims that residual fusion lets a conditional Poisson flow generative model enforce projection-domain data consistency without destroying the ODE sampling trajectory, yielding fast, physically consistent sparse-view CT…

desk verdict The framework is promising and the experiments are extensive, but the algorithm contradicts its own trajectory-preservation claim and the EDM baseline table looks wrong. read the letter →

arxiv 2506.06400 v1 pith:IYKWAI5R submitted 2025-06-06 eess.IV cs.CV

classification eess.IVcs.CV
keywords sparse-viewCTPoissonflowgenerativemodelconditionalgenerationresidualfusiondataconsistencyhijackedsamplingreconstructionPFGM++
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

This paper introduces ResPF, a generative method that reconstructs sparse-view CT from far fewer projections than usual, and claims it is the first use of Poisson flow generative models for this task. The method builds on PFGM++ by turning it into a conditional model guided by a fast FBP image, skipping the early random-noise stages of sampling through a 'hijacking' strategy, and then enforcing agreement with the measured projections using the ASD-POCS algorithm. To keep these physics corrections from destroying the deterministic ODE trajectory that Poisson flow sampling follows, the paper inserts a residual fusion step that linearly combines the generative output with the data-consistent output at every iteration. On simulated data with 63 and 125 views and on a clinical 123-view cardiac scan, it reports better PSNR, SSIM, and LPIPS than the compared CNN, iterative, and diffusion baselines, at about 1.5 seconds per slice. The point that would matter in practice is that strong projection-domain constraints can be combined with a fast learned generative prior without losing either speed or physical fidelity.

What carries the argument

The load-bearing element is the residual linear fusion module, which at each sampling step linearly combines the generative output with the ASD-POCS data-consistent output and feeds the result back as the next state of the Poisson-flow ODE. Its job is to allow hard data consistency (total-variation minimization with nonnegativity) to enter the loop without, in the paper's argument, kicking the sample off the deterministic electric-field trajectory. The other components are the conditional PFGM++ backbone trained on triples of perturbed image, FBP condition, and noise scale, the hijacking initialization that starts sampling at an intermediate timestep, and the EDM-style Heun sampler with a 16-step noise schedule. The fusion coefficient $\alpha = 0.4$ and the hijack start (step 14 of 16) are set by grid search on the simulation dataset.

What would settle it

Run ResPF and record, at each of its 16 steps, the angle between the model-predicted Poisson-field direction $d_i$ and the actual state update $x_{i+1} - x_i$ after residual fusion. If the trajectory-preservation claim is right, these angles stay small; if the fused update diverges strongly from $d_i$ at the hijack or fusion points, the central mechanism is not doing what the paper says.

Watch

Extended reading notes

Core claim

On the paper's own terms, the central discovery is that conditional Poisson flow sampling can be truncated (hijacked) and corrected with projection-domain data consistency without sacrificing reconstruction quality, provided the correction is fused back into the trajectory rather than injected as a hard replacement. The authors extend PFGM++ with a conditional denoiser $f_\theta(x_t, x_{\mathrm{sp}}, \sigma)$ trained to map a perturbed image plus an FBP reconstruction to the clean image, then run 16 Heun steps starting from a noisy version of the FBP image at step 14. Each step computes a generative estimate, refines it with 10 ASD-POCS iterations, and forms $\hat{x}_{\mathrm{fused}} = \alpha \hat{x}_{\mathrm{gen}} + (1-\alpha) \hat{x}_{\mathrm{phys}}$ with $\alpha = 0.4$, which becomes the next ODE state. The reported result is consistent best-in-class metrics across all tested sparsity levels, with the largest gains at 63 views, where the compared EDM diffusion baseline degrades sharply. The authors conclude that the residual fusion preserves the continuity of the electrostatic-field trajectory, making the method both fast and physically consistent.

Load-bearing premise

The assumption that carries the argument is that feeding the fused image back as the next ODE state keeps the sample on the learned Poisson-flow trajectory, so the structural prior is not degraded by the physics correction.

Editorial extensions

If this is right

  • Sparse-view CT could be reconstructed in about 1.5–1.7 seconds per slice at 63–125 views while beating FBPConvNet, SwinIR, RED-CNN, DDS, PPFM, and EDM on PSNR, SSIM, and LPIPS in this paper's experiments.
  • Because PFGM++ reduces to EDM as the augmented dimension $D \to \infty$, the hijacking plus residual fusion recipe should transfer to diffusion-style samplers, not just Poisson flow models.
  • The 16-step, 2-hijack-step sampling budget makes generative reconstruction practical for clinical timeframes, where the compared diffusion baselines need more steps or much longer runtimes.
  • The ASD-POCS data-consistency branch is essential for suppressing hallucinated structures under extreme undersampling, so the method is best understood as a hybrid of a learned generative prior and TV-regularized physical correction.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The paper tunes $\alpha$ and the hijack start on simulated data and validates on a single clinical patient, so a reader should expect these hyperparameters to need re-tuning for other scanners, geometries, or dose levels.
  • A direct test of the trajectory-continuity claim would be to compare, at each sampling step, the model-predicted Poisson-field direction with the actual fused update direction; if fusion preserves the trajectory, the angle between them should stay small throughout.
  • Because the residual fusion effectively blends a generative prior with a TV-regularized reconstruction, the method's edge preservation may owe much to the ASD-POCS branch; ablating the generative branch entirely would clarify how much the learned prior contributes.
  • The same scheme—a fast ODE generative prior plus projection-domain consistency via residual fusion—should transfer to other linear inverse problems such as MRI or PET, where a forward operator and undersampling mask are available.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

6 major / 4 minor

Summary. The paper proposes ResPF, a conditional Poisson Flow Generative Model (based on PFGM++) for sparse-view CT reconstruction. The method uses an FBP reconstruction as a conditioning input, a 'hijacking' strategy to start sampling from an intermediate timestep, ASD-POCS data consistency updates, and a linear residual fusion between the generative output and the physically refined image. Experiments on simulated AAPM datasets (63 and 125 views) and one GE clinical dataset (123 views) report state-of-the-art PSNR, SSIM, and LPIPS with an inference time of about 1.5 s per slice. The central claims are that ResPF preserves the PFGM ODE trajectory while enforcing data consistency, that it is the first Poisson-flow application to sparse-view CT, and that it outperforms iterative, learning-based, and diffusion baselines.

Significance. If the central claims hold, ResPF would be a practically valuable contribution: it demonstrates a fast generative reconstruction pipeline for sparse-view CT, with the first conditional extension of PFGM++ to this task, a hijacked sampling scheme that reduces the number of ODE steps, and a residual fusion module intended to reconcile generative priors with physical consistency. The manuscript includes a broad baseline comparison, an ablation study, runtime measurements, and a real clinical-data test, all of which are strengths. However, several load-bearing issues remain: an internal contradiction between the claimed trajectory-preserving fusion and the in-loop operations in Algorithm 1, an inconsistency between Eq. (24) and Algorithm 1 in the fusion weighting, an implausible EDM result in Table I, and the absence of error bars on the main quality metrics. These issues need to be resolved before the state-of-the-art and physical-consistency claims can be considered established.

major comments (6)
  1. [Section III-D, Eq. (24), and Algorithm 1] The fusion rule in Eq. (24) is \hat{x}_fused = α·\hat{x}_gen + (1−α)·\hat{x}_phys, but Algorithm 1, line 15 sets \hat{x}_fused ← α·\hat{x}_phys + (1−α)·\hat{x}_gen. These assign opposite weights to the generative and physically consistent terms, and the surrounding text ('When α is large, the result favors the generative output') agrees with Eq. (24), not with the algorithm. This is load-bearing because α=0.4 is used in all experiments and the ablation conclusions depend on which term is weighted. The equation and pseudocode must be reconciled, and the reported results must be re-examined if the implementation used the swapped weighting.
  2. [Section III-D vs. Algorithm 1] The paper argues that residual fusion 'decouples data fidelity from the intermediate steps' and preserves the ODE trajectory, but Algorithm 1 applies ASD-POCS (lines 12–14) and sets x_{i+1} to the fused image (lines 15–16) at every sampling step. This is exactly the type of step-wise data-consistency insertion that Section III-C warns disrupts the continuous generative flow. Either the pseudocode is wrong or the trajectory-preservation claim is unsubstantiated. The authors should clarify whether fusion is intended as an in-loop operation or only as a final output, and support the actual mechanism by ablating in-loop versus final-only fusion.
  3. [Section III-B, Eq. (13), and Algorithm 1, line 2] Eq. (13) defines the hijacked initial state as x_{t0} = x_sp + σ(t0)·ε, while Algorithm 1 sets x_τ ← x_sp with no noise perturbation. These are different initializations: the algorithm starts from the clean FBP image, not from a noisy intermediate sample. Please state which implementation was used and, if line 2 is correct, reconcile the description of starting from an intermediate timestep with the noise-free initialization.
  4. [Table I] For EDM on the simulation dataset, PSNR is 38.5046 dB at 63 views but only 31.1091 dB at 125 views. This 7.4 dB decrease with more projection views is physically implausible and suggests an experimental or reporting error (e.g., a misconfigured baseline, data leakage, or swapped cells). Since the cross-sparsity comparison underpins the robustness claim, the authors must correct or explain this result before the state-of-the-art claim can be accepted.
  5. [Table I and Section IV-B] All PSNR, SSIM, and LPIPS results are reported as point estimates without standard deviations or significance tests, despite the simulation datasets comprising hundreds of slices. Because many baseline differences in Table I are small (e.g., SSIM differences below 0.01), the claim that ResPF 'consistently' outperforms all baselines is not statistically supported. Provide error bars or per-slice distributions, and state the number of test images over which the metrics are averaged.
  6. [Section IV-A and Section III-D] Key hyperparameters (fusion coefficient α=0.4, hijack start step, number of sampling steps, D=128, and ASD-POCS parameters) are selected on the simulation dataset, and the same simulation dataset is used for the headline results in Table I. The only external evidence is one clinical patient, so the robustness and transferability claims are currently supported by limited independent data. Please clarify whether any held-out simulation slices were used for hyperparameter selection, report validation metrics on those, or temper the generalization claims accordingly.
minor comments (4)
  1. [Section IV-A] In the Implementation Details paragraph, '2 hijcak steps' should read '2 hijack steps'.
  2. [Section III-B] The sentence 'Relative EDM sampler is shown in Fig. 3(B)' appears to be a typo; it should likely read 'The EDM sampler is shown in Fig. 3(B)'.
  3. [Section IV-C] The NPS maps are presented qualitatively, but no quantitative NPS summary (e.g., band-limited variance or a roughness measure) is reported; a brief quantitative statement would strengthen the claim of 'natural texture recovery'.
  4. [Algorithm 1, line 4] The update direction d_i is defined as (x_i − f_θ(x_i, t_i, x_sp))/t_i, while the text in Eqs. (14)–(17) uses the model-predicted Poisson field direction \hat{E}_θ; the relationship between these two expressions should be stated explicitly.

Circularity Check

1 steps flagged · score 4.0 of 10

Partial circularity: the fusion weight is tuned on the simulation benchmark, and the same simulation benchmark is then reported as a headline superiority result.

  1. fitted input called prediction [Section III-D (Eq. 24 and grid-search paragraph); Section IV-B (Table I simulation results)]
    "Through a grid search over α∈[0,1] conducted on the simulation dataset, we find that α=0.4 provides an effective trade-off between perceptual quality and data fidelity in sparse-view CT reconstruction. ... Quantitative evaluations further substantiate these findings, as summarized in Table I. In the case of severely undersampled 63-view, the ResPF framework consistently achieves the best performance across all metrics, demonstrating significant advantages in both structural fidelity and perceptual quality."

    The fusion weight α in Eq. (24) is the free parameter that balances the generative output against the ASD-POCS-corrected image. It is selected by grid search on the simulation dataset using the same PSNR/SSIM/LPIPS criteria that are later reported in Table I as evidence that ResPF achieves superior reconstruction quality on that dataset. The simulation-column numbers are therefore in-sample fits for α, not out-of-sample predictions. The clinical results are partly independent, but they cover a single patient and do not remove the circularity from the simulation headline.

full rationale

The only clear circular step is the selection of the fusion coefficient α on the simulation dataset followed by reporting simulation metrics as evidence of superiority. This is a genuine fitted-input-called-prediction issue, though it affects one scalar parameter rather than the whole architecture. The rest of the method is an empirical application of external PFGM++/EDM machinery: cPFGM training follows the standard PFGM++/EDM objectives, the hijack step is a heuristic initialization, and ASD-POCS is a standard iterative data-consistency module. No load-bearing uniqueness theorem or ansatz is imported through self-citation; references [32] and [34] are external and independently established. A separate coherence risk, not counted as circularity, is that Algorithm 1 performs ASD-POCS and fused-state replacement inside every sampling step, whereas Section III-D claims the fusion 'decouples data fidelity from the intermediate steps' and preserves ODE trajectory continuity; this is an internal contradiction or an unsupported claim, but it is not a reduction of an output to an input by construction. Overall, the central contribution is not equivalent to its inputs, but the headline simulation comparison is partly circular because its fusion weight was tuned on the same benchmark.

Assumptions & free parameters 5 free parameters · 4 assumptions · 0 invented entities

The central claim rests on PFGM++'s ODE as a prior, on FBP images as sufficient conditioning, and on an unproven trajectory-preservation property of the residual fusion step. The method's free parameters (alpha, D, hijack time, step count, ASD-POCS settings) are tuned on the simulation dataset.

free parameters (5)
  • Fusion coefficient alpha = 0.4
    Chosen by grid search on the simulation dataset (Sec. III-D, Fig. 4). It weights the linear combination of generative output and ASD-POCS output.
  • Hijack start timestep t0 = step 14 (within 20-40% of the noise schedule)
    Empirically selected to balance quality and speed (Sec. III-B, ablation in Sec. IV-E).
  • Number of sampling steps = 16
    Selected from {8,16,32} as a trade-off (Sec. IV-A3).
  • Augmented dimension D = 128
    Selected from {64,128,infinity} for numerical stability (Sec. IV-A3, Fig. 5).
  • ASD-POCS iterations, subsets, TV steps = 10 iterations, 8 subsets, 5 TV steps
    Hand-chosen convergence parameters (Sec. IV-A3).
assumptions (4)
  • domain assumption PFGM++ defines a valid generative ODE (Eq. 7) that maps noise to the data distribution.
    The method builds on the PFGM++ framework [34] without re-deriving it; correctness of the ODE is assumed from prior literature.
  • domain assumption The FBP image x_sp provides sufficient conditioning information for the generative model to converge to the true posterior.
    The authors state this is 'widely adopted in traditional CT reconstruction literature' (Sec. III-A) but offer no quantitative support.
  • ad hoc to paper Linearly mixing the generative output and the ASD-POCS-refined image at every step preserves the ODE trajectory and the generative prior.
    This is the core methodological claim; no proof or dedicated ablation is given, and it conflicts with Algorithm 1's in-loop fusion.
  • domain assumption ASD-POCS with 10 iterations and 8 subsets provides sufficient data consistency.
    Borrowed from iterative CT literature; the specific settings are chosen by hand (Sec. IV-A3).

how reviews work

0 comments
Cite this review

Pith. "Pith review of ResPF: Residual Poisson Flow for Efficient and Physically Consistent Sparse-View CT Reconstruction." pith.science (2026). https://pith.science/paper/IYKWAI5R

@misc{pith2026250606400,
  author       = {Pith},
  title        = {Pith review of: ResPF: Residual Poisson Flow for Efficient and Physically Consistent Sparse-View CT Reconstruction},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/IYKWAI5R}},
  note         = {Machine review of arXiv:2506.06400}
}
read the original abstract

Sparse-view computed tomography (CT) is a practical solution to reduce radiation dose, but the resulting ill-posed inverse problem poses significant challenges for accurate image reconstruction. Although deep learning and diffusion-based methods have shown promising results, they often lack physical interpretability or suffer from high computational costs due to iterative sampling starting from random noise. Recent advances in generative modeling, particularly Poisson Flow Generative Models (PFGM), enable high-fidelity image synthesis by modeling the full data distribution. In this work, we propose Residual Poisson Flow (ResPF) Generative Models for efficient and accurate sparse-view CT reconstruction. Based on PFGM++, ResPF integrates conditional guidance from sparse measurements and employs a hijacking strategy to significantly reduce sampling cost by skipping redundant initial steps. However, skipping early stages can degrade reconstruction quality and introduce unrealistic structures. To address this, we embed a data-consistency into each iteration, ensuring fidelity to sparse-view measurements. Yet, PFGM sampling relies on a fixed ordinary differential equation (ODE) trajectory induced by electrostatic fields, which can be disrupted by step-wise data consistency, resulting in unstable or degraded reconstructions. Inspired by ResNet, we introduce a residual fusion module to linearly combine generative outputs with data-consistent reconstructions, effectively preserving trajectory continuity. To the best of our knowledge, this is the first application of Poisson flow models to sparse-view CT. Extensive experiments on synthetic and clinical datasets demonstrate that ResPF achieves superior reconstruction quality, faster inference, and stronger robustness compared to state-of-the-art iterative, learning-based, and diffusion models.

Figures

Figures reproduced from arXiv: 2506.06400 by the authors.

Figure 1
Figure 1. Modeling of sparse-view CT data acquisition. [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Illustration of 3D Poisson field trajectories for a 2D CT image [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Sampling process of the proposed Residual Poisson Flow (ResPF) model. (A) A sparse-view reconstruction [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figures from the paper (6 more)
Figure 4
Figure 4. Figure 4: Effect of fusion coefficient α on SSIM and PSNR. IV. EXPERIMENTS AND RESULTS A. Experimental Setup To comprehensively evaluate the performance of the pro￾posed ResPF in sparse-view CT reconstruction, experiments are conducted on two simulated datasets and one clinical …
Figure 5
Figure 5. Figure 5: Effect of D for reconstruction under different sparse-view projections. • EDM [32]: EDM (Elucidated Diffusion Model) includes a family of diffusion models, and Poisson flow generative models are based on it, which offers fast sampling via denoising optimization. As the…
Figure 6
Figure 6. Figure 6: Visual comparison of reconstruction results on two slices from the AAPM low-dose CT simulation dataset under a 125-view setting. The first and [PITH_FULL_IMAGE:figures/full_fig_p009_6.png]
Figure 7
Figure 7. Figure 7: Same as Fig.6 but reconstructed from highly sparse 63-view setting. [PITH_FULL_IMAGE:figures/full_fig_p010_7.png]
Figure 8
Figure 8. Figure 8: Qualitative and frequency-domain comparison of reconstruction methods on clinical CT dataset under a 123-view sparse-view setting. The first row [PITH_FULL_IMAGE:figures/full_fig_p011_8.png]
Figure 9
Figure 9. Figure 9: Ablation study results under 125-view sparse setting. Each column [PITH_FULL_IMAGE:figures/full_fig_p011_9.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

45 extracted references · 31 canonical work pages

  1. [1]

    Representation of a function by its line integrals, with some radiological applications,

    A. M. Cormack, “Representation of a function by its line integrals, with some radiological applications,”Journal of applied physics, vol. 34, no. 9, pp. 2722–2727, 1963

  2. [2]

    Computerized transverse axial scanning (tomogra- phy): Part 1. description of system,

    G. N. Hounsfield, “Computerized transverse axial scanning (tomogra- phy): Part 1. description of system,”The British journal of radiology, vol. 46, no. 552, pp. 1016–1022, 1973

  3. [3]

    An outlook on x-ray ct research and development,

    G. Wang, H. Yu, and B. De Man, “An outlook on x-ray ct research and development,”Medical physics, vol. 35, no. 3, pp. 1051–1064, 2008

  4. [4]

    Low-dose ct with a residual encoder-decoder convolutional neural network,

    H. Chen, Y . Zhang, M. K. Kalra, F. Lin, Y . Chen, P. Liao, J. Zhou, and G. Wang, “Low-dose ct with a residual encoder-decoder convolutional neural network,”IEEE transactions on medical imaging, vol. 36, no. 12, pp. 2524–2535, 2017

  5. [5]

    Computed tomography—an increasing source of radiation exposure,

    D. J. Brenner and E. J. Hall, “Computed tomography—an increasing source of radiation exposure,”New England journal of medicine, vol. 357, no. 22, pp. 2277–2284, 2007

  6. [6]

    Strategies for ct radiation dose optimization,

    K. Mk, “Strategies for ct radiation dose optimization,”Radiology, vol. 230, no. 3, pp. 619–628, 2004

  7. [7]

    The alara concept in pediatric ct: myth or reality?

    T. L. Slovis, “The alara concept in pediatric ct: myth or reality?” Radiology, vol. 223, no. 1, pp. 5–6, 2002

  8. [8]

    Low-dose x-ray ct reconstruction via dictionary learning,

    Q. Xu, H. Yu, X. Mou, L. Zhang, J. Hsieh, and G. Wang, “Low-dose x-ray ct reconstruction via dictionary learning,”IEEE transactions on medical imaging, vol. 31, no. 9, pp. 1682–1697, 2012

Show all 45 references
  1. [9]

    Artifact suppressed dictionary learning for low-dose ct image processing,

    Y . Chen, L. Shi, Q. Feng, J. Yang, H. Shu, L. Luo, J.-L. Coatrieux, and W. Chen, “Artifact suppressed dictionary learning for low-dose ct image processing,”IEEE transactions on medical imaging, vol. 33, no. 12, pp. 2271–2292, 2014

  2. [10]

    Accurate image reconstruction from few-views and limited-angle data in divergent-beam ct,

    E. Y . Sidky, C.-M. Kao, and X. Pan, “Accurate image reconstruction from few-views and limited-angle data in divergent-beam ct,”Journal of X-ray Science and Technology, vol. 14, no. 2, pp. 119–139, 2006

  3. [11]

    Prior image constrained compressed sensing (piccs): a method to accurately reconstruct dynamic ct im- ages 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 im- ages from highly undersampled projection data sets,”Medical physics, vol. 35, no. 2, pp. 660–663, 2008

  4. [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,” Physics in Medicine & Biology, vol. 53, no. 17, p. 4777, 2008

  5. [13]

    Evaluation of sparse-view reconstruction from flat-panel-detector cone-beam ct,

    J. Bian, J. H. Siewerdsen, X. Han, E. Y . Sidky, J. L. Prince, C. A. Pelizzari, and X. Pan, “Evaluation of sparse-view reconstruction from flat-panel-detector cone-beam ct,”Physics in Medicine & Biology, vol. 55, no. 22, p. 6575, 2010

  6. [14]

    The ill-conditioned nature of the limited angle tomog- raphy problem,

    M. E. Davison, “The ill-conditioned nature of the limited angle tomog- raphy problem,”SIAM Journal on Applied Mathematics, vol. 43, no. 2, pp. 428–448, 1983

  7. [15]

    Dictionary learning based sinogram inpainting for ct sparse reconstruction,

    S. Li, Q. Cao, Y . Chen, Y . Hu, L. Luo, and C. Toumoulin, “Dictionary learning based sinogram inpainting for ct sparse reconstruction,”Optik, vol. 125, no. 12, pp. 2862–2867, 2014

  8. [16]

    K-svd: An algorithm for designing overcomplete dictionaries for sparse representation,

    M. Aharon, M. Elad, and A. Bruckstein, “K-svd: An algorithm for designing overcomplete dictionaries for sparse representation,”IEEE Transactions on signal processing, vol. 54, no. 11, pp. 4311–4322, 2006

  9. [17]

    View-interpolation of sparsely sampled sinogram using convolutional neural network,

    H. Lee, J. Lee, and S. Cho, “View-interpolation of sparsely sampled sinogram using convolutional neural network,” inMedical Imaging 2017: Image Processing, vol. 10133. SPIE, 2017, pp. 617–624

  10. [18]

    Imagenet classification with deep convolutional neural networks,

    A. Krizhevsky, I. Sutskever, and G. E. Hinton, “Imagenet classification with deep convolutional neural networks,”Advances in neural informa- tion processing systems, vol. 25, 2012

  11. [19]

    Cine cone beam ct reconstruction using low-rank matrix factorization: algorithm and a proof-of-principle study,

    J.-F. Cai, X. Jia, H. Gao, S. B. Jiang, Z. Shen, and H. Zhao, “Cine cone beam ct reconstruction using low-rank matrix factorization: algorithm and a proof-of-principle study,”IEEE transactions on medical imaging, vol. 33, no. 8, pp. 1581–1591, 2014

  12. [20]

    Non- local total-variation (nltv) minimization combined with reweighted l1- norm for compressed sensing ct reconstruction,

    H. Kim, J. Chen, A. Wang, C. Chuang, M. Held, and J. Pouliot, “Non- local total-variation (nltv) minimization combined with reweighted l1- norm for compressed sensing ct reconstruction,”Physics in Medicine & Biology, vol. 61, no. 18, p. 6878, 2016

  13. [21]

    Sparse-view x-ray ct reconstruction via total generalized variation regularization,

    S. Niu, Y . Gao, Z. Bian, J. Huang, W. Chen, G. Yu, Z. Liang, and J. Ma, “Sparse-view x-ray ct reconstruction via total generalized variation regularization,”Physics in Medicine & Biology, vol. 59, no. 12, p. 2997, 2014

  14. [22]

    Relaxed linearized algorithms for faster x- ray ct image reconstruction,

    H. Nien and J. A. Fessler, “Relaxed linearized algorithms for faster x- ray ct image reconstruction,”IEEE transactions on medical imaging, vol. 35, no. 4, pp. 1090–1098, 2015

  15. [23]

    Combining ordered subsets and momentum for accelerated x-ray ct image reconstruction,

    D. Kim, S. Ramani, and J. A. Fessler, “Combining ordered subsets and momentum for accelerated x-ray ct image reconstruction,”IEEE transactions on medical imaging, vol. 34, no. 1, pp. 167–178, 2014

  16. [24]

    Sparse-view x-ray spectral ct reconstruction using annihilating filter-based low rank hankel matrix approach,

    Y . S. Han, K. H. Jin, K. Kim, and J. C. Ye, “Sparse-view x-ray spectral ct reconstruction using annihilating filter-based low rank hankel matrix approach,” in2016 IEEE 13th International Symposium on Biomedical Imaging (ISBI). IEEE, 2016, pp. 573–576

  17. [25]

    ρ-nerf: Leveraging attenuation priors in neural radiance field for 3d computed tomography reconstruction,

    L. Zhou, C. Fang, B. Morovati, Y . Liu, S. Han, Y . Xu, and H. Yu, “ρ-nerf: Leveraging attenuation priors in neural radiance field for 3d computed tomography reconstruction,” 2024. [Online]. Available: https://arxiv.org/abs/2412.05322

  18. [26]

    Diffusion probabilistic priors for zero-shot low-dose ct image denoising,

    X. Liu, Y . Xie, J. Cheng, S. Diao, S. Tan, and X. Liang, “Diffusion probabilistic priors for zero-shot low-dose ct image denoising,”arXiv preprint arXiv:2305.15887, 2023

  19. [27]

    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,”arXiv preprint arXiv:2111.08005, 2021

  20. [28]

    Mr image denoising and super- resolution using regularized reverse diffusion,

    H. Chung, E. S. Lee, and J. C. Ye, “Mr image denoising and super- resolution using regularized reverse diffusion,”IEEE Transactions on Medical Imaging, vol. 42, no. 4, pp. 922–934, 2022

  21. [29]

    Jccs-pfgm: A novel circle- supervision based poisson flow generative model for multiphase cect progressive low-dose reconstruction with joint condition,

    R. Ge, Y . He, C. Xia, and D. Zhang, “Jccs-pfgm: A novel circle- supervision based poisson flow generative model for multiphase cect progressive low-dose reconstruction with joint condition,” inInterna- tional Conference on Medical Image Computing and Computer-Assisted Interve...

  22. [30]

    Physics-informed score-based diffusion model for limited-angle reconstruction of cardiac computed tomography,

    S. Han, Y . Xu, D. Wang, B. Morovati, L. Zhou, J. S. Maltz, G. Wang, and H. Yu, “Physics-informed score-based diffusion model for limited-angle reconstruction of cardiac computed tomography,”IEEE Transactions on Medical Imaging, 2024

  23. [31]

    Im- pact of network architecture and training strategy for photon counting ct data correction,

    B. Morovati, M. Li, S. Han, Y . Xu, L. Zhou, G. Wang, and H. Yu, “Im- pact of network architecture and training strategy for photon counting ct data correction,” inDevelopments in X-Ray Tomography XV, vol. 13152. SPIE, 2024, p. 131520D

  24. [32]

    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,”Advances in neural infor- mation processing systems, vol. 35, pp. 26 565–26 577, 2022

  25. [33]

    Poisson flow generative models,

    Y . Xu, Z. Liu, M. Tegmark, and T. Jaakkola, “Poisson flow generative models,”Advances in Neural Information Processing Systems, vol. 35, pp. 16 782–16 795, 2022

  26. [34]

    Pfgm++: Unlocking the potential of physics-inspired generative models,

    Y . Xu, Z. Liu, Y . Tian, S. Tong, M. Tegmark, and T. Jaakkola, “Pfgm++: Unlocking the potential of physics-inspired generative models,” in International Conference on Machine Learning. PMLR, 2023, pp. 38 566–38 591

  27. [35]

    U. M. Ascher and L. R. Petzold,Computer methods for ordinary differential equations and differential-algebraic equations. SIAM, 1998

  28. [36]

    Framing u-net via deep convolutional framelets: Application to sparse-view ct,

    Y . Han and J. C. Ye, “Framing u-net via deep convolutional framelets: Application to sparse-view ct,”IEEE transactions on medical imaging, vol. 37, no. 6, pp. 1418–1429, 2018

  29. [37]

    Sgd- net: Efficient model-based deep learning with theoretical guarantees,

    J. Liu, Y . Sun, W. Gan, X. Xu, B. Wohlberg, and U. S. Kamilov, “Sgd- net: Efficient model-based deep learning with theoretical guarantees,” IEEE Transactions on Computational Imaging, vol. 7, pp. 598–610, 2021

  30. [38]

    Deep con- volutional neural network for inverse problems in imaging,

    K. H. Jin, M. T. McCann, E. Froustey, and M. Unser, “Deep con- volutional neural network for inverse problems in imaging,”IEEE transactions on image processing, vol. 26, no. 9, pp. 4509–4522, 2017

  31. [39]

    Low Dose CT Grand Challenge,

    AAPM, “Low Dose CT Grand Challenge,” https://www.aapm.org/ grandchallenge/lowdosect/, 2017, [Online; accessed 2024-12-28]

  32. [40]

    Fast calculation of the exact radiological path for a three- dimensional ct array,

    R. L. Siddon, “Fast calculation of the exact radiological path for a three- dimensional ct array,”Medical physics, vol. 12, no. 2, pp. 252–255, 1985

  33. [41]

    Swinir: Image restoration using swin transformer,

    J. Liang, J. Cao, G. Sun, K. Zhang, L. Van Gool, and R. Timofte, “Swinir: Image restoration using swin transformer,” inProceedings of the IEEE/CVF international conference on computer vision, 2021, pp. 1833–1844

  34. [42]

    Decomposed diffusion sam- pler for accelerating large-scale inverse problems,

    H. Chung, S. Lee, and J. C. Ye, “Decomposed diffusion sam- pler for accelerating large-scale inverse problems,”arXiv preprint arXiv:2303.05754, 2023

  35. [43]

    Ppfm: Image denoising in photon-counting ct using single-step posterior sampling poisson flow generative models,

    D. Hein, S. Holmin, T. Szczykutowicz, J. S. Maltz, M. Danielsson, G. Wang, and M. Persson, “Ppfm: Image denoising in photon-counting ct using single-step posterior sampling poisson flow generative models,” IEEE Transactions on Radiation and Plasma Medical Sciences, 2024

  36. [44]

    Optimal design of fir filter based on improved artificial bee colony algorithm,

    C. Fang, X. Huang, and K. Xia, “Optimal design of fir filter based on improved artificial bee colony algorithm,” in2022 21st International Symposium on Distributed Computing and Applications for Business Engineering and Science (DCABES). IEEE, 2022, pp. 155–158

  37. [45]

    A hybrid equilibrium optimizer algorithm and its optimization for lstm,

    L. Li, Y . Yang, C. Fang, and K. Xia, “A hybrid equilibrium optimizer algorithm and its optimization for lstm,” in2023 16th International Conference on Advanced Computer Theory and Engineering (ICACTE). IEEE, 2023, pp. 213–217

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.