Pith. sign in

REVIEW 5 major objections 5 minor 42 references

PWD: Prior-Guided and Wavelet-Enhanced Diffusion Model for Limited-Angle CT

T0 review · 5 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

Pith's one-line read A diffusion model that conditions every sampling step on a cheap filtered back-projection prior can reconstruct limited-angle dental CT in 50 steps, reporting higher PSNR and SSIM than diffusion baselines that use 200 or 1000 steps.

desk verdict A promising fast-sampling recipe for limited-angle CT that is undermined by an overstated headline claim and shaky derivations; the experiments show real gains but not the advertised 10% SSIM. read the letter →

arxiv 2507.05317 v2 pith:7OLQC5EL submitted 2025-06-30 eess.IV cs.AIcs.CV

classification eess.IVcs.AIcs.CV
keywords limited-angleCTdiffusionmodelfastsamplingguidedwaveletfeaturefusiondentalCBCTimagereconstruction
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 argues that limited-angle CT reconstruction can be made both fast and faithful by conditioning a diffusion model on a cheap structural prior: the filtered back-projection (FBP) image computed from the available projections. During training the denoiser learns to map noisy limited-angle images to full-angle targets while seeing that FBP prior; during inference, each denoising step pulls the predicted clean image a small distance toward the prior instead of relying on the model alone. A wavelet-convolution branch adds multi-scale frequency features so that skipped sampling steps do not blur edges. The authors report that at 50 sampling steps the method improves PSNR by at least 1.7 dB and SSIM by at least 10% over existing diffusion and supervised baselines on clinical dental-arch and periapical CT, with per-slice inference around four seconds. If correct, this makes high-quality limited-angle dental CT practical in clinical settings where many-step diffusion was previously too slow.

What carries the argument

The load-bearing mechanism is the guided-DDIM interpolation formula: the predicted clean image x*_0 is pulled toward the FBP prior c by a scalar weight w, which is presented as the closed-form minimizer of ||x0 - x*_0||^2 + w||x0 - c||^2. This O(1) guidance replaces projection-domain data-consistency operations, and its validity rests on the linearity of FBP, which keeps the limited-angle-plus-missing-angle combination structurally consistent with full-angle reconstruction. The second component is the WTConv module: a discrete wavelet transform splits features into LL, LH, HL, and HH bands, each processed by depthwise separable convolution and recombined by inverse transform, then fused with a standard convolution branch so high-frequency detail survives skip sampling.

What would settle it

Run PWD at 50 sampling steps on a limited-angle test set where the FBP prior is known to be poor, such as scans with metal implants or angular coverage reduced to 60 degrees: if the method no longer beats an unguided DDIM baseline at the same step count, the claimed advantage depends on favorable prior consistency rather than on the sampling mechanism itself. A simpler check is a guidance-weight sweep on such data: an optimal weight of zero would indicate that the prior term is not doing the work.

Watch

Extended reading notes

Core claim

The central claim is that a diffusion model can preserve fine structure under aggressive step reduction if each sampling step is explicitly guided by a prior image rather than relying on learned statistics alone. PWD embeds the limited-angle CT image c as a condition in the denoising network and, at inference, replaces the DDIM (denoising diffusion implicit model) clean-image estimate x*_0 with x_g = x*_0 + w(c - x*_0), the minimizer of a quadratic energy that balances model prediction against the prior. A wavelet feature fusion branch (WTConv) processes low- and high-frequency subbands separately so edges survive 50-step sampling. On clinical dental-arch and periapical datasets at 90 degrees and 120 degrees scan ranges, the authors report PSNR and SSIM above all compared methods under the same or fewer sampling steps: at 50 steps, at least 1.7 dB PSNR and 10% SSIM improvement over the best baselines, with reconstruction time around 3.9 seconds per slice.

Load-bearing premise

The load-bearing premise is that the FBP image computed from the limited-angle projections alone is structurally consistent with the unknown full-angle image, so that pulling every sampling step toward that prior improves rather than corrupts the result; the linearity reasoning for that consistency is sketched rather than proven, and the Gaussian form of the prior on x0 given c is asserted without validation.

Editorial extensions

If this is right

  • At 50 sampling steps, PWD reaches higher PSNR and SSIM than DOLCE at 200 steps and DPS at 1000 steps, so the step budget no longer needs to be sacrificed for quality.
  • The guidance term is a constant-time operation, so per-slice inference drops to about 3.9 seconds at 50 steps on a single GPU, moving diffusion reconstruction toward interactive clinical use.
  • Wavelet feature fusion contributes most when the step budget is small and saturates near full sampling, meaning the frequency branch is specifically a remedy for skip-sampling degradation.
  • With w = 0.05 optimal and w = 1 harmful, the prior must be applied weakly; overly strong guidance suppresses the generative model and lowers both PSNR and SSIM.

Reading between the lines

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

  • If correct, the same O(1) prior-guide recipe should transfer to other ill-posed inverse problems with cheap linear reconstructions, such as sparse-view CT, MRI with analytic backprojection priors, or PET; the test would be whether a weak prior anchor improves DDIM at low step counts.
  • The paper's Gaussian prior on x0 given c is asserted rather than validated; a harder test is whether calibrating the guidance weight to the local reliability of the FBP prior improves robustness at very narrow angles or with metal implants.
  • Because gains are measured on paired full-angle references from the same scanner, a natural extension is to test on out-of-distribution anatomy or scan geometry; the structural-consistency argument predicts graceful degradation as long as FBP retains coarse directional structure.
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

5 major / 5 minor

Summary. The manuscript proposes PWD, a diffusion-model method for limited-angle CT (LACT) reconstruction. During training, a conditional denoising network is trained with the LACT filtered back-projection (FBP) image as a structural condition, and the network is enhanced with wavelet-domain multi-scale feature fusion. During inference, the authors propose a guided DDIM sampling update in which the predicted clean image is interpolated with the LACT prior image, controlled by a scalar weight w, together with a skip-sampling schedule that reduces the inference to 50 steps. Experiments on dental-arch and periapical CBCT datasets compare PWD with FBP, FBPConvNet, IRON, DPS, DDIM, DOLCE, and TIFA, reporting improved PSNR/SSIM and lower runtime, with an advertised guarantee of 'at least 1.7 dB improvement in PSNR and 10% gain in SSIM' at 50 sampling steps.

Significance. If substantiated, the work would be practically valuable: it targets a real deployment scenario (fast limited-angle dental CBCT on a commercial scanner) and proposes a simple mechanism to retain fidelity when the diffusion sampling budget is cut to 50 steps. The use of a real clinical scanner, multiple baseline comparisons, and dedicated ablation studies are positive aspects. The principal claims, however, are not supported by the evidence in the manuscript: the quantitative guarantee in the abstract is contradicted by the paper's own tables, and the theoretical derivation of the guided update contains multiple load-bearing errors. The paper does not provide code or data, so the empirical results are not independently verifiable.

major comments (5)
  1. [Abstract; Tables I-III] The headline claim 'Using only 50 sampling steps, PWD achieves at least 1.7 dB improvement in PSNR and 10% gain in SSIM' is not supported by the paper's own tables. In Table II (periapical dataset), the gain over the best baseline DOLCE is +1.25 dB / +0.29% SSIM at 90 degrees and +0.33 dB / +1.50% SSIM at 120 degrees. In Table III, the matched 50-step dental-arch comparison gives +2.99 dB but only +4.37% relative SSIM over DOLCE. Table I reports a 10-slice subset without error bars, and even there the SSIM gain is +7.68 percentage points at 90 degrees and +2.04 percentage points at 120 degrees, not 10%. As written, the abstract's 'at least' statement is false against the best baseline in several configurations.
  2. [Section III-B, Eqs. (6)-(8)] The score-function scaling in Eqs. (6) and (7) is incorrect. For the forward process in Eq. (4), the score is approximately -epsilon_theta(x_t,t)/sqrt(1-alpha_bar_t), not -epsilon_theta(x_t,t)/(1-alpha_bar_t). The error appears in both the unconditional and conditional scores, so Eq. (8) is accidentally correct after cancellation, but the derivation as written is invalid and the claimed relationship between conditional and unconditional noise predictions is not derived from the stated equations.
  3. [Section III-C, Eqs. (17)-(21)] Eq. (19) is not the maximizer of the posterior in Eq. (17)-(18). With p(xt|x0) = N(sqrt(alpha_bar_t) x0, (1-alpha_bar_t)I) and p(x0|c) = N(c, sigma_c^2 I), the posterior maximizer is x_hat = (sqrt(alpha_bar_t)/(1-alpha_bar_t) x_t + c/sigma_c^2) / (alpha_bar_t/(1-alpha_bar_t) + 1/sigma_c^2), which is a timestep-dependent weighted average, not x*_0 + w(c - x*_0). Moreover, the quadratic energy in Eq. (21) has minimizer (x*_0 + w c)/(1+w), which equals Eq. (19) only for w=0. Since w is then tuned to 0.05 in the ablation study, the claimed closed-form Bayesian derivation is not load-bearing for the actual update.
  4. [Section III-C, Eq. (16) and Algorithm 1] The DDIM update in Eq. (16) uses the coefficient alpha_{t-1}/alpha_t in front of the predicted-clean-image term; the standard DDIM update requires sqrt(alpha_{t-1}/alpha_t). The same inconsistency appears in Eq. (20) and in Algorithm 1's reconstruction loop. If the implementation uses the correct DDIM coefficient, the paper should state the exact update actually implemented; if it uses the equations as written, the sampling scheme is not DDIM and its properties do not follow from the cited DDIM derivation.
  5. [Section III-A] The linearity argument justifying the prior c is not applicable to the actual input. The authors argue that FBP from the sum of available and missing projections is structurally consistent with the full-angle reconstruction, but c is computed from the limited-angle projections only; the missing projections are not available. The assumption that the FBP-limited-angle image is a reliable structural prior for the full-angle target is therefore unvalidated, and this is especially questionable at 90 degrees where the FBP baseline has severe artifacts (e.g., PSNR 22.56 dB in Table I). The small tuned weight w=0.05 may mitigate this, but the paper's motivation in Section III-A overstates the theoretical support for the prior.
minor comments (5)
  1. [Algorithm 1] Line 3 of the training stage, 'epsilon_theta(xt,t) = epsilon_theta(xt,c,t)', is not a meaningful training update; it appears to be a typo for the conditioning of the network. Line 5 of the reconstruction stage samples 'epsilon_theta ~ N(0,I)', which conflates the network prediction with the noise sample.
  2. [Section III-B] The architecture used to inject the prior c into the denoising network is not specified. The reader is not told whether c is concatenated with xt, used as a cross-attention condition, or processed by a separate encoder; this is essential for reproducibility.
  3. [Table I caption] Reporting results on '10 randomly selected slices' without standard deviations or statistical tests is insufficient to support the paper's global quantitative claims; the full test sets (4,000 and 8,000 slices) should be used and error bars reported.
  4. [Section IV-A, periapical dataset] The description says 90-degree data were retained from real CBCT acquisitions but full-view CT images were reconstructed using FBP; please clarify whether the full-angle reference is real or simulated, since this affects the interpretation of the paired training data.
  5. [Section III-B, Eq. (10)] Eq. (10) uses a proportionality sign without defining the normalization or the schedule of the guidance weight; the relation between the conditioning in Eq. (9) and the explicit guidance term in Eq. (10) should be stated precisely.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the derivation chain is empirical and self-contained, with the prior image, wavelet branch, and held-out evaluation acting as independent components.

full rationale

PWD's derivation is not circular. The structural prior c is an FBP reconstruction of the observed limited-angle sinogram, i.e., an external input; conditioning the denoiser on c is a standard conditional-modeling choice, not a quantity that is itself defined by the target reconstruction. The inference update in Eq. (19) is a weighted interpolation between the DDIM estimate x0* and c, and although the paper's Gaussian-posterior derivation is mathematically loose (the MAP weight should be sigma_0^2/(sigma_0^2+sigma_c^2), not sigma_0^2/sigma_c^2, and Eq. (8) has a missing (1 - alpha_t) factor), a flawed derivation is not the same as a circular one: the update is an explicit algorithmic definition, not a prediction that reduces to the training data. The wavelet branch is a separate architectural component tested by ablation against a no-WTConv variant, and the reported test numbers come from held-out slices; there is no load-bearing self-citation chain or imported uniqueness theorem. The main rigor concerns are correctness issues, not circularity: the abstract's 'at least 1.7 dB and 10% SSIM' overstates the tables (best-baseline SSIM relative gains are roughly 0.3-8.7% depending on table, and Table I is restricted to 10 randomly selected slices without error bars), and the guidance weight w is tuned in an ablation rather than derived from the stated posterior. None of these makes the central claim equivalent to its inputs by construction.

Assumptions & free parameters 1 free parameters · 3 assumptions · 0 invented entities

The central claim rests on one tuned scalar (w), two unvalidated modeling assumptions about the FBP prior, and the standard DDIM/score machinery. No new physical entities are introduced; WTConv is an architectural module.

free parameters (1)
  • guidance weight w = 0.05 (optimal in ablation, Fig. 12)
    Presented as sigma0^2/sigmac^2, but the variance ratio is never specified and w is selected by tuning. It controls how much of the LACT FBP prior is mixed into every DDIM step, and the headline results depend on it.
assumptions (3)
  • domain assumption The LACT FBP reconstruction serves as a reliable structural prior for the full-angle image.
    Invoked in Section III-A: summing available and missing projections retains structural consistency with full-angle reconstruction, but missing projections are not available at inference.
  • ad hoc to paper The prior p(x0|c) is an isotropic Gaussian centered at the LACT FBP image.
    Eq. (18). This assumption is used to derive the guided update; no empirical justification for Gaussianity or for the variance sigma_c^2 is given.
  • standard math The score function is approximated by the model output with a specific scaling.
    Eq. (6) prints the score as -1/(1-bara_t) times the noise estimate, but the standard VP schedule uses -1/sqrt(1-bara_t). Eq. (8) uses the sqrt form, so the printed Eq. (6) is internally inconsistent.

how reviews work

0 comments
Cite this review

Pith. "Pith review of PWD: Prior-Guided and Wavelet-Enhanced Diffusion Model for Limited-Angle CT." pith.science (2026). https://pith.science/paper/7OLQC5EL

@misc{pith2026250705317,
  author       = {Pith},
  title        = {Pith review of: PWD: Prior-Guided and Wavelet-Enhanced Diffusion Model for Limited-Angle CT},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/7OLQC5EL}},
  note         = {Machine review of arXiv:2507.05317}
}
read the original abstract

Generative diffusion models have received increasing attention in medical imaging, particularly in limited-angle computed tomography (LACT). Standard diffusion models achieve high-quality image reconstruction but require a large number of sampling steps during inference, resulting in substantial computational overhead. Although skip-sampling strategies have been proposed to improve efficiency, they often lead to loss of fine structural details. To address this issue, we propose a prior information embedding and wavelet feature fusion fast sampling diffusion model for LACT reconstruction. The PWD enables efficient sampling while preserving reconstruction fidelity in LACT, and effectively mitigates the degradation typically introduced by skip-sampling. Specifically, during the training phase, PWD maps the distribution of LACT images to that of fully sampled target images, enabling the model to learn structural correspondences between them. During inference, the LACT image serves as an explicit prior to guide the sampling trajectory, allowing for high-quality reconstruction with significantly fewer steps. In addition, PWD performs multi-scale feature fusion in the wavelet domain, effectively enhancing the reconstruction of fine details by leveraging both low-frequency and high-frequency information. Quantitative and qualitative evaluations on clinical dental arch CBCT and periapical datasets demonstrate that PWD outperforms existing methods under the same sampling condition. Using only 50 sampling steps, PWD achieves at least 1.7 dB improvement in PSNR and 10% gain in SSIM.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

42 extracted references · 38 canonical work pages

  1. [1]

    Directional- TV algorithm for image reconstruction from limited-angular-range data,

    Z. Zhang, B. Chen, D. Xia, E. Y . Sidky, and X. Pan, “Directional- TV algorithm for image reconstruction from limited-angular-range data,” Med. Image Anal. , vol. 70, p. 102030, 2021

  2. [2]

    Spectral2Spectral: Image-spectral similarity assisted deep spectral CT reconstruction without reference,

    X. Guo and et al., “Spectral2Spectral: Image-spectral similarity assisted deep spectral CT reconstruction without reference,” IEEE Trans. Com- put. Imaging , vol. 9, pp. 1031–1042, 2023

  3. [3]

    Deep learning computed tomography: Learning projection-domain weights from image domain in limited angle prob- lems,

    T. W ¨urfl and et al., “Deep learning computed tomography: Learning projection-domain weights from image domain in limited angle prob- lems,” IEEE Trans. Med. Imaging , vol. 37, no. 6, pp. 1454–1463, 2018

  4. [4]

    DIOR: Deep iterative optimization-based residual-learning for limited-angle CT reconstruc- tion,

    D. Hu, Y . Zhang, J. Liu, S. Luo, and Y . Chen, “DIOR: Deep iterative optimization-based residual-learning for limited-angle CT reconstruc- tion,” IEEE Trans. Med. Imaging , vol. 41, no. 7, pp. 1778–1790, 2022

  5. [5]

    Image quality and localization accuracy in c-arm tomosynthesis-guided head and neck surgery,

    G. Bachar, J. H. Siewerdsen, M. J. Daly, D. A. Jaffray, and J. C. Irish, “Image quality and localization accuracy in c-arm tomosynthesis-guided head and neck surgery,”Med. Phys., vol. 34, no. 12, pp. 4664–4677, Nov. 2007

  6. [6]

    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, p. 4777, 2008

  7. [7]

    An improved statistical iterative algorithm for sparse- view and limited-angle CT image reconstruction,

    Z. Hu and et al., “An improved statistical iterative algorithm for sparse- view and limited-angle CT image reconstruction,” Sci. Rep., vol. 7, no. 1, p. 10747, 2017

  8. [8]

    An image reconstruction method based on total variation and wavelet tight frame for limited-angle CT,

    X. Luo, W. Yu, and C. Wang, “An image reconstruction method based on total variation and wavelet tight frame for limited-angle CT,” IEEE Access, vol. 6, pp. 1461–1470, 2017

Show all 42 references
  1. [9]

    Limited-angle ct reconstruction with generative adversarial network sinogram inpainting and unsupervised artifact removal,

    E. Xie, P. Ni, R. Zhang, and X. Li, “Limited-angle ct reconstruction with generative adversarial network sinogram inpainting and unsupervised artifact removal,” Appl. Sci. , vol. 12, no. 12, p. 6268, 2022

  2. [10]

    Limited-angle tomography reconstruction via deep end-to-end learning on synthetic data,

    T. Germer, J. Robine, S. Konietzny, S. Harmeling, and T. Uelwer, “Limited-angle tomography reconstruction via deep end-to-end learning on synthetic data,” Applied Mathematics for Modern Challenges , vol. 1, no. 2, pp. 126–142, 2023

  3. [11]

    BM3D frames and variational image deblurring,

    A. Danielyan, V . Katkovnik, and K. Egiazarian, “BM3D frames and variational image deblurring,” IEEE Trans. Image Process. , vol. 21, no. 4, pp. 1715–1728, 2011

  4. [12]

    Image denoising via sparse and redundant representations over learned dictionaries,

    M. Elad and M. Aharon, “Image denoising via sparse and redundant representations over learned dictionaries,” IEEE Trans. Image Process. , vol. 15, no. 12, pp. 3736–3745, 2006

  5. [13]

    Iterative residual optimization network for limited- angle tomographic reconstruction,

    J. Pan and et al., “Iterative residual optimization network for limited- angle tomographic reconstruction,” IEEE Trans. Image Process., vol. 33, pp. 910–925, 2024

  6. [14]

    A survey on generative diffusion models,

    H. Cao and et al., “A survey on generative diffusion models,” IEEE Trans. Knowl. Data Eng. , 2024

  7. [15]

    Efficient diffusion models: A Survey,

    H. e. Shen, “Efficient diffusion models: A Survey,” arXiv:2502.06805, 2025

  8. [16]

    Diffusion model quantization: A Review,

    Q. Zeng and et al., “Diffusion model quantization: A Review,” arXiv:2505.05215, 2025

  9. [17]

    Tackling the generative learning trilemma with denoising diffusion GANs,

    Z. Xiao, K. Kreis, and A. Vahdat, “Tackling the generative learning trilemma with denoising diffusion GANs,” in Proc. Int. Conf. Learn. Represent. (ICLR), 2022

  10. [18]

    Continual learning of diffusion models with generative distillation,

    S. Masip, P. Rodr ´ıguez, T. Tuytelaars, and G. M. van de Ven, “Continual learning of diffusion models with generative distillation,” Comput. Res. Repository (CoRR) , vol. abs/2302.12255, 2023

  11. [19]

    PrideDiff: Physics-regularized generalized diffusion model for CT reconstruction,

    Z. Lu and et al., “PrideDiff: Physics-regularized generalized diffusion model for CT reconstruction,” IEEE Trans. Radiat. Plasma Med. Sci. , 2024

  12. [20]

    DPER: Diffusion prior driven neural representation for limited angle and sparse view CT Reconstruction,

    C. Du and et al., “DPER: Diffusion prior driven neural representation for limited angle and sparse view CT Reconstruction,” Comput. Res. Repository (CoRR) , vol. abs/2401.07516, 2024

  13. [21]

    Deep learning for tomographic image reconstruction,

    G. Wang, J. C. Ye, and B. De Man, “Deep learning for tomographic image reconstruction,” Nat. Mach. Intell. , vol. 2, no. 12, pp. 737–748, 2020

  14. [22]

    Prior frequency guided diffusion model for limited angle (LA)-CBCT reconstruction,

    J. Xie, H.-C. Shao, Y . Li, and Y . Zhang, “Prior frequency guided diffusion model for limited angle (LA)-CBCT reconstruction,” Phys. Med. Biol. , vol. 69, no. 13, p. 135008, 2024

  15. [23]

    Scalable diffusion models with transformers,

    W. Peebles and S. Xie, “Scalable diffusion models with transformers,” in Proc. IEEE/CVF Int. Conf. Comput. Vis. (ICCV) , 2023, pp. 4195–4205

  16. [24]

    Lose the views: Limited angle CT reconstruction via implicit sinogram completion,

    R. Anirudh and et al., “Lose the views: Limited angle CT reconstruction via implicit sinogram completion,” in Proc. IEEE Conf. Comput. Vis. Pattern Recognit. (CVPR) , 2018, pp. 6343–6352

  17. [25]

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

    J. Liu and et al., “Dolce: A model-based probabilistic diffusion frame- work for limited-angle ct reconstruction,” in Proc. IEEE/CVF Int. Conf. Comput. Vis. (ICCV) , 2023, pp. 10 498–10 508

  18. [26]

    Time-reversion fast-sampling score-based model for limited-angle CT reconstruction,

    Y . Wang, Z. Li, and W. Wu, “Time-reversion fast-sampling score-based model for limited-angle CT reconstruction,” IEEE Trans. Med. Imaging , 2024

  19. [27]

    Wavelet-inspired multi-channel score-based model for limited-angle CT reconstruction,

    J. Zhang, H. Mao, X. Wang, Y . Guo, and W. Wu, “Wavelet-inspired multi-channel score-based model for limited-angle CT reconstruction,” IEEE Trans. Med. Imaging , 2024

  20. [28]

    Diffusion probabilistic model made slim,

    X. Yang, D. Zhou, J. Feng, and X. Wang, “Diffusion probabilistic model made slim,” in Proc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR), 2023, pp. 22 552–22 562

  21. [29]

    Frequency-Aware diffusion model for multi-modal MRI image synthesis,

    M. Jiang and et al., “Frequency-Aware diffusion model for multi-modal MRI image synthesis,” J. Imaging, vol. 11, no. 5, p. 152, 2025

  22. [30]

    Fast controllable diffusion models for undersampled MRI reconstruction,

    W. Jiang, Z. Xiong, F. Liu, N. Ye, and H. Sun, “Fast controllable diffusion models for undersampled MRI reconstruction,” in Proc. IEEE Int. Symp. Biomed. Imaging (ISBI) . IEEE, 2024, pp. 1–5

  23. [31]

    Fast-DDPM: Fast denoising diffusion probabilistic models for medical image-to-image generation,

    H. Jiang and et al., “Fast-DDPM: Fast denoising diffusion probabilistic models for medical image-to-image generation,” IEEE J. Biomed. Health Inform., 2025

  24. [32]

    A fast iterative shrinkage-thresholding algorithm for linear inverse problems,

    A. Beck and M. Teboulle, “A fast iterative shrinkage-thresholding algorithm for linear inverse problems,” SIAM J. Imaging Sci. , vol. 2, no. 1, pp. 183–202, 2009

  25. [33]

    Deep convo- lutional neural network for inverse problems in imaging,

    K. H. Jin, M. T. McCann, E. Froustey, and M. Unser, “Deep convo- lutional neural network for inverse problems in imaging,” IEEE Trans. Image Process., vol. 26, no. 9, pp. 4509–4522, 2017

  26. [34]

    Low-dose CT image denoising using a generative adversarial network with Wasserstein distance and perceptual loss,

    Q. Yang and et al., “Low-dose CT image denoising using a generative adversarial network with Wasserstein distance and perceptual loss,” IEEE Trans. Med. Imaging , vol. 37, no. 6, pp. 1348–1357, 2018

  27. [35]

    Diffusion posterior sampling for general noisy inverse problems,

    H. Chung, J. Song, and S. Ermon, “Diffusion posterior sampling for general noisy inverse problems,” in Proc. Int. Conf. Learn. Represent. (ICLR), 2023

  28. [36]

    Denoising diffusion probabilistic models,

    J. Ho, A. Jain, and P. Abbeel, “Denoising diffusion probabilistic models,” Adv. Neural Inf. Process. Syst. , vol. 33, pp. 6840–6851, 2020

  29. [37]

    Score-based generative modeling through stochastic differential equations,

    Y . Song and et al., “Score-based generative modeling through stochastic differential equations,” 2020, arXiv:2011.13456

  30. [38]

    Denoising diffusion implicit models,

    J. Song, C. Meng, and S. Ermon, “Denoising diffusion implicit models,” in Proc. Int. Conf. Learn. Represent. (ICLR) , 2021

  31. [39]

    On distillation of guided diffusion models,

    C. Meng and et al., “On distillation of guided diffusion models,” in Proc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR) , 2023, pp. 14 297–14 306

  32. [40]

    Progressive distillation for fast sampling of diffusion models,

    T. Salimans and J. Ho, “Progressive distillation for fast sampling of diffusion models,” in Proc. Int. Conf. Learn. Represent. (ICLR) , 2022

  33. [41]

    Error analysis for filtered back projection reconstructions in besov spaces,

    M. Beckmann, P. Maass, and J. Nickel, “Error analysis for filtered back projection reconstructions in besov spaces,” Inverse Problems, vol. 37, no. 1, p. 014002, 2020

  34. [42]

    A. C. Kak and M. Slaney, Principles of computerized tomographic imaging. Philadelphia, PA: SIAM, 2001

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.