Pith. sign in

REVIEW 4 major objections 5 minor 2 cited by

Zero-shot CT Super-Resolution using Diffusion-based 2D Projection Priors and Signed 3D Gaussians

T0 review · 4 major / 5 minor · reviewed 2026-08-05 · deepseek-v4-flash

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

Pith's one-line read Zero-shot 3D CT super-resolution can beat a supervised baseline, the paper claims, by upsampling 2D projections with a chest-X-ray diffusion prior and reconstructing with a Gaussian-splatting variant that learns signed residuals.

desk verdict A genuinely new combination of 2D diffusion priors and signed-residual 3DGS, but the evaluation overstates the evidence—especially the unmeasured transfer to oblique projections and test-set-tuned hyperparameters. read the letter →

arxiv 2508.15151 v3 pith:3SN4IJ3Z submitted 2025-08-21 eess.IV cs.CV

Jeonghyun Noh , Hyun-Jic Oh , Won-Ki Jeong This is my paper
classification eess.IVcs.CV
keywords zero-shotsuper-resolutionCTdiffusionmodelprior3DGaussiansplattingnegativealphablendingresiduallearningX-rayprojectionDDNM
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 tries to establish that high-resolution CT volumes can be recovered from a single low-resolution scan without any paired high-resolution training data, by borrowing a high-frequency prior from ordinary 2D X-ray images. The strategy is to split the task: first super-resolve each of the LR volume's 2D projections with a diffusion model trained on large chest-X-ray sets, then reconstruct the 3D volume with Gaussian splatting. The new reconstruction mechanism, NAB-GS, allows Gaussian densities to be negative so the splats can literally learn the residual between the diffusion-enhanced projections and the original low-resolution ones. If the claim holds, any hospital with a routine low-dose CT scan and a pretrained 2D X-ray model could sharpen volume detail at 4x without ever collecting paired LR-HR CT data. The paper reports exactly that outcome on two public datasets: it beats the established zero-shot method on every reported metric and a fully supervised method on PSNR.

What carries the argument

The load-bearing object is NAB-GS (negative alpha blending Gaussian splatting): radiative 3D Gaussian splatting where each Gaussian density is activated by leaky ReLU (negative slope gamma) instead of softplus. Negative densities yield negative opacities in alpha compositing, so the field can represent signed residuals between the diffusion-enhanced target projections and the reprojections of the upsampled LR volume, adding structure and subtracting blur. Standard 3DGS's non-negativity made residuals inexpressible; gamma controls gradient magnitude and near-zero densities are pruned. The companion mechanism is DDNM with per-projection adaptive sampling (PAS): the diffusion refinement start s

What would settle it

On a CT volume outside the chest, render true HR projections and compare per gantry angle: if diffusion-upsampled projections beat cubic interpolation near frontal and lateral angles but not at oblique angles, or if the reconstructed volume's PSNR falls below cubic interpolation, the transfer premise fails. A cheaper check: retrain the same diffusion stage on CT-derived projections rather than chest X-rays; a large PSNR gain would show the domain gap is eating the prior's value.

Watch

Extended reading notes

Core claim

The central claim: fine detail for a low-resolution CT volume can come from outside the volume — abundant 2D high-resolution X-ray images, not paired 3D CT data. The pipeline super-resolves each 2D projection of the LR scan with a denoising diffusion null-space model (DDNM) trained on chest X-rays, using per-projection adaptive sampling to suppress domain-gap artifacts. It then reconstructs the volume with radiative Gaussian splatting extended to negative densities (NAB-GS, via leaky ReLU), so splats learn the residual between diffusion-enhanced and LR projections, adding structure where the volume is smooth and subtracting blur elsewhere. Results reported: it beats zero-shot CuNeRF on every

Load-bearing premise

Everything rests on a diffusion model trained on frontal and lateral chest X-rays being a trustworthy source of fine detail for CT projections of other anatomies, scanners, and angles; if the prior does not transfer, the upsampled projections invent structure and the final volume is no better than interpolation (the paper itself concedes the full 0-180 degree range is not consistently covered).

Editorial extensions

If this is right

  • If the claim holds, zero-shot CT super-resolution no longer needs paired HR volumes: any LR scan plus a diffusion model trained on ordinary 2D chest X-rays is enough to sharpen the volume at 4x and 8x.
  • NAB-GS's signed densities make residual learning expressible in 3DGS for the first time; the paper's ablation attributes a roughly 0.4 dB PSNR gain on MELA 4x to this change alone.
  • Per-projection adaptive sampling beats every fixed diffusion start step tested, so the framework adapts restoration strength per projection rather than assuming one schedule fits all angles.
  • On both public datasets the method outperforms the zero-shot baseline on all metrics and the supervised ArSSR on PSNR (with competitive SSIM), and the abstract reports expert evaluation supporting clinical potential at 4x.

Reading between the lines

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

  • A natural untested corollary: the two-stage split (2D projection SR plus signed-density 3D reconstruction) transfers to other tomographic modalities — PET, SPECT, interventional cone-beam CT — wherever paired HR volumes are scarce but 2D projections are cheap.
  • Because the paper concedes its prior is trained on frontal and lateral views, per-angle gains are likely uneven: a per-projection PSNR curve against true HR projections would show where the prior stops helping, and PAS probably matters most at oblique angles where the prior is weakest.
  • The signed-density mechanism is a general device: any 3DGS task that models differences — motion- or artifact-corrected imaging, change detection, subtraction studies — could adopt leaky-ReLU densities the same way.
  • The residual-learning design invites a cheaper alternative reading: a large part of the volume-level gain may come from NAB-GS itself rather than the diffusion prior; replacing DDNM with cubic upsampling inside NAB-GS would isolate the prior's marginal contribution across anatomies.
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

4 major / 5 minor

Summary. The paper proposes a zero-shot 3D CT super-resolution framework that first upsamples low-resolution 2D CT projections using a diffusion model trained on large-scale chest X-ray images, with a per-projection adaptive sampling (PAS) strategy, and then reconstructs the 3D volume with 3D Gaussian splatting. To allow residual learning between the upsampled LR projections and the diffusion-enhanced projections, the authors introduce NAB-GS, which extends 3DGS by allowing negative Gaussian densities via a leaky ReLU activation. The method is evaluated on MELA and UHRCT datasets against traditional interpolation, supervised ArSSR, and zero-shot CuNeRF, and is supported by ablations on the diffusion starting step, activation function, negative slope, and reconstruction backbones.

Significance. If the reported results are robust, the paper makes a useful contribution: it provides a practical way to inject external 2D X-ray priors into 3D CT SR without paired HR CT volumes, and NAB-GS is a technically interesting extension of 3DGS that enables signed residual learning. The ablations are reasonably thorough and the pipeline components are clearly described. However, the strength of the 3D claim depends on the transferability of a chest X-ray diffusion prior to CT projections from other anatomies and across all angles, and the current evaluation provides only weak direct evidence for that transfer. The hyperparameter selection and small test sets further limit the confidence in the quantitative conclusions.

major comments (4)
  1. [Supplementary §9.3, Table 9; §5.3] The direct evidence for the 2D diffusion prior is weaker than the 3D results imply. At 8×, the proposed DDNM+PAS upsampling has SSIM 0.9107, below cubic (0.9180) and trilinear (0.9205); at 4× it is below trilinear (0.9643 vs 0.9652). Yet Table 1 reports that the full pipeline improves 3D SSIM over interpolation at both scales (e.g., MELA 8×: 0.9153 vs 0.9005 cubic). Because the only external information entering the 3D reconstruction is these 2D projections, the origin of the 3D gain is not established. Please report per-angle or angular-bin 2D metrics and an analysis of how much of the final 3D improvement is attributable to genuine projection enhancement versus hallucinated oblique-angle content fitted by NAB-GS.
  2. [§5.1, Supplementary §9.1] Hyperparameters appear to be selected on the evaluation data. τthr is set to 7 (4×) and 11 (8×); γ is tuned separately for MELA 4×/8× and UHRCT 4×/8× (0.09/0.003/0.03/0.003) after inspecting PSNR/SSIM on the test volumes (Tables 6–7). With only 20 and 10 test volumes and no error bars or significance tests, the reported margins over cubic (0.13–0.80 dB PSNR; 0.013–0.024 SSIM) may not be stable. Section 6 also concedes that the optimal γ varies with dataset. Please provide a validation-based hyperparameter selection protocol, report per-volume metrics with variance, and run significance tests (or state uncertainty).
  3. [§6, §5.1] The paper acknowledges that the diffusion prior is 'limited to frontal and lateral views' and that enhancement across 0–180° is non-trivial, yet the pipeline runs the same prior on 100 projections uniformly spaced between 0° and 180°. PAS (Eq. 10) only chooses a starting timestep; it does not add missing high-frequency content for off-distribution angles. No angle-resolved analysis is provided, so the central premise—that an external 2D X-ray prior transfers to all projection angles—remains unsupported. At minimum, report metrics broken down by angular bin and the distribution of tstart across angles.
  4. [Abstract, §5] The abstract states that 'expert evaluations present the framework's clinical potential at 4x', but no expert evaluation protocol, results, or statistical analysis is included in the main text or supplementary materials. This claim is either missing support or should be removed from the abstract.
minor comments (5)
  1. [§1] Typo: 'resulting in to a degradation' should be 'resulting in a degradation'.
  2. [Title] The arXiv title ('Zero-shot CT Super-Resolution using Diffusion-based 2D Projection Priors and Signed 3D Gaussians') and the main-text title ('Zero-shot Volumetric CT Super-Resolution using 3D Gaussian Splatting with Upsampled 2D X-ray Projection Priors') should be aligned.
  3. [§5.1] The sentence 'Note that unsupervised learning only uses a testset' is ambiguous. Clarify whether any validation split is used for selecting PAS thresholds and γ, and whether the reported numbers are on the same volumes used for hyperparameter tuning.
  4. [Eq. (4)] The notation for µi is unclear: 'µi is the covariance-related factor,(2π|Σi/Σ̂i|1/2)' appears to have a missing 'where' and possible missing parentheses. Please rewrite.
  5. [Table 1] The table caption does not state the number of test volumes (20 for MELA, 10 for UHRCT). Please add this information and report per-volume standard deviations.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the derivation chain is driven by an external diffusion prior and independent reconstruction baselines; noted test-set hyperparameter tuning is a rigor concern, not a circular step.

full rationale

The paper's central derivation is not circular. The 2D projection SR uses an unconditional DDPM trained on external X-ray datasets (ChestX-ray14, CheXpert), then applies DDNM with data consistency (Eq. 6: x̂0|t = A†y + (I − A†A)x0|t), so the upsampled projections are constrained by the LR projections and the external prior. The 3D reconstruction via NAB-GS fits Gaussian residuals between the upsampled LR volume and the diffusion-upsampled projections (Section 4.2); this is a supervised fitting to externally generated targets, not a self-consistency loop. No equation in the paper reduces to another by construction, and no load-bearing claim is justified only by a self-citation: references to DDNM, R2-GS, CuNeRF, ArSSR, and the diffusion literature are all prior external work. The per-dataset/per-resolution selection of the leaky-ReLU slope γ (Supplementary Tables 6–7) and the PAS threshold τthr is hyperparameter tuning; even if performed on evaluation data, it is a model-selection weakness that can overstate absolute gains, not a fitted parameter renamed as a prediction. Similarly, the modest 2D projection gains and lower SSIM at 8× (Table 9) raise a substantive question about whether the diffusion prior transfers across the full 0–180° angular range, but that is a correctness/validity concern about the external prior, not circularity. The framework is self-contained against external benchmarks and the reported improvements are not forced by the paper's own definitions.

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

The central claim rests on a pre-trained diffusion prior (external data) and several hand-tuned hyperparameters, especially gamma and tau_thr, which are selected on the evaluation datasets. This limits the zero-shot generalization claim.

free parameters (4)
  • negative slope gamma of leaky ReLU = 0.09 (MELA 4x), 0.003 (MELA 8x), 0.03 (UHRCT 4x), 0.003 (UHRCT 8x)
    Controls gradient magnitude for residual learning with negative densities; optimal value varies per dataset and resolution and is selected on the evaluation data (Supplementary 9.1, Section 6).
  • PAS threshold tau_thr = 7 (4x), 11 (8x)
    Determines the diffusion starting timestep per projection; set by hand for each upsampling factor (Section 5.1).
  • Loss weights lambda1, lambda2 = 0.5, 0.05
    Set by hand for the total loss (Section 4.2).
  • DDNM noise level sigma_y = 0.0015
    Noise level for DDNM+ data consistency (Section 5.1).
assumptions (5)
  • domain assumption The 2D X-ray diffusion prior transfers to CT projections from other anatomies and scanners
    DDNM is applied to LR projections of MELA and UHRCT using a model trained on chest X-rays; Section 6 acknowledges the domain gap and the limitation to frontal/lateral views.
  • domain assumption The downsampling degradation (sinc interpolation with Gaussian smoothing) accurately simulates low-resolution CT
    LR volumes are generated this way in Section 5.1; the method's evaluation depends on this model of degradation.
  • domain assumption The radiative Gaussian model from R2-GS adequately represents CT density fields
    Adopted in Section 3 as the basis for 3D reconstruction.
  • ad hoc to paper Negative density values in the Gaussian field yield a stable optimization for residual learning
    NAB-GS replaces softplus with leaky ReLU; the authors show it helps empirically but provide no theoretical justification that negative densities preserve the X-ray transform inversion.
  • domain assumption Diffusion-upsampled projections are reliable targets for the residual loss
    The reconstruction is supervised with the diffusion output as pseudo-ground truth; errors in the diffusion output propagate into the volume (Section 4.2).

how reviews work

0 comments
Cite this review

Pith. "Pith review of Zero-shot CT Super-Resolution using Diffusion-based 2D Projection Priors and Signed 3D Gaussians." pith.science (2026). https://pith.science/paper/3SN4IJ3Z

@misc{pith2026250815151,
  author       = {Pith},
  title        = {Pith review of: Zero-shot CT Super-Resolution using Diffusion-based 2D Projection Priors and Signed 3D Gaussians},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/3SN4IJ3Z}},
  note         = {Machine review of arXiv:2508.15151}
}
read the original abstract

Computed tomography (CT) is important in clinical diagnosis, but acquiring high-resolution (HR) CT is constrained by radiation exposure risks. While deep learning-based super-resolution (SR) methods have shown promise for reconstructing HR CT from low-resolution (LR) inputs, supervised approaches require paired datasets that are often unavailable. Zero-shot methods address this limitation by operating on single LR inputs; however, they frequently fail to recover fine structural details due to limited LR information within individual volumes. To overcome these limitations, we propose a novel zero-shot 3D CT SR framework that integrates diffusion-based upsampled 2D projection priors into the 3D reconstruction process. Specifically, our framework consists of two stages: (1) LR CT projection SR, training a diffusion model on abundant X-ray data to upsample LR projections, thereby enhancing the scarce information inherent in the LR inputs. (2) 3D CT volume reconstruction, using 3D Gaussian splatting with our novel Negative Alpha Blending (NAB-GS), which models positive and negative Gaussian densities to learn signed residuals between diffusion-generated HR and upsampled LR projections. Our framework demonstrates superior quantitative and qualitative performance on two public datasets, and expert evaluations present the framework's clinical potential at 4x.

Figures

Figures reproduced from arXiv: 2508.15151 by the authors.

Figure 1
Figure 1. Overview of the proposed zero-shot 3D CT SR framework. (a) 2D X-ray Projection SR: LR 2D X-ray projection [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Visual comparisons of 3D CT reconstruction results. (a) Ground truth, (b) Trilinear, (c) Cubic, (d) ArSSR (Wu et al. [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 3
Figure 3. Visualization results of 3D CT reconstruction across different activation functions. (a) GT, (b) [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: Visual comparisons of datasets. (a) Diffusion train [PITH_FULL_IMAGE:figures/full_fig_p011_4.png]
Figure 5
Figure 5. Figure 5: Visual comparisons of 3D CT reconstruction results on the MELA. (a) Ground truth, (b) Trilinear, (c) Cubic, (d) [PITH_FULL_IMAGE:figures/full_fig_p012_5.png]
Figure 6
Figure 6. Figure 6: Visual comparisons of 3D CT reconstruction results on the UHRCT. (a) Ground truth, (b) Trilinear, (c) Cubic, (d) [PITH_FULL_IMAGE:figures/full_fig_p013_6.png]
Figure 7
Figure 7. Figure 7: Visual comparisons using 3D CT reconstruction methods on the MELA. (a) Ground truth, (b) FDK (Feldkamp, Davis, [PITH_FULL_IMAGE:figures/full_fig_p014_7.png]

Discussion (0). Sign in to comment.

Forward citations

Cited by 2 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. K-space Gaussian Representation for Parallel MRI

    cs.CV 2026-07 conditional novelty 6.0 of 10

    K-space Gaussian Representation (KGR) fits measured multi-coil k-space with shared Gabor-Gaussian primitives and refines the result with Hankel low-rank projection, improving accelerated MRI reconstruction over baseli...

  2. K-space Gaussian Representation for Parallel MRI

    cs.CV 2026-07 conditional novelty 6.0 of 10

    KGR fits shared Gabor-Gaussian primitives directly in multi-coil k-space and applies low-rank projection, consistently outperforming representative parallel-MRI baselines on brain datasets.

Reference graph

Works this paper leans on

40 extracted references · 35 canonical work pages · cited by 1 Pith paper

  1. [1]

    , " * write output.state after.block = add.period write newline

    ENTRY address archivePrefix author booktitle chapter edition editor eid eprint howpublished institution isbn journal key month note number organization pages publisher school series title type volume year label extra.label sort.label short.list INTEGERS output.state before.all mid.sentence after.sentence after.block FUNCTION init.state.consts #0 'before.a...

  2. [2]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION word.in bbl.in capitalize " " * FUNCT...

  3. [3]

    H.; and Kak, A

    Andersen, A. H.; and Kak, A. C. 1984. Simultaneous algebraic reconstruction technique (SART): a superior implementation of the ART algorithm. Ultrasonic imaging, 6(1): 81--94

    work page 1984

  4. [4]

    Biguri, A.; Dosanjh, M.; Hancock, S.; and Soleimani, M. 2016. TIGRE : a MATLAB-GPU toolbox for CBCT image reconstruction. Biomed. Phys. Eng. Express (BPEE), 2(5): 055010

    work page 2016

  5. [5]

    Cai, Y.; Wang, J.; Yuille, A.; Zhou, Z.; and Wang, A. 2024. Structure-aware sparse-view x-ray 3d reconstruction. In Proc. Comput. Vis. Pattern Recognit. (CVPR) , 11174--11183

    work page 2024

  6. [6]

    Chen, X.; Wang, X.; Zhou, J.; Qiao, Y.; and Dong, C. 2023 a . Activating more pixels in image super-resolution transformer. In Proc. Comput. Vis. Pattern Recognit. (CVPR) , 22367--22377

    work page 2023

  7. [7]

    Chen, Z.; Yang, L.; Lai, J.-H.; and Xie, X. 2023 b . Cu NeRF : Cube-based neural radiance field for zero-shot medical image arbitrary-scale super resolution. In Proc. Comput. Vis. Pattern Recognit. (CVPR) , 21185--21195

    work page 2023

  8. [8]

    Chu, Y.; Zhou, L.; Luo, G.; Qiu, Z.; and Gao, X. 2023. Topology-Preserving Computed Tomography Super-Resolution Based on Dual-Stream Diffusion Model. In Proc. Int'l Conf. Med. Image Comput. Comput. Assist. Interv. (MICCAI) , 260--270

    work page 2023

Show all 40 references
  1. [9]

    C.; He, K.; and Tang, X

    Dong, C.; Loy, C. C.; He, K.; and Tang, X. 2015. Image super-resolution using deep convolutional networks. IEEE Trans. Pattern Anal. Mach. Intell. (TPAMI) , 38(2): 295--307

    2015

  2. [10]

    C.; Yan, K.; Yao, J.; Chen, X.; Liu, Z.; Lu, L.; et al

    Fang, W.; Tang, Y.; Guo, H.; Yuan, M.; Mok, T. C.; Yan, K.; Yao, J.; Chen, X.; Liu, Z.; Lu, L.; et al. 2024. Cycle INR : Cycle Implicit Neural Representation for Arbitrary-Scale Volumetric Super-Resolution of Medical Data. In Proc. Comput. Vis. Pattern Recognit. (CVPR) , 11631--11641

    2024

  3. [11]

    A.; Davis, L

    Feldkamp, L. A.; Davis, L. C.; and Kress, J. W. 1984. Practical cone-beam algorithm. Journal of the Optical Society of America A, 1(6): 612--619

    1984

  4. [12]

    L.; Louis, N.; Zbijewski, W.; Vaishnav, J.; Clark, K.; and Nicolella, D

    Frazer, L. L.; Louis, N.; Zbijewski, W.; Vaishnav, J.; Clark, K.; and Nicolella, D. P. 2024. Super-resolution of clinical CT : Revealing microarchitecture in whole bone clinical CT image data. Bone, 185: 117115

    2024

  5. [13]

    Ho, J.; Jain, A.; and Abbeel, P. 2020. Denoising diffusion probabilistic models. Adv. Neural Inf. Process. Syst. (NeurIPS) , 33: 6840--6851

    2020

  6. [14]

    Kawar, B.; Elad, M.; Ermon, S.; and Song, J. 2022. Denoising diffusion restoration models. Adv. Neural Inf. Process. Syst. (NeurIPS) , 35: 23593--23606

    2022

  7. [15]

    Kerbl, B.; Kopanas, G.; Leimk \"u hler, T.; and Drettakis, G. 2023. 3D Gaussian splatting for real-time radiance field rendering. ACM Trans. on Graph. (ToG) , 42(4): 139--1

    2023

  8. [16]

    Ledig, C.; Theis, L.; Husz \'a r, F.; Caballero, J.; Cunningham, A.; Acosta, A.; Aitken, A.; Tejani, A.; Totz, J.; Wang, Z.; et al. 2017. Photo-realistic single image super-resolution using a generative adversarial network. In Proc. Comput. Vis. Pattern Recognit. (CVPR) , 4681--4690

    2017

  9. [17]

    Liang, J.; Cao, J.; Sun, G.; Zhang, K.; Van Gool, L.; and Timofte, R. 2021. Swin IR : Image restoration using swin transformer. In Proc. of Int'l Conf. on Comput. Vis. (ICCV) , 1833--1844

    2021

  10. [18]

    Liu, G.; Sun, H.; Li, J.; Yin, F.; and Yang, Y. 2023. Accelerating diffusion models for inverse problems through shortcut sampling. arXiv preprint arXiv:2305.16965

    2023 arXiv

  11. [19]

    Lu, Z.; Li, J.; Liu, H.; Huang, C.; Zhang, L.; and Zeng, T. 2022. Transformer for single image super-resolution. In Proc. Comput. Vis. Pattern Recognit. (CVPR) , 457--466

    2022

  12. [20]

    R.; and Semelka, R

    Martin, D. R.; and Semelka, R. C. 2006. Health effects of ionising radiation from diagnostic CT . The Lancet, 367(9524): 1712--1714

    2006

  13. [21]

    H.; Yu, L.; Kofler, J

    McCollough, C. H.; Yu, L.; Kofler, J. M.; Leng, S.; Zhang, Y.; Li, Z.; and Carter, R. E. 2015. Degradation of CT low-contrast spatial resolution due to the use of iterative reconstruction and reduced dose levels. Radiology, 276(2): 499--506

    2015

  14. [22]

    Q.; and Dhariwal, P

    Nichol, A. Q.; and Dhariwal, P. 2021. Improved denoising diffusion probabilistic models. In International conference on machine learning, 8162--8171. PMLR

    2021

  15. [23]

    Peng, C.; Lin, W.-A.; Liao, H.; Chellappa, R.; and Zhou, S. K. 2020. Saint: Spatially aware interpolation network for medical slice synthesis. In Proc. Comput. Vis. Pattern Recognit. (CVPR) , 7750--7759

    2020

  16. [24]

    Rombach, R.; Blattmann, A.; Lorenz, D.; Esser, P.; and Ommer, B. 2022. High-resolution image synthesis with latent diffusion models. In Proc. Comput. Vis. Pattern Recognit. (CVPR) , 10684--10695

    2022

  17. [25]

    J.; and Norouzi, M

    Saharia, C.; Ho, J.; Chan, W.; Salimans, T.; Fleet, D. J.; and Norouzi, M. 2022. Image super-resolution via iterative refinement. IEEE Trans. Pattern Anal. Mach. Intell. (TPAMI) , 45(4): 4713--4726

    2022

  18. [26]

    Shocher, A.; Cohen, N.; and Irani, M. 2018. Zero-shot super-resolution using deep internal learning. In Proc. Comput. Vis. Pattern Recognit. (CVPR) , 3118--3126

    2018

  19. [27]

    Y.; and Pan, X

    Sidky, E. Y.; and Pan, X. 2008. Image reconstruction in circular cone-beam computed tomography by constrained, total-variation minimization. Physics in Medicine & Biology, 53(17): 4777

    2008

  20. [28]

    Sitzmann, V.; Martel, J.; Bergman, A.; Lindell, D.; and Wetzstein, G. 2020. Implicit neural representations with periodic activation functions. Adv. Neural Inf. Process. Syst. (NeurIPS) , 33: 7462--7473

    2020

  21. [29]

    W.; Cho, S.; and Cho, N

    Soh, J. W.; Cho, S.; and Cho, N. I. 2020. Meta-transfer learning for zero-shot super-resolution. In Proc. Comput. Vis. Pattern Recognit. (CVPR) , 3516--3525

    2020

  22. [30]

    Song, J.; Meng, C.; and Ermon, S. 2020. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502

    2020 arXiv

  23. [31]

    Wang, J.; Wang, R.; Tao, R.; and Zheng, G. 2022. UASSR : Unsupervised arbitrary scale super-resolution reconstruction of single anisotropic 3D images via disentangled representation learning. In Proc. Int'l Conf. Med. Image Comput. Comput. Assist. Interv. (MICCAI) , 453--462

    2022

  24. [32]

    C.; and Wen, B

    Wang, Y.; Yang, W.; Chen, X.; Wang, Y.; Guo, L.; Chau, L.-P.; Liu, Z.; Qiao, Y.; Kot, A. C.; and Wen, B. 2024. Sin SR : Diffusion-based image super-resolution in a single step. In Proc. Comput. Vis. Pattern Recognit. (CVPR) , 25796--25805

    2024

  25. [33]

    Wang, Y.; Yu, J.; and Zhang, J. 2022. Zero-shot image restoration using denoising diffusion null-space model. arXiv preprint arXiv:2212.00490

    2022 arXiv

  26. [34]

    C.; Sheikh, H

    Wang, Z.; Bovik, A. C.; Sheikh, H. R.; and Simoncelli, E. P. 2004. Image quality assessment: from error visibility to structural similarity. IEEE Trans. Image Process. (TIP) , 13(4): 600--612

    2004

  27. [35]

    Wu, Q.; Li, Y.; Sun, Y.; Zhou, Y.; Wei, H.; Yu, J.; and Zhang, Y. 2022. An arbitrary scale super-resolution approach for 3d MR images via implicit neural representation. IEEE J. Biomed. Health Inform. (JBHI), 27(2): 1004--1015

    2022

  28. [36]

    W.; and Zhang, R

    Yu, P.; Zhang, H.; Kang, H.; Tang, W.; Arnold, C. W.; and Zhang, R. 2022. RPLHR-CT dataset and transformer baseline for volumetric super-resolution from CT scans. In Proc. Int'l Conf. Med. Image Comput. Comput. Assist. Interv. (MICCAI) , 344--353

    2022

  29. [37]

    J.; Cai, Y.; Cao, J.; Zhang, Y.; and Li, H

    Zha, R.; Lin, T. J.; Cai, Y.; Cao, J.; Zhang, Y.; and Li, H. 2024. R ^2 - G aussian: Rectifying Radiative Gaussian Splatting for Tomographic Reconstruction. Adv. Neural Inf. Process. Syst. (NeurIPS)

    2024

  30. [38]

    Zha, R.; Zhang, Y.; and Li, H. 2022. NAF : N eural attenuation fields for sparse-view CBCT reconstruction. In Proc. Int'l Conf. Med. Image Comput. Comput. Assist. Interv. (MICCAI) , 442--452

    2022

  31. [39]

    Zhang, Y.; Tian, Y.; Kong, Y.; Zhong, B.; and Fu, Y. 2018. Residual dense network for image super-resolution. In Proc. Comput. Vis. Pattern Recognit. (CVPR) , 2472--2481

    2018

  32. [40]

    Zhao, K.; Pang, K.; Hung, A. L. Y.; Zheng, H.; Yan, R.; and Sung, K. 2024. Mri super-resolution with partial diffusion models. IEEE Transactions on Medical Imaging (TMI)

    2024

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.