Pith. sign in

REVIEW 4 major objections 4 minor 12 references

Noise Controlled CT Super-Resolution with Conditional Diffusion Model

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

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

Pith's one-line read The paper proposes a conditional diffusion model for CT super-resolution that improves spatial resolution without amplifying noise, by training on a hybrid of noise-matched simulations and segmented real bone details.

desk verdict A sensible hybrid training recipe for CT super-resolution, but the noise-control claim outruns the evidence: two test cases and two ROIs. read the letter →

arxiv 2502.09793 v1 pith:VRLYHSPJ submitted 2025-02-13 cs.CV

Yuang Wang , Siyeop Yoon , Rui Hu , Baihui Yu , Duhgoon Lee , Rajiv Gupta , Li Zhang , Zhiqiang Chen
show 1 more author
Dufan Wu
This is my paper
classification cs.CV
keywords CTsuper-resolutionconditionaldiffusionmodelnoisecontrolhybridtrainingdatanoise-matchedsimulationbonesegmentationtrabeculardetailmedicalimaging
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

The paper proposes a way to do CT super-resolution that avoids the usual noise-amplification trade-off: train a conditional diffusion model on two kinds of data at once. The first kind is simulated high- and low-resolution CT pairs whose injected Gaussian noise is tuned to match the real scanner, so the model learns resolution enhancement without learning to add noise. The second kind is real bone regions, segmented from unmatched CT pairs, which supply trabecular detail that phantoms do not contain. On real temporal-bone CT, the method improves texture similarity to the high-resolution reference while keeping uniform-region noise standard deviation near the low-resolution input, and it outperforms both simulation-only training and noise-unmatched training on that balance.

What carries the argument

The central mechanism is the conditional Denoising Diffusion Probabilistic Model (conditional DDPM), a generative model that learns to reverse a 1,000-step Gaussian noising process while conditioned on the low-resolution image. The noise-control property specifically comes from the training-data construction: correlated Gaussian noise is injected into simulated post-log projections through a single shared noise vector $z$ with hand-set amplitudes $k_{\mathrm{HR}}=0.12$ and $k_{\mathrm{LR}}=0.60$, so the simulated HR and LR images carry noise levels matched to the real LR image. Real bone regions are then segmented and added to the training batch (16 of 64 slices per batch) so the model learns trabecular detail that the numerical phantoms lack. At test time, the model is applied twice, once to the full LR image and once to the bone-only image, and the bone-mask output pixels replace the corresponding pixels of the full-image output.

What would settle it

Scan a uniform phantom with the same photon-counting CT at several dose levels, estimate the noise parameters of Eqs. (14)--(15) directly from the measured noise power spectra instead of setting them by hand, and retrain. If the super-resolved output's noise standard deviation in uniform regions is higher than the LR input's at any of those dose levels, the noise-control claim is falsified; a second check would be to perturb the bone segmentation mask and see whether the claimed texture-detail gain disappears.

Watch

Extended reading notes

Core claim

The central claim is that noise amplification in CT super-resolution is not an unavoidable cost of sharper images; it comes from a mismatch between the noise statistics of the training pairs and the noise statistics of the real scanner. The paper's conditional Denoising Diffusion Probabilistic Model, conditioned on the low-resolution image, is trained on hybrid data: simulated HR-LR pairs generated from numerical phantoms with correlated Gaussian noise matched to the real LR image ($k_{\mathrm{HR}}=0.12$, $k_{\mathrm{LR}}=0.60$), plus real HR-LR bone patches segmented by thresholding and morphological operations. At test time, the model super-resolves the whole image and the segmented bone region separately, then replaces the bone-mask pixels with the bone-only result. On two real temporal-bone test cases this produced texture-feature distances of 325 and 535 in detail ROIs, down from 1367 and 1148 in the LR input, while uniform-ROI standard deviations stayed at 41.5 and 39.9 HU compared with 45.7 and 38.1 HU for the LR image. The paper states this outcome as 'spatial resolution improvement without amplification of the noise.'

Load-bearing premise

The result stands on the assumption that the real scanner's noise is Gaussian, correlated, and well represented by the shared noise vector $z$ with the hand-set constants $k_{\mathrm{HR}}=0.12$ and $k_{\mathrm{LR}}=0.60$, and that the bone segmentation used to inject real details is accurate; the paper itself notes that segmentation errors can add artifacts or lose detail.

Editorial extensions

If this is right

  • On the two real temporal-bone test cases, output noise standard deviations in uniform ROIs (41.5 and 39.9 HU) stayed close to or below the LR input (45.7 and 38.1 HU), while texture-feature distances in detail ROIs dropped from 1367/1148 to 325/535.
  • Training on noise-matched simulation data alone oversmoothes trabecular bone, giving the worst texture distances and the lowest PSNR of the compared methods.
  • Training with unmatched noise levels restores bony detail but amplifies noise, with uniform-ROI standard deviations rising to 54.1 and 51.9 HU.
  • The proposed method leaves PSNR close to the LR input (23.79 and 24.33 dB versus 23.88 and 24.30 dB), indicating the super-resolution does not drastically change large-scale structures.
  • The two-branch design means the final image combines global structure learned from noise-matched simulation with real trabecular detail learned from segmented bones.

Reading between the lines

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

  • The method hand-sets $k_{\mathrm{HR}}$ and $k_{\mathrm{LR}}$; transferring it to another scanner or protocol would require re-estimating these constants from measured system noise, a calibration step the paper does not describe.
  • The use of a single shared noise vector $z$ in Eqs. (14)--(15) means the model is trained assuming the HR and LR noises are perfectly correlated; an experiment drawing independent noise for LR and HR would test whether that correlation is necessary for the noise-control effect.
  • The same hybrid recipe, combining a physics-based noise simulator with real structural detail patches, could transfer to other inverse problems in medical imaging where paired high-quality data are scarce.
  • A testable improvement would be to replace hard threshold-based bone segmentation with soft masks or uncertainty-aware segmentation, directly addressing the artifact risk the paper acknowledges.
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 / 4 minor

Summary. The paper proposes a conditional denoising diffusion probabilistic model (DDPM) for CT super-resolution with explicit noise control. The key idea is to train the model on a hybrid dataset: noise-matched simulated HR/LR projection pairs, generated with a hand-tuned Gaussian correlated noise model, plus bone details segmented from real noise-unmatched CT data. At test time, the model super-resolves the whole LR image and the segmented bone region separately, then fuses the bone output into the whole-image output. The authors evaluate on two real temporal-bone CT cases, comparing their method with a simulation-only baseline (M1) and a hybrid baseline without matched noise (M2), using STD in uniform ROIs, Haralick texture features, and PSNR. They conclude that the proposed method improves spatial resolution without amplifying noise, while M1 oversmooths bony structure and M2 amplifies noise.

Significance. If the central claim is correct, the hybrid training strategy is a practical contribution: it uses simulation data for noise control and a small amount of real segmented data to recover anatomy that phantoms lack, thereby avoiding the need for large noise-matched real training sets. The paper is honest about the dependence on segmentation quality and mentions the black-background training caveat in the limitations. However, the evidence base is narrow. The quantitative support consists of two test cases, two uniform ROIs for noise, and two detail ROIs for texture, with no error bars, no statistical tests, and no noise-power-spectrum validation of the simulation model. The literal claim of 'without amplification of the noise' is already contradicted by the ROI 2 STD increase from 38.1 HU to 39.9 HU. The core idea is plausible, but the current experimental support is not yet commensurate with the strength of the claimed conclusions.

major comments (4)
  1. [Section III, Table I] The quantitative results do not fully support the literal claim that the proposed method achieves super-resolution 'without amplification of the noise.' In ROI 2, the proposed output has an STD of 39.9 HU versus 38.1 HU for the LR input, an increase of 1.8 HU, and this is not discussed. Noise is measured in only two hand-selected uniform ROIs, with no error bars, no per-slice variation, and no statistical test. A whole-image or larger-ROI noise analysis, or at least an explicit explanation for the ROI 2 increase, is needed before the headline claim can be accepted.
  2. [Section II.B, Eqs. (14)-(15)] The noise-control mechanism depends on the assumption that the correlated Gaussian noise model with a shared random vector z, with kHR=0.12 and kLR=0.60 chosen to match the scalar noise level of the real LR image, faithfully represents the real CT system's noise statistics. No comparison of noise power spectra, spatial correlation lengths, or higher-order statistics between simulation and real data is reported. If the real noise is non-Gaussian or has a different correlation structure, the model trained on this simulation data may not transfer the noise-control behavior to real scans. A noise-power-spectrum or covariance comparison between simulated and real repeated measurements would directly test this load-bearing assumption.
  3. [Section II.C and Table I] The simulation HR noise is matched to the real LR image rather than to the real HR image that is used as the evaluation target. In ROI 1 the real HR has an STD of 60.4 HU while the real LR has 45.7 HU. Training the model with HR labels whose noise level is matched to the LR input, and then computing PSNR against the noisier real HR, conflates two possible goals: producing a low-noise HR image (matching the LR noise level) or reproducing the actual noisy HR acquisition. The paper should state explicitly which target distribution the method is designed to match, and either adjust the simulation noise matching or change the evaluation metric accordingly. This issue affects the interpretation of the PSNR values and the comparison with M2.
  4. [Section III, experimental design] The central comparison among the proposed method, M1, and M2 relies on two test cases from a single temporal bone. Several differences are small (e.g., Haralick distance 535 vs. 550 in ROI 4, PSNR 24.33 vs. 24.26 dB in Case 2 for the proposed method versus M2), and no confidence intervals or significance tests are provided. To support the claim that the proposed method is systematically better than M2, the authors should either add more test cases/slices and report variance, or soften the comparative conclusion to a proof-of-concept demonstration.
minor comments (4)
  1. [Section III] The phrase 'Regions of interests' should be 'Regions of interest', and 'M1 oversmoothes' should be 'M1 oversmooths'.
  2. [Section II.C] The multiplication spacing in '3×3 and 6 ×5' is inconsistent; please use a uniform format for binning factors throughout the paper.
  3. [Figure 2] The block labels in Figure 2 (e.g., 'Contract Step (2c)', 'Medium Step (16c)') are not explained in the text; a short description in Section II.C of how these labels map to the U-Net stages would improve readability.
  4. [Section IV] The limitations paragraph mentions that training with segmented details on a completely black background may challenge the conditional diffusion model, but it does not describe any mitigation or analysis of this effect; adding a brief comment on how the model was observed to behave in the black-background regions would be useful.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the noise-control claim is an empirical outcome, not an identity imposed by the training-data construction.

full rationale

The paper's derivation chain is a standard conditional DDPM training pipeline (Eqs. 1-13) from external references; no load-bearing self-citation appears. The only potentially self-referential element is the calibration of kHR=0.12 and kLR=0.60 in Eqs. (14)-(15) to match the simulation noise level to that of the real LR image, followed by evaluation of STD on real LR images. This is not circular: the hyperparameters set the noise level of simulated training projections, but they do not algebraically constrain the diffusion model's output noise. The reported reduction (LR 45.7 HU vs Proposed 41.5 HU in ROI 1) is an empirical result of training and inference, and the paper includes controls (M1, M2) that isolate the effect of the hybrid training and noise matching. The bone-segmentation limitation is acknowledged in the conclusion and concerns robustness, not circularity. No equation or citation reduces the central claim to its own inputs, so the appropriate finding is no significant circularity.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

The central method relies on several modeling choices: Gaussian correlated noise in projection data, matching of simulation geometry to a real scanner, an operational definition of HR/LR via detector rebinning, and accurate bone segmentation. The noise-matching constants kHR and kLR are free parameters fitted to the real data domain. No new physical entities are introduced.

free parameters (3)
  • kHR = 0.12
    Noise scaling in Eq. (14) chosen by hand to match simulated HR noise to real LR noise; fitted to the target dataset.
  • kLR = 0.60
    Noise scaling in Eq. (15) chosen by hand to match simulated LR noise to real LR noise; fitted to the target dataset.
  • bone segmentation threshold = not reported
    Threshold plus morphological operations used to segment bone details from real data; the threshold value is not stated, leaving the preprocessing under-specified.
assumptions (4)
  • domain assumption Quantum noise in CT projections is Gaussian with a shared correlated noise vector z.
    Equations (14)-(15) assume this noise model; no validation against the real scanner's noise statistics is provided.
  • domain assumption Simulation geometry and reconstruction parameters match the real CT acquisition.
    The method relies on XCIST phantoms with the same geometry parameters as the real system; any mismatch would reduce transfer to real data.
  • domain assumption Real HR/LR pairs created by detector rebinning are valid training targets.
    HR is 1x1 binning and LR is 3x3 or 6x5 rebinning; the 1x1 HR has higher noise, so the model may learn to reduce noise while sharpening. The validity of this operational definition is assumed.
  • domain assumption Segmented bone details with a black background can be used as training patches without harming diffusion model performance.
    The authors themselves flag this in Section IV, noting that segmentation errors can introduce artifacts and that the black background may pose challenges.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Noise Controlled CT Super-Resolution with Conditional Diffusion Model." pith.science (2026). https://pith.science/paper/VRLYHSPJ

@misc{pith2026250209793,
  author       = {Pith},
  title        = {Pith review of: Noise Controlled CT Super-Resolution with Conditional Diffusion Model},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/VRLYHSPJ}},
  note         = {Machine review of arXiv:2502.09793}
}
read the original abstract

Improving the spatial resolution of CT images is a meaningful yet challenging task, often accompanied by the issue of noise amplification. This article introduces an innovative framework for noise-controlled CT super-resolution utilizing the conditional diffusion model. The model is trained on hybrid datasets, combining noise-matched simulation data with segmented details from real data. Experimental results with real CT images validate the effectiveness of our proposed framework, showing its potential for practical applications in CT imaging.

Figures

Figures reproduced from arXiv: 2502.09793 by the authors.

Figure 1
Figure 1. Framework of Noise Controlled CT Super-Resolution [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Architecture of the conditional noise predictor. The content in parentheses (c, 2c, 4c, 8c and 16c) after the block name indicates the number of output [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Results on LR CT images of a temporal bone using our proposed framework and two comparison methods M1 and M2. [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

12 extracted references · 9 canonical work pages

  1. [1]

    Song and S

    Y. Song and S. Ermon, ``Generative modeling by estimating gradients of the data distribution,'' Advances in neural information processing systems, vol. 32, 2019

    work page 2019

  2. [2]

    J. Ho, A. Jain, and P. Abbeel, ``Denoising diffusion probabilistic models,'' Advances in neural information processing systems, vol. 33, pp. 6840--6851, 2020

    2020

  3. [3]

    J. Choi, S. Kim, Y. Jeong, Y. Gwon, and S. Yoon, ``Ilvr: Conditioning method for denoising diffusion probabilistic models,'' arXiv preprint arXiv:2108.02938, 2021

    arXiv 2021

  4. [4]

    Saharia, J

    C. Saharia, J. Ho, W. Chan, T. Salimans, D. J. Fleet, and M. Norouzi, ``Image super-resolution via iterative refinement,'' IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 45, no. 4, pp. 4713--4726, 2022

    work page 2022

  5. [5]

    H. Li, Y. Yang, M. Chang, S. Chen, H. Feng, Z. Xu, Q. Li, and Y. Chen, ``Srdiff: Single image super-resolution with diffusion probabilistic models,'' Neurocomputing, vol. 479, pp. 47--59, 2022

    work page 2022

  6. [6]

    Zhang, C

    X. Zhang, C. Feng, A. Wang, L. Yang, and Y. Hao, ``Ct super-resolution using multiple dense residual block based gan,'' Signal, Image and Video Processing, vol. 15, pp. 725--733, 2021

    work page 2021

  7. [7]

    C. You, G. Li, Y. Zhang, X. Zhang, H. Shan, M. Li, S. Ju, Z. Zhao, Z. Zhang, W. Cong, M. W. Vannier, P. K. Saha, E. A. Hoffman, and G. Wang, ``Ct super-resolution gan constrained by the identical, residual, and cycle learning ensemble (gan-circle),'' IEEE Transactions on Medical Imaging, vol. 39, no. 1, pp. 188--203, 2020

    work page 2020

  8. [8]

    M. Wu, P. FitzGerald, J. Zhang, W. P. Segars, H. Yu, Y. Xu, and B. De Man, ``Xcist—an open access x-ray/ct simulation toolkit,'' Physics in Medicine & Biology, vol. 67, no. 19, p. 194002, 2022

    work page 2022

Show all 12 references
  1. [9]

    X. Wang, L. Xie, C. Dong, and Y. Shan, ``Real-esrgan: Training real-world blind super-resolution with pure synthetic data,'' in Proceedings of the IEEE/CVF international conference on computer vision, 2021, pp. 1905--1914

    2021

  2. [10]

    Jabri, D

    A. Jabri, D. Fleet, and T. Chen, ``Scalable adaptive computation for iterative generation,'' arXiv preprint arXiv:2212.11972, 2022

    2022 arXiv

  3. [11]

    R. M. Haralick, K. Shanmugam, and I. H. Dinstein, ``Textural features for image classification,'' IEEE Transactions on systems, man, and cybernetics, no. 6, pp. 610--621, 1973

    1973

  4. [12]

    4O,d (@Yq UK ! #i ' Jb Xb,[ ' ,V>* a JQM 'v Mb /p 34<q>F ?fŬF^bP K

    11em plus .33em minus .07em 4000 4000 100 4000 4000 500 `\.=1000 = #1 \@IEEEnotcompsoconly \@IEEEcompsoconly #1 * [1] 0pt [0pt][0pt] #1 * [1] 0pt [0pt][0pt] #1 * \| ** #1 \@IEEEauthorblockNstyle \@IEEEcompsocnotconfonly \@IEEEauthorblockAstyle \@IEEEcompsocnotconfonly \@IEEEco...

    1988

Pith tools

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