Pith. sign in

REVIEW 3 major objections 5 minor 25 references

ENCORE: Efficient Noise Context-Aware Representation for Low-Dose CT Denoising

T0 review · 3 major / 5 minor · reviewed 2026-08-15 · deepseek-v4-flash

Pith's one-line read Telling CT denoisers about their own noise cuts cost and lifts quality.

desk verdict A well-engineered CT denoising paper that convincingly shows noise-context conditioning helps in simulation, but the headline quality and efficiency claims outrun the evidence: real-world gains are only qualitative and the latency benefit is not realized. read the letter →

arxiv 2608.10343 v1 pith:XSNENLYX submitted 2026-08-11 cs.CV physics.med-ph

classification cs.CVphysics.med-ph
keywords low-doseCTdenoisingnoiseautocovariancecontextadaptiveconvolutionCornish-Fisherexpansionself-supervisedNoise2Noisezero-shotconditionalcomputedtomography
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 a CT denoising network performs better and runs faster when it is explicitly told what its own noise looks like, rather than when it is merely handed a noisy image or a raw random noise sample. The ENCORE framework estimates a local noise autocovariance map from a single synthesized noise realization, feeds that map into adaptive convolution layers whose weights change per image region, and trains the full pipeline under a Noise2Noise self-supervised scheme with a skewness-corrected noise model. In the reported experiments on simulated low-dose CT benchmarks and a real tabletop cone-beam scan, ENCORE improves PSNR, SSIM, and a hallucination metric over vanilla, noise-augmented, and autocovariance-concatenation baselines, with the largest gains at the unseen 10% dose level and on cross-vendor data, while reducing multiply-accumulate operations. If correct, the work shows that packaging noise statistics into a stable context is a more direct route to better CT denoising than merely enlarging the model or adding attention modules.

What carries the argument

The load-bearing object is the noise autocovariance map $V$, computed by Eq. (6) as a local average of products of noise-map values at small spatial lags, with defaults $p=5$, $w=5$, $N=1$ and signed-log normalization. This map is a stable surrogate for the non-stationary, spatially correlated CT noise that generic denoisers assume away. The second load-bearing piece is FlyingConv, an adaptive convolution that predicts spatially varying kernel weights from pooled anatomical features and the autocovariance map, sharing each kernel across two channels and fusing interpolation with convolution to cut memory traffic. The supporting machine is the Cornish-Fisher skewness correction, which replaces the Gaussian noise term in training-pair synthesis with $W(0,V_{\mathrm{target}})=\sqrt{V_{\mathrm{target}}}(\alpha N(0,1)+\beta(N(0,1)^2-1))$.

What would settle it

Run ENCORE on real raw projection data from a scanner with significant scatter and beam hardening and compare it against the simpler autocovariance-concatenation baseline: if ENCORE does not beat it, the FlyingConv mechanism is not carrying the claimed advantage. Separately, acquire many repeated scans of a static phantom and compare the true measured autocovariance with the paper's single-realization $5\times5$ estimator to test the local-stationarity premise directly.

Watch

Extended reading notes

Core claim

The central claim is that CT noise can be turned into a stable, model-friendly input: instead of feeding raw stochastic noise realizations, the method aggregates them into local autocovariance maps that capture noise power and spatial correlation, and it makes those maps actively steer the network through FlyingConv, a convolution whose per-region kernel weights are predicted from anatomy and noise-context features. The paper reports that this combination outperforms all tested variants, including a version that simply concatenates the autocovariance map to the input, most clearly in the ultra-low-dose regime at 10% dose that the model never saw during training. It also reports that scaling the intensity of the noise context at inference changes the output texture in a zero-shot way, letting a single trained model trade denoising strength against texture preservation. A skewness-corrected noise model based on a Cornish-Fisher expansion is used during training-pair generation to better match the physics of photon starvation.

Load-bearing premise

The entire context pipeline rests on the claim that a single synthesized noise realization, averaged over a $5\times5$ window, faithfully represents the true local noise correlation of real CT scans; if scatter, beam hardening, or anatomy-dependent non-stationarity makes that estimate wrong, the context misleads the network.

Editorial extensions

If this is right

  • Self-supervised CT denoising can be improved by changing what is fed into the network rather than by enlarging the model or adding perceptual losses.
  • A single trained model can produce outputs with different residual noise and texture levels by scaling context maps at inference, with no retraining or extra post-processing.
  • The reported multiply-accumulate reductions suggest the approach could fit into clinical pipelines where the full reconstruction-to-denoise chain must run on a single GPU.
  • Because the context is estimated per image, the method should transfer across scanners and reconstruction kernels more gracefully than a static denoiser, consistent with the cross-vendor results.

Reading between the lines

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

  • The authors demonstrate zero-shot dose control only on simulated data; it remains an open extension to show that the same scaling behaves predictably on real clinical raw data where the noise model is imperfect.
  • The fixed $5\times5$ window invites a learned or anatomy-guided aggregation of noise statistics, which could improve accuracy where the local-stationarity assumption is weakest.
  • The same autocovariance-context idea could transfer to other tasks dominated by non-stationary noise, such as iterative reconstruction, metal artifact reduction, or low-dose super-resolution.
  • The gap between multiply-accumulate operations and measured latency, which the paper attributes to memory bandwidth, suggests that hardware-side optimizations for adaptive-weight convolutions are the next lever for speed.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 5 minor

Summary. The paper proposes ENCORE, a low-dose CT denoising framework that explicitly models CT noise statistics. It reformulates noise synthesis using a Cornish-Fisher expansion to capture Poisson skewness, extracts local autocovariance maps as noise context, and introduces a FlyingConv module that adaptively modulates convolution weights based on this context. The method is trained in a Noise2Noise setting and supports zero-shot conditional denoising by scaling the context maps at inference. Evaluations on simulated Mayo2016 and Mayo2020 datasets show consistent PSNR/SSIM/AUHOC improvements over vanilla, NADD, and autocovariance-concatenation baselines, with reduced MACs; a real tabletop phantom study is included for qualitative assessment.

Significance. If the results hold, ENCORE contributes a physically motivated way to inject CT noise statistics into a denoising network, with a clear component-wise ablation and an interesting zero-shot control mechanism. The paper is reproducible in principle (code is provided), and the experimental design is thorough within the simulation setting. The main strength is the coherent pipeline from noise synthesis to context estimation to adaptive convolution. However, the quantitative evidence is limited to simulated data generated with the same noise model used in training, and there is a technical inconsistency in the appendix derivation that must be resolved before the noise-synthesis claims can be accepted.

major comments (3)
  1. [Section III-A / Appendix A] The definition of the skewness-corrected noise W is internally inconsistent. In Eq. (9), V_target is set to d(1-d)P_ND, and this value is used as the variance of W in Eq. (7). When W replaces N(0,P_ND) in Eq. (3), the injected term d a W has variance (d a)^2 V_target = d(1-d) * d(1-d)P_ND, rather than the required d(1-d)P_ND. The final β formula is correct only if W is intended to have variance P_ND, i.e., V_target should be P_ND, not d(1-d)P_ND. As written, the equations would produce noise with the wrong magnitude if implemented literally. Please correct the definitions and clarify whether V_target refers to the variance of W or to the variance of the scaled injected term.
  2. [Section IV-A, Section VI, and Fig. 3] The central claim of substantial denoising gains is quantitatively established only under the same simulated Poisson+Gaussian noise model used for training. The Mayo2020 results use forward-projected reference images with the same noise injection, differing only in acquisition geometry. The only real-data evaluation (Tabletop) is qualitative, has no reference image, and the Discussion explicitly concedes that scatter, beam hardening, and other physical factors are omitted. Because the method's entire mechanism is to feed noise-context maps derived from the simulated model into the network, the external validity of the method on real CT noise is not yet demonstrated. I recommend adding at least one quantitative real-data experiment (e.g., a phantom with known ground truth or a validated no-reference metric) or clearly stating in the abstract and conclusion that the quantitative gains are demonstrated on simulated data.
  3. [Abstract and Table I] The claim of 'substantial gains in ... computational efficiency' is overstated. In Table I, ENCORE consistently has higher latency than the Vanilla baseline (e.g., UNet Base at 10% dose: 10.24 ms vs 8.38 ms), and its MACs are lower only than the +COV and +NADD variants, not than Vanilla. The Discussion (Section VI) attributes this to memory-bandwidth limitations. Please clarify that the efficiency advantage is in MACs and relative to noise-aware baselines, not in end-to-end inference speed against the vanilla network.
minor comments (5)
  1. [Section III-A, Eq. (4)] The text states that the Gaussian term N in Eq. (4) is replaced by W in the final implementation, but the equation itself still shows N(0,P_LD). Please update the equation or add an explicit note to avoid confusion.
  2. [Table I] The underline notation for statistical significance is defined, but in several rows the best-performing ENCORE value is not underlined (e.g., UNet Base 25% SSIM). Please report the actual p-values or specify which comparisons achieved significance to support the 'most cases' statement.
  3. [Section V-C1, Table II] The ablation for the skewness-corrected noise model reports only PSNR and SSIM. Since the correction targets higher-order statistics, consider reporting a metric more sensitive to texture or noise statistics, such as a local noise power spectrum error, to better demonstrate the effect.
  4. [Section III-B, Eq. (6)] The autocovariance estimator with N=1 and w=5 is justified by the local-stationarity assumption [9], but the paper does not quantify the bias/variance of this estimator under simulated noise. A small synthetic experiment showing the estimator's error as a function of w and N would strengthen the default parameter choice.
  5. [References] Reference [24] appears to be titled 'sfrc for assessing hallucinations...'; this is likely a typo for 'SFRC' (structural frequency-domain hallucination metric). Please verify the reference title and the acronym usage.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the noise synthesis, autocovariance context, and FlyingConv modules are derived from explicit physical models and evaluated against independent baselines; same-model simulation is an external-validity caveat, not a circular reduction.

full rationale

The derivation chain is self-contained. The Cornish-Fisher term W in Eq. (7) is constructed to match target variance and skewness from the Poisson model (Eqs. (8)-(10)); beta is derived analytically, not fitted to outcome metrics. The autocovariance maps V in Eq. (6) are computed directly from synthesized noise maps X_n, which are defined from the stated Poisson-plus-Gaussian forward model (Eq. (4)); no parameter is tuned to the reported PSNR/SSIM/AUHOC. The FlyingConv module's adaptive weights are produced by a Predictor network trained end-to-end under the standard N2N MSE loss, so its output is not a renamed input. Citations [5] and [9] are prior external publications by co-authors, but they are used only to motivate the Gaussian baseline and the local-stationarity approximation; the present paper derives its skewness correction and context representation explicitly, and neither cited result is invoked as a uniqueness theorem or as a substitute for the current evaluation. The main caveat is external validity: all quantitative tests use the same synthetic noise model as training, and the real-data evaluation is qualitative, with the paper itself conceding that scatter and beam hardening are omitted (Section VI). That is a generalizability limitation, not a circular reduction of the claimed derivation.

Assumptions & free parameters 4 free parameters · 7 assumptions · 0 invented entities

The framework introduces no new physical entities. The free parameters are empirically chosen hyperparameters. The central derivation rests on the standard CT noise model, the local-stationarity assumption, and the Noise2Noise learning rule.

free parameters (4)
  • Spatial window size w = 5
    Chosen via ablation in Table III on the Mayo2016 dataset. Larger w improves estimation slightly but increases latency; w=5 balances them.
  • Lag patch size p = 5
    Chosen via ablation in Table III; p=5 gives the best PSNR among tested values and matches the receptive field of a 3x3 convolution.
  • Number of noise maps N = 1
    Ablation shows no noticeable gain for N=3 or N=10, while latency increases; N=1 is adopted as default.
  • FlyingConv group size g = 2
    Ablation in Table IV shows g=2 gives the best PSNR/latency balance; larger g reduces memory and latency but hurts PSNR.
assumptions (7)
  • domain assumption CT noise is modeled as Poisson quantum noise plus Gaussian electronic noise (Eq. 1).
    Standard model in CT noise simulation; cited in Section III-A.
  • domain assumption CT noise is approximately stationary within a small region of interest (a few millimeters).
    Used to justify N=1 and w=5 in the autocovariance estimate; cited from Baek and Pelc [9] in Section III-B.
  • domain assumption The Gaussian approximation of Poisson noise is valid for the baseline pair generation in the absence of skewness correction.
    Used in Eq. 3 and in the prior work [5]; the paper argues this fails at low photon counts.
  • standard math Noise2Noise training learns the conditional mean of the noisy target, which is the clean signal.
    The N2N framework (Lehtinen et al., Ref. [3]) relies on two independent noisy observations with the same expectation.
  • domain assumption Forward projection under a monoenergetic X-ray beam assumption is used to simulate raw CT data from reference images.
    Stated in Section IV-A; ignores beam hardening, scatter, and detector energy dependence.
  • standard math The small-beta approximation in the Cornish-Fisher expansion is valid for the noise levels considered.
    The appendix approximates 6*alpha^2*beta + 8*beta^3 as 6*beta, assuming beta is small; checked against photon counts in Table II.
  • domain assumption FBP is linear, so the difference of log-transformed projections reconstructs the image-domain noise map.
    Used in Eq. 4 to define the noise maps X_n; relies on the standard linearity of FBP.

how reviews work

0 comments
Cite this review

Pith. "Pith review of ENCORE: Efficient Noise Context-Aware Representation for Low-Dose CT Denoising." pith.science (2026). https://pith.science/paper/XSNENLYX

@misc{pith2026260810343,
  author       = {Pith},
  title        = {Pith review of: ENCORE: Efficient Noise Context-Aware Representation for Low-Dose CT Denoising},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/XSNENLYX}},
  note         = {Machine review of arXiv:2608.10343}
}
read the original abstract

While deep learning-based denoising has become widely adopted in low-dose CT, conventional models use generic architectures designed for natural images, failing to account for non-stationary and spatially correlated CT noise characteristics. To address this, we propose an Efficient Noise COntext-aware REpresentation (ENCORE) framework that explicitly leverages CT noise characteristics and anatomical features. First, we reformulate the noise synthesis procedure based on a realistic noise distribution beyond the conventional Gaussian approximation, establishing a rigorous foundation for training pair generation. Next, we extract local noise power and correlation contexts to guide the denoising process. To fully leverage the potential of noise context, we propose a FlyingConv module, which adaptively changes convolution weights for each local image region. Notably, our approach demonstrates substantial gains in both denoising quality and computational efficiency. Furthermore, manipulating the intensity of the noise context maps at inference time enables zero-shot conditional denoising, allowing for dynamic control over the output image texture. The entire pipeline is available at https://github.com/minwoo-yu/ENCORE.git

Figures

Figures reproduced from arXiv: 2608.10343 by the authors.

Figure 1
Figure 1. Performance comparison of PSNR against MACs (left) and latency [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Overall scheme of our proposed ENCORE framework for low-dose CT denoising. The ENCORE framework is built upon two core components: noise [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Qualitative comparison of each configuration with zoomed-in ROI patches. For Mayo2016 (first row) and Mayo2020 (second row), results are evaluated [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Result for sweeping dtarget values, evaluated on the Mayo2016 dataset at a 25% dose level with the Base UNet+ENCORE. (a) Image quality metrics (b) ROI patches. additional FBP executions and increases the computational overhead. Since no noticeable performance gain is o…
Figure 5
Figure 5. Figure 5: Benchmark comparison between customized CUDA kernels and the [PITH_FULL_IMAGE:figures/full_fig_p009_5.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

25 extracted references · 14 canonical work pages

  1. [1]

    Ct super-resolution gan constrained by the identical, residual, and cycle learning ensemble (gan-circle),

    C. You, G. Li, Y . Zhang, X. Zhang, H. Shan, M. Li, S. Ju, Z. Zhao, Z. Zhang, W. Conget al., “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, 2019

  2. [2]

    Low-dose ct denoising via convolutional neural network with an observer loss function,

    M. Han, H. Shim, and J. Baek, “Low-dose ct denoising via convolutional neural network with an observer loss function,”Medical physics, vol. 48, no. 10, pp. 5727–5742, 2021

  3. [3]

    Noise2noise: Learning image restoration without clean data,

    J. Lehtinen, J. Munkberg, J. Hasselgren, S. Laine, T. Karras, M. Aittala, and T. Aila, “Noise2noise: Learning image restoration without clean data,”arXiv preprint arXiv:1803.04189, 2018

  4. [4]

    Half2half: deep neural network based ct image denoising without independent reference data,

    N. Yuan, J. Zhou, and J. Qi, “Half2half: deep neural network based ct image denoising without independent reference data,”Physics in Medicine & Biology, vol. 65, no. 21, p. 215020, 2020

  5. [5]

    Simulating arbitrary dose levels and inde- pendent noise image pairs from a single ct scan,

    S. Wang and A. S. Wang, “Simulating arbitrary dose levels and inde- pendent noise image pairs from a single ct scan,” in7th International Conference on Image Formation in X-Ray Computed Tomography, vol. 12304. SPIE, 2022, pp. 460–466

  6. [6]

    Benchmarking denoising algorithms with real photographs,

    T. Plotz and S. Roth, “Benchmarking denoising algorithms with real photographs,” inProceedings of the IEEE conference on computer vision and pattern recognition, 2017, pp. 1586–1595

  7. [7]

    Noise-augmented deep denois- ing: A method to boost ct image denoising networks,

    G. Kristof, E. Eulig, and M. Kachelrieß, “Noise-augmented deep denois- ing: A method to boost ct image denoising networks,”Medical Physics, vol. 52, no. 10, p. e18121, 2025

  8. [8]

    Moments and cumulants in the speci- fication of distributions,

    E. A. Cornish and R. A. Fisher, “Moments and cumulants in the speci- fication of distributions,”Revue de l’Institut international de Statistique, pp. 307–320, 1938

Show all 25 references
  1. [9]

    The noise power spectrum in ct with direct fan beam reconstruction,

    J. Baek and N. J. Pelc, “The noise power spectrum in ct with direct fan beam reconstruction,”Medical physics, vol. 37, no. 5, pp. 2074–2081, 2010

  2. [10]

    Pixel-wise estimation of noise statistics on iterative ct reconstruction from a single scan,

    T. Wang and L. Zhu, “Pixel-wise estimation of noise statistics on iterative ct reconstruction from a single scan,”Medical physics, vol. 44, no. 7, pp. 3525–3533, 2017

  3. [11]

    May denoising remove structures? how to reconstruct invariances of ct denoising algorithms,

    E. Eulig, J. Maier, B. Ommer, and M. Kachelrieß, “May denoising remove structures? how to reconstruct invariances of ct denoising algorithms,” inMedical Imaging 2024: Physics of Medical Imaging, vol. 12925. SPIE, 2024, pp. 23–28

  4. [12]

    Bilateral filtering for gray and color images,

    C. Tomasi and R. Manduchi, “Bilateral filtering for gray and color images,” inSixth international conference on computer vision (IEEE Cat. No. 98CH36271). IEEE, 1998, pp. 839–846

  5. [13]

    Fast and high quality image denoising via malleable convolution,

    Y . Jiang, B. Wronski, B. Mildenhall, J. T. Barron, Z. Wang, and T. Xue, “Fast and high quality image denoising via malleable convolution,” in European Conference on Computer Vision. Springer, 2022, pp. 429– 446

  6. [14]

    Run, don’t walk: chasing higher flops for faster neural networks,

    J. Chen, S.-h. Kao, H. He, W. Zhuo, S. Wen, C.-H. Lee, and S.-H. G. Chan, “Run, don’t walk: chasing higher flops for faster neural networks,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2023, pp. 12 021–12 031

  7. [15]

    Differentiable forward projector for x-ray computed tomography,

    H. Kim and K. Champley, “Differentiable forward projector for x-ray computed tomography,”arXiv preprint arXiv:2307.05801, 2023

  8. [16]

    Tigre: a matlab- gpu toolbox for cbct image reconstruction,

    A. Biguri, M. Dosanjh, S. Hancock, and M. Soleimani, “Tigre: a matlab- gpu toolbox for cbct image reconstruction,”Biomedical Physics & Engineering Express, vol. 2, no. 5, p. 055010, 2016

  9. [17]

    Low-dose ct for the detection and classification of metastatic liver lesions: results of the 2016 low dose ct grand challenge,

    C. H. McCollough, A. C. Bartley, R. E. Carter, B. Chen, T. A. Drees, P. Edwards, D. R. Holmes III, A. E. Huang, F. Khan, S. Lenget al., “Low-dose ct for the detection and classification of metastatic liver lesions: results of the 2016 low dose ct grand challenge,”Medical physi...

  10. [18]

    Variance analysis of x-ray ct sinograms in the presence of electronic noise background,

    J. Ma, Z. Liang, Y . Fan, Y . Liu, J. Huang, W. Chen, and H. Lu, “Variance analysis of x-ray ct sinograms in the presence of electronic noise background,”Medical physics, vol. 39, no. 7Part1, pp. 4051–4065, 2012

  11. [19]

    Low-dose ct image and projec- tion dataset,

    T. R. Moen, B. Chen, D. R. Holmes, X. Duan, Z. Yu, L. Yu, S. Leng, J. G. Fletcher, and C. H. McCollough, “Low-dose ct image and projec- tion dataset,”Medical Physics, vol. 48, no. 2, pp. 902–911, Feb. 2021, epub 2020 Dec 16

  12. [20]

    Beyond a gaussian denoiser: Residual learning of deep cnn for image denoising,

    K. Zhang, W. Zuo, Y . Chen, D. Meng, and L. Zhang, “Beyond a gaussian denoiser: Residual learning of deep cnn for image denoising,”IEEE transactions on image processing, vol. 26, no. 7, pp. 3142–3155, 2017

  13. [21]

    U-net: Convolutional networks for biomedical image segmentation,

    O. Ronneberger, P. Fischer, and T. Brox, “U-net: Convolutional networks for biomedical image segmentation,” inInternational Conference on Medical image computing and computer-assisted intervention. Springer, 2015, pp. 234–241

  14. [22]

    Simple baselines for image restoration,

    L. Chen, X. Chu, X. Zhang, and J. Sun, “Simple baselines for image restoration,” inEuropean conference on computer vision. Springer, 2022, pp. 17–33

  15. [23]

    Uformer: A general u-shaped transformer for image restoration,

    Z. Wang, X. Cun, J. Bao, W. Zhou, J. Liu, and H. Li, “Uformer: A general u-shaped transformer for image restoration,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2022, pp. 17 683–17 693

  16. [24]

    sfrc for assessing hallucina- tions in medical image restoration,

    P. Kc, R. Zeng, N. Soni, and A. Badano, “sfrc for assessing hallucina- tions in medical image restoration,”arXiv preprint arXiv:2603.04673, 2026

  17. [25]

    Perceptual losses for real-time style transfer and super-resolution,

    J. Johnson, A. Alahi, and L. Fei-Fei, “Perceptual losses for real-time style transfer and super-resolution,” inEuropean conference on computer vision. Springer, 2016, pp. 694–711

Pith tools

Reviewed August 15, 2026 · model on record in the stance chip above.