pith. sign in

arxiv: 2604.16925 · v3 · pith:7FWJ2TCPnew · submitted 2026-04-18 · 💻 cs.CV

Rethinking Cross-Dose PET Denoising: Mitigating Averaging Effects via Residual Noise Learning

Yichao Liu , Zongru Shao , Yueyang Teng , Junwen Guo This is my paper

Pith reviewed 2026-05-19 16:47 UTC · model grok-4.3

classification 💻 cs.CV
keywords cross-dose PET denoisingresidual noise learninglow-dose imagingnoise estimationdeep learningmedical image denoisinggeneralization in denoisingaveraging effect
0
0 comments X

The pith

Estimating noise directly from low-dose PET images avoids the averaging effect in cross-dose denoising models.

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

The paper seeks to show that the poor generalization of PET denoising models across different radiation dose levels stems from them learning an average of the denoising functions for each noise level. Instead of having the network predict the full high-quality image, the new framework has it predict only the noise pattern present in the low-dose scan so that subtracting it yields the denoised result. This is claimed to prevent the network from settling on a compromise mapping that works okay for all doses but not well for any. Experiments on big datasets collected at two hospitals show the new method works better than training one model for all doses, better than training separate models for each dose, and better than models that take dose level as input. Readers might care because it offers a single model that handles the range of doses used in practice without performance loss.

Core claim

Standard training formulations for cross-dose denoising implicitly optimize an expectation over heterogeneous noise distributions, causing the network to learn an averaged denoising mapping that cannot accurately model dose-specific noise characteristics. The proposed unified residual noise learning framework estimates noise directly from low-dose PET images rather than predicting full-dose images. Experiments on large-scale multi-dose PET datasets from two medical centers demonstrate that this approach outperforms the one-size-for-all model, individual dose-specific U-Net models, and dose-conditioned approaches.

What carries the argument

Residual noise learning framework, which has the network predict the additive noise residual from the low-dose PET input for direct subtraction.

If this is right

  • A single model trained on mixed-dose data can match or exceed the performance of separate dose-specific models.
  • The network avoids learning compromise mappings that degrade results at every individual dose level.
  • Generalization across dose conditions improves without requiring dose level as an explicit input.
  • Denoising performance gains hold on large datasets collected at multiple medical centers.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The same residual prediction idea could simplify workflows in other modalities where noise strength changes with scan parameters.
  • Clinics could maintain one denoising network instead of retraining or storing versions for each common dose protocol.
  • If residuals prove stable enough, they might combine with physics-based reconstruction steps to further reduce artifacts.

Load-bearing premise

The noise component in low-dose PET can be treated as an additive residual whose statistical properties are sufficiently independent of the underlying anatomy that a network can learn to predict it directly from the noisy input alone.

What would settle it

A test on held-out multi-dose PET scans showing whether subtracting the predicted residual from low-dose images consistently yields lower error to full-dose ground truth than baseline methods across varied anatomies and dose levels; lack of consistent improvement would indicate the averaging problem persists.

Figures

Figures reproduced from arXiv: 2604.16925 by Junwen Guo, Yichao Liu, YueYang Teng, Zongru Shao.

Figure 1
Figure 1. Figure 1: Noise analysis for 1/100 noise level and 1/2 noise level LDPET images. Noise image is calculated by the difference between FDPET and LDPET. [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: The overview of the training network. The model takes a LDPET image as input and predicts the underlying noise. To prevent the loss of negative [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Quantitative comparison of PSNR, SSIM and RMSE of standard deviation on two datasets, University of Bern (left) and Shanghai Ruijing hospital [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Qualitative comparison of denoised upper abdomen using different methods for 1/50 noise level and 1/2 noise level LDPET images. The sample is [PITH_FULL_IMAGE:figures/full_fig_p006_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Qualitative comparison of denoised lower abdomen using different methods for 1/50 noise level and 1/2 noise level LDPET images. The sample is [PITH_FULL_IMAGE:figures/full_fig_p007_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Comparison of denoised lower abdomen using different methods for 1/50 noise level and 1/2 noise level LDPET images. The sample is selected [PITH_FULL_IMAGE:figures/full_fig_p008_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Qualitative comparison of whole-body PET denoising using different methods at [PITH_FULL_IMAGE:figures/full_fig_p010_7.png] view at source ↗
read the original abstract

Cross-dose denoising for low-dose positron emission tomography (LDPET) has been proposed to address the limited generalization of models trained at a single noise level. However, neural networks trained on a specific dose level often fail to generalize to other dose conditions due to variations in noise magnitude and statistical properties. Conventional "one-size-for-all" models attempt to mitigate this variability but tend to learn averaged representations across noise levels, resulting in degraded performance. In this work, we analyze this limitation and show that standard training formulations implicitly optimize an expectation over heterogeneous noise distributions, causing the network to learn an averaged denoising mapping that cannot accurately model dose-specific noise characteristics. We propose a unified residual noise learning framework that estimates noise directly from low-dose PET images rather than predicting full-dose images. Experiments on large-scale multi-dose PET datasets from two medical centers demonstrate that the proposed method outperforms the "one-size-for-all" model, individual dose-specific U-Net models, and dose-conditioned approaches, achieving improved denoising performance. These results indicate that residual noise learning effectively mitigates the averaging effect and enhances generalization for cross-dose PET denoising.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 2 minor

Summary. The manuscript analyzes limitations of one-size-for-all cross-dose PET denoising models, which implicitly optimize over heterogeneous noise distributions and learn averaged mappings. It proposes a residual noise learning framework that directly estimates the additive noise residual from low-dose PET images rather than predicting full-dose images, and reports superior performance over one-size-for-all U-Nets, dose-specific models, and dose-conditioned baselines on large-scale multi-dose datasets from two medical centers.

Significance. If the residual noise learning approach holds, it offers a practical route to improved generalization across dose levels without training separate models or conditioning on dose, which is relevant for clinical low-dose PET where radiation exposure must be minimized while preserving diagnostic quality.

major comments (2)
  1. [Framework / residual noise learning] Framework section (residual noise learning): The central premise that the difference between low-dose and full-dose images forms an additive residual whose statistics are sufficiently independent of the underlying activity map is not accompanied by a derivation or empirical check. Given that PET noise is Poisson (variance proportional to local mean count), the residual variance remains modulated by anatomy; this dependence risks reintroducing the dose-anatomy coupling the method claims to bypass.
  2. [Experiments] Experiments and results: The abstract states outperformance on multi-center data, yet the manuscript provides no quantitative tables, loss formulation details, or statistical significance tests (e.g., paired t-tests or confidence intervals on PSNR/SSIM differences). Without these, it is impossible to verify whether reported gains are robust or sensitive to post-hoc hyperparameter choices.
minor comments (2)
  1. [Method] Notation for the residual term is introduced without an explicit equation linking it to the Poisson noise model; adding a short derivation or reference to the noise statistics would improve clarity.
  2. [Figures] Figure captions for qualitative results should explicitly state the dose levels shown and the metrics reported in each panel.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We are grateful to the referee for the constructive and detailed feedback. We address each major comment below and describe the revisions planned for the manuscript.

read point-by-point responses

  1. Referee: [Framework / residual noise learning] Framework section (residual noise learning): The central premise that the difference between low-dose and full-dose images forms an additive residual whose statistics are sufficiently independent of the underlying activity map is not accompanied by a derivation or empirical check. Given that PET noise is Poisson (variance proportional to local mean count), the residual variance remains modulated by anatomy; this dependence risks reintroducing the dose-anatomy coupling the method claims to bypass.

    Authors: We thank the referee for this observation. The additive residual is indeed an approximation, and Poisson statistics imply intensity-dependent variance. Nevertheless, predicting the residual rather than the clean image allows the network to focus on dose-specific noise patterns and avoids the averaged mappings learned by standard cross-dose models. We will revise the Framework section to include a brief derivation under the Poisson model and add empirical analysis (e.g., residual variance stratified by local intensity and anatomy) to quantify the practical utility of the approach despite the theoretical dependence. revision: yes

  2. Referee: [Experiments] Experiments and results: The abstract states outperformance on multi-center data, yet the manuscript provides no quantitative tables, loss formulation details, or statistical significance tests (e.g., paired t-tests or confidence intervals on PSNR/SSIM differences). Without these, it is impossible to verify whether reported gains are robust or sensitive to post-hoc hyperparameter choices.

    Authors: We appreciate the referee highlighting the need for clearer presentation. While performance is discussed in the text, dedicated tables, explicit loss details, and statistical tests were not included. We will add comprehensive tables reporting PSNR and SSIM for all baselines on both centers, state the loss function (L1 on the residual) in the Methods, and include paired t-tests with p-values and 95% confidence intervals on the metric differences to demonstrate robustness. revision: yes

Circularity Check

0 steps flagged

Empirical framework with no derivation chain circularity

full rationale

The paper proposes a residual noise learning framework for cross-dose PET denoising and validates it through experiments on multi-center datasets. No mathematical derivation, first-principles result, or prediction is presented that reduces by construction to fitted parameters, self-citations, or inputs defined within the paper. The central claim rests on empirical performance comparisons rather than any closed-form equivalence or load-bearing self-referential step, satisfying the criteria for a self-contained empirical contribution.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

The paper introduces no new physical entities or untested axioms. It relies on standard assumptions that PET noise is approximately additive and that a convolutional network can learn a residual mapping. No free parameters are explicitly fitted beyond ordinary network training.

axioms (1)
  • domain assumption Noise in low-dose PET can be modeled as an additive residual whose statistics are learnable directly from the noisy image.
    This premise is invoked when the authors shift the learning target from full-dose image to noise residual (abstract).

pith-pipeline@v0.9.0 · 5731 in / 1374 out tokens · 37717 ms · 2026-05-19T16:47:52.275789+00:00 · methodology

Review history (2 revisions) →

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

What do these tags mean?
matches
The paper's claim is directly supported by a theorem in the formal canon.
supports
The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends
The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses
The paper appears to rely on the theorem as machinery.
contradicts
The paper's claim conflicts with a theorem or certificate in the canon.
unclear
Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

32 extracted references · 32 canonical work pages

  1. [1]

    Positron-emission tomography and assessment of cancer therapy

    M. Juweid and B. Cheson, “Positron-emission tomography and assessment of cancer therapy.”The New England journal of medicine, vol. 354 5, pp. 496–507, 2006. [Online]. Available: https://www.ncbi.nlm.nih.gov/pubmed/16452561

    work page arXiv 2006

  2. [2]

    A systematic review of pet and pet/ct in oncology: A way to personalize cancer treatment in a cost-effective manner?

    A. Langer, “A systematic review of pet and pet/ct in oncology: A way to personalize cancer treatment in a cost-effective manner?”BMC Health Services Research, vol. 10, pp. 283 – 283, 2010. [Online]. Available: https://api.semanticscholar.org/CorpusId:16124679 JOURNAL OF TRANSACTIONS ON RADIATION AND PLASMA MEDICAL SCIENCES, VOL. 14, NO. 8, AUGUST 2021 8 F...

    work page 2010

  3. [3]

    The role of radiology in head and neck tumours in children,

    C. Lloyd and K. McHugh, “The role of radiology in head and neck tumours in children,”Cancer Imaging, vol. 10, pp. 49 – 61, 2010. [Online]. Available: https://api.semanticscholar.org/CorpusId:16451286

    work page 2010

  4. [4]

    Present and future roles of fdg-pet/ct imaging in the management of gastrointestinal cancer: an update,

    K. Kitajima, M. Nakajo, H. Kaida, R. Minamimoto, K. Hirata, M. Tsurusaki, H. Doi, Y . Ueno, K. Sofue, Y . Tamaki, and K. Yamakado, “Present and future roles of fdg-pet/ct imaging in the management of gastrointestinal cancer: an update,”Nagoya Journal of Medical Science, vol. 79, pp. 527 – 543, 2017. [Online]. Available: https://api.semanticscholar.org/Cor...

    work page 2017

  5. [5]

    Nsclc biomarkers to predict response to immunotherapy with checkpoint inhibitors (ici): From the cells to in vivo images,

    V . Liberini, A. Mariniello, L. Righi, M. Capozza, M. D. Delcuratolo, E. Terreno, M. Farsad, M. V olante, S. Novello, and D. Deandreis, “Nsclc biomarkers to predict response to immunotherapy with checkpoint inhibitors (ici): From the cells to in vivo images,”Cancers, vol. 13, 2021. [Online]. Available: https://api.semanticscholar.org/CorpusId:237212796

    work page 2021

  6. [6]

    Pet-ct in clinical adult oncology—iv. gynecologic and genitourinary malignancies,

    A. E. Salem, G. Fine, M. F. Covington, B. Koppula, R. Wiggins, J. Hoffman, and K. A. Morton, “Pet-ct in clinical adult oncology—iv. gynecologic and genitourinary malignancies,”Cancers, vol. 14, 2022. [Online]. Available: https://api.semanticscholar.org/CorpusId:249860940

    work page 2022

  7. [7]

    State of the art in total body pet,

    S. Vandenberghe, P. Moskal, and J. Karp, “State of the art in total body pet,”EJNMMI Physics, vol. 7, 2020. [Online]. Available: https://api.semanticscholar.org/CorpusId:218877646

    work page 2020

  8. [8]

    Machine learning in quantitative pet imaging,

    T. Wang, Y . Lei, Y . Fu, W. Curran, T. xing Liu, and X. Yang, “Machine learning in quantitative pet imaging,”ArXiv, vol. abs/2001.06597, 2020. [Online]. Available: https://api.semanticscholar.org/CorpusId:210839460

    work page arXiv 2001

  9. [9]

    Deeptofsino: A deep learning model for synthesizing full-dose time-of-flight bin sinograms from their corresponding low- dose sinograms,

    A. Sanaat, H. Shooli, S. Ferdowsi, I. Shiri, H. Arabi, and H. Zaidi, “Deeptofsino: A deep learning model for synthesizing full-dose time-of-flight bin sinograms from their corresponding low- dose sinograms,”NeuroImage, p. 118697, 2021. [Online]. Available: https://api.semanticscholar.org/CorpusId:241905571

    work page 2021

  10. [10]

    Sharpness-aware low-dose ct denoising using conditional generative adversarial network,

    X. Yi and P. Babyn, “Sharpness-aware low-dose ct denoising using conditional generative adversarial network,”Journal of Digital Imaging, vol. 31, pp. 655–669, 2017. [Online]. Available: https: //api.semanticscholar.org/CorpusId:3397048

    work page 2017

  11. [11]

    Mr-guided dynamic pet reconstruction with the kernel method and spectral temporal basis functions,

    P. Novosad and A. J. Reader, “Mr-guided dynamic pet reconstruction with the kernel method and spectral temporal basis functions,”Physics in Medicine & Biology, vol. 61, no. 12, pp. 4624–4644, 2016

    work page 2016

  12. [12]

    Partial volume correction of brain pet studies using iterative deconvolution in combination with hypr denoising,

    S. S. Golla, M. Lubberink, B. N. van Berckel, A. A. Lammertsma, and R. Boellaard, “Partial volume correction of brain pet studies using iterative deconvolution in combination with hypr denoising,”EJNMMI research, vol. 7, no. 1, p. 36, 2017

    work page 2017

  13. [13]

    Maximum likelihood reconstruction for emission tomography,

    L. A. Shepp and Y . Vardi, “Maximum likelihood reconstruction for emission tomography,”IEEE transactions on medical imaging, vol. 1, no. 2, pp. 113–122, 1982

    work page 1982

  14. [14]

    Iterative reconstruc- tion of simulated low count data: a comparison of post-filtering versus regularised osem,

    K. Karaoglanis, N. Efthimiou, and C. Tsoumpas, “Iterative reconstruc- tion of simulated low count data: a comparison of post-filtering versus regularised osem,” inJournal of Physics: Conference Series, vol. 637, no. 1. IOP Publishing, 2015, p. 012006

    work page 2015

  15. [15]

    Pet image reconstruction in- corporating deep image prior and a forward projection model,

    F. Hashimoto, K. Ote, and Y . Onishi, “Pet image reconstruction in- corporating deep image prior and a forward projection model,”IEEE Transactions on Radiation and Plasma Medical Sciences, vol. 6, no. 8, pp. 841–846, 2022

    work page 2022

  16. [16]

    Full-count pet recovery from low-count image using a dilated convolu- tional neural network,

    K. Spuhler, M. Serrano-Sosa, R. Cattell, C. DeLorenzo, and C. Huang, “Full-count pet recovery from low-count image using a dilated convolu- tional neural network,”Medical Physics, vol. 47, no. 10, pp. 4928–4938, 2020

    work page 2020

  17. [17]

    Deep auto-context convolutional neural networks for standard-dose pet image estimation from low-dose pet/mri,

    L. Xiang, Y . Qiao, D. Nie, L. An, W. Lin, Q. Wang, and D. Shen, “Deep auto-context convolutional neural networks for standard-dose pet image estimation from low-dose pet/mri,”Neurocomputing, vol. 267, pp. 406–416, 2017

    work page 2017

  18. [18]

    Dynamic pet image denoising with deep learning-based joint filtering,

    Y . He, S. Cao, H. Zhang, H. Sun, F. Wang, H. Zhu, W. Lv, and L. Lu, “Dynamic pet image denoising with deep learning-based joint filtering,” IEEE Access, vol. 9, pp. 41 998–42 012, 2021

    work page 2021

  19. [19]

    Leqmod: adaptable lesion-quantification- consistent modulation for deep learning low-count pet image denoising,

    M. Xia, H. Xie, Q. Liu, B. Zhou, H. Wang, B. Li, A. Rominger, Q. Li, R. D. Badawi, K. Shiet al., “Leqmod: adaptable lesion-quantification- consistent modulation for deep learning low-count pet image denoising,” IEEE Transactions on Medical Imaging, 2025

    work page 2025

  20. [20]

    Ultra- low-dose pet reconstruction using generative adversarial network with feature matching and task-specific perceptual loss,

    J. Ouyang, K. T. Chen, E. Gong, J. Pauly, and G. Zaharchuk, “Ultra- low-dose pet reconstruction using generative adversarial network with feature matching and task-specific perceptual loss,”Medical physics, vol. 46, no. 8, pp. 3555–3564, 2019

    work page 2019

  21. [21]

    Su- pervised learning with cyclegan for low-dose fdg pet image denoising,

    L. Zhou, J. D. Schaefferkoetter, I. W. Tham, G. Huang, and J. Yan, “Su- pervised learning with cyclegan for low-dose fdg pet image denoising,” Medical image analysis, vol. 65, p. 101770, 2020

    work page 2020

  22. [22]

    S. Xue, R. Guo, K. P. Bohn, J. Matzke, M. Viscione, I. Alberts, H. Meng, C. Sun, M. Zhang, M. Zhanget al., “A cross-scanner and cross- tracer deep learning method for the recovery of standard-dose imaging JOURNAL OF TRANSACTIONS ON RADIATION AND PLASMA MEDICAL SCIENCES, VOL. 14, NO. 8, AUGUST 2021 9 quality from low-dose pet,”European journal of nuclear m...

    work page 2021

  23. [23]

    Unified noise-aware network for low-count pet denoising with varying count levels,

    H. Xie, Q. Liu, B. Zhou, X. Chen, X. Guo, H. Wang, B. Li, A. Rominger, K. Shi, and C. Liu, “Unified noise-aware network for low-count pet denoising with varying count levels,”IEEE transactions on radiation and plasma medical sciences, vol. 8, no. 4, pp. 366–378, 2023

    work page 2023

  24. [24]

    A personalized deep learning denoising strategy for low-count pet images,

    Q. Liu, H. Liu, N. Mirian, S. Ren, V . Viswanath, J. Karp, S. Surti, and C. Liu, “A personalized deep learning denoising strategy for low-count pet images,”Physics in Medicine & Biology, vol. 67, no. 14, p. 145014, 2022

    work page 2022

  25. [25]

    Deep supervised transformer-based noise-aware network for low-dose pet denoising across varying count levels,

    M. S. Azimi, V . Felfelian, N. Zeraatkar, H. Dadgar, H. Arabi, and H. Zaidi, “Deep supervised transformer-based noise-aware network for low-dose pet denoising across varying count levels,”Computers in biology and medicine, vol. 196, p. 110733, 2025

    work page 2025

  26. [26]

    Cross noise level pet denoising with continuous adversarial domain generalization,

    X. Liu, S. Vafay Eslahi, T. Marin, A. Tiss, Y . Chemli, Y . Huang, K. A. Johnson, G. El Fakhri, and J. Ouyang, “Cross noise level pet denoising with continuous adversarial domain generalization,”Physics in Medicine & Biology, vol. 69, no. 8, p. 085001, 2024

    work page 2024

  27. [27]

    Domain adaptation for medical image analysis: a survey,

    H. Guan and M. Liu, “Domain adaptation for medical image analysis: a survey,”IEEE Transactions on Biomedical Engineering, vol. 69, no. 3, pp. 1173–1185, 2021

    work page 2021

  28. [28]

    Domain general- ization: A survey,

    K. Zhou, Z. Liu, Y . Qiao, T. Xiang, and C. C. Loy, “Domain general- ization: A survey,”IEEE transactions on pattern analysis and machine intelligence, vol. 45, no. 4, pp. 4396–4415, 2022

    work page 2022

  29. [29]

    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

    work page 2017

  30. [30]

    A stochastic approximation method,

    H. Robbins and S. Monro, “A stochastic approximation method,”The annals of mathematical statistics, pp. 400–407, 1951

    work page 1951

  31. [31]

    Swin-unet: Unet-like pure transformer for medical image segmenta- tion,

    H. Cao, Y . Wang, J. Chen, D. Jiang, X. Zhang, Q. Tian, and M. Wang, “Swin-unet: Unet-like pure transformer for medical image segmenta- tion,” inEuropean conference on computer vision. Springer, 2022, pp. 205–218

    work page 2022

  32. [32]

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

    H. Chen, Y . Zhang, M. K. Kalra, F. Lin, Y . Chen, P. Liao, J. Zhou, and G. Wang, “Low-dose ct with a residual encoder-decoder convolutional neural network,”IEEE transactions on medical imaging, vol. 36, no. 12, pp. 2524–2535, 2017. JOURNAL OF TRANSACTIONS ON RADIATION AND PLASMA MEDICAL SCIENCES, VOL. 14, NO. 8, AUGUST 2021 10 Fig. 7. Qualitative compari...

    work page 2017