Pith. sign in

REVIEW 2 major objections 2 minor 37 references

Unpaired Deep Learning Synthesis of Photon-Counting CT Material Basis Maps from Non-contrast Energy-Integrating Abdominal CT Images

T0 review · 2 major / 2 minor · reviewed 2026-06-25 · grok-4.3

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

Pith's one-line read A two-step deep learning pipeline can synthesize photon-counting CT material maps from unpaired standard EID CT scans.

desk verdict The paper outlines a two-step unpaired pipeline (DDIM then domain-adversarial U-Net) to make synthetic PCCT material maps from EID CT, but supplies no numbers to show the maps are accurate on real unpaired scans. read the letter →

arxiv 2606.24317 v1 pith:Y5EIVAOX submitted 2026-06-23 physics.med-ph

Ruihan Huang , Staffan Holmin , Mats Persson This is my paper · ORCID
classification physics.med-ph
keywords photon-countingCTmaterialbasismapsunpairedlearningDDIMdomain-adversarialU-Netvirtualmonoenergeticimagesenergy-integratingdetectorspectralsynthesis
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

Photon-counting CT provides better spatial resolution and spectral information than conventional energy-integrating detector CT, but public PCCT datasets are scarce because the technology is new. This paper presents a method that first uses a denoising diffusion implicit model to translate between PCCT and EID image domains, then trains a domain-adversarial U-Net on the generated EID images to predict water and iodine material maps. The resulting maps are used to reconstruct virtual monoenergetic images at 40 and 70 keV. Evaluation shows these images maintain anatomical detail from the original EID scans while gaining resolution, indicating the approach works without any paired PCCT-EID training examples.

What carries the argument

Two-step unpaired pipeline: DDIM generation of EID-style images from PCCT followed by domain-adversarial U-Net material decomposition into water and iodine maps.

What would settle it

Quantitative comparison on any set of real paired PCCT and EID scans showing large errors in Hounsfield unit values for the predicted water or iodine maps or failure to recover the claimed spatial-resolution gain in the 40 and 70 keV reconstructions.

Watch

Extended reading notes

Core claim

The proposed framework provides a feasible approach for synthesizing PCCT spectral material-basis images from conventional EID CT without requiring paired images by using DDIM to generate EID CT images from PCCT followed by a domain-adversarial U-Net to predict water and iodine maps, with reconstructed 40 and 70 keV images showing higher spatial resolution while preserving anatomical structures and textures of the original EID CT images.

Load-bearing premise

The distribution shift between DDIM-generated EID images and real EID images is small enough that a network trained on the generated images will still produce accurate material maps on real unpaired EID scans.

Editorial extensions

If this is right

  • Large synthetic PCCT material-map datasets become available for algorithm development without waiting for paired clinical acquisitions.
  • Reconstructed virtual monoenergetic images at 40 and 70 keV retain original EID textures while exhibiting measurably higher spatial resolution.
  • The method supplies training and evaluation data for PCCT-specific tasks in environments where real photon-counting scanners remain limited.
  • Water and iodine basis maps can be obtained directly from existing public EID CT collections.

Reading between the lines

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

  • The same unpaired translation strategy could be tested on other emerging CT modalities that also lack paired reference data.
  • Once real paired PCCT-EID cases appear in limited numbers they could serve as a direct validation set for the synthetic maps.
  • Downstream PCCT reconstruction algorithms could be pretrained on the generated material maps before fine-tuning on scarce real data.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

2 major / 2 minor

Summary. The manuscript presents a two-step unpaired deep learning pipeline to synthesize photon-counting CT (PCCT) material-basis maps (water and iodine) from energy-integrating detector (EID) CT images. A DDIM first generates synthetic EID images from PCCT data; a domain-adversarial U-Net then predicts the material maps from these generated EID images. Virtual monoenergetic images (VMIs) at 40 and 70 keV are reconstructed from the predicted maps and evaluated via HU accuracy, MTF, NPS, and qualitative appearance, with the central claim that the resulting VMIs exhibit higher spatial resolution while preserving anatomical structures and textures of the original EID images.

Significance. If the generalization to real unpaired EID data holds, the method could enable creation of large synthetic PCCT training sets from existing EID archives, supporting algorithm development where paired PCCT-EID data are unavailable. The two-stage design (diffusion-based domain translation followed by adversarial material decomposition) directly targets the unpaired-data constraint.

major comments (2)
  1. [Results/Evaluation] Results section (and abstract): All reported quantitative metrics (HU accuracy, MTF, NPS) and the qualitative preservation claim are demonstrated exclusively on DDIM-generated EID images paired with PCCT-derived material maps. No quantitative or qualitative results are supplied for the U-Net+DANN applied to real unpaired EID CT scans, so the load-bearing assumption that the domain-adversarial loss closes the gap between synthetic and real EID distributions remains untested on the target domain.
  2. [Methods] Methods (second-stage network): The domain-adversarial loss is described as mitigating the domain shift, yet no ablation is presented that isolates its contribution (e.g., U-Net trained with vs. without the adversarial term) or that quantifies residual mismatch in noise texture or beam-hardening between DDIM outputs and real EID images; without such evidence the feasibility claim for real EID inputs rests on an unverified assumption.
minor comments (2)
  1. [Abstract] Abstract: 'clinc' is a typographical error for 'clinic'.
  2. [Abstract] Abstract: The evaluation metrics are listed but no numerical values, confidence intervals, or statistical comparisons are supplied, reducing the abstract's informativeness.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive and detailed comments. We address each major point below and indicate where revisions will be made to strengthen the manuscript.

read point-by-point responses

  1. Referee: [Results/Evaluation] Results section (and abstract): All reported quantitative metrics (HU accuracy, MTF, NPS) and the qualitative preservation claim are demonstrated exclusively on DDIM-generated EID images paired with PCCT-derived material maps. No quantitative or qualitative results are supplied for the U-Net+DANN applied to real unpaired EID CT scans, so the load-bearing assumption that the domain-adversarial loss closes the gap between synthetic and real EID distributions remains untested on the target domain.

    Authors: We agree that the reported quantitative metrics rely on DDIM-generated EID images paired with PCCT-derived material maps as ground truth, which is feasible only in this controlled setting. Real unpaired EID scans lack corresponding material maps, precluding the same quantitative evaluation. The domain-adversarial loss is intended to bridge the distribution gap, but we acknowledge that direct evidence on real EID inputs would better support the generalization claim. In the revised manuscript we will add qualitative examples of material-basis maps and VMIs produced from real unpaired EID CT scans to illustrate applicability to the target domain. revision: yes

  2. Referee: [Methods] Methods (second-stage network): The domain-adversarial loss is described as mitigating the domain shift, yet no ablation is presented that isolates its contribution (e.g., U-Net trained with vs. without the adversarial term) or that quantifies residual mismatch in noise texture or beam-hardening between DDIM outputs and real EID images; without such evidence the feasibility claim for real EID inputs rests on an unverified assumption.

    Authors: We agree that an ablation isolating the domain-adversarial term and additional quantification of residual domain mismatch would strengthen the methods section. In the revised manuscript we will include an ablation study comparing the U-Net trained with and without the adversarial loss, together with analysis of noise texture and beam-hardening characteristics between DDIM outputs and real EID images. revision: yes

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: empirical pipeline with independent evaluation metrics

full rationale

The paper describes a two-stage unpaired synthesis pipeline (DDIM for EID generation from PCCT followed by adversarial U-Net for material map prediction) whose central feasibility claim rests on post-hoc image-quality metrics (HU accuracy, MTF, NPS) and qualitative preservation of anatomical structures. These metrics are standard, externally defined image properties computed on output volumes and are not algebraically or statistically forced by the training losses or fitted parameters. No self-citations appear in the provided text to justify uniqueness theorems or ansatzes, and no step renames a fitted quantity as a prediction or reduces the reported result to an input definition by construction. The derivation chain is therefore self-contained against external benchmarks.

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

The central claim rests on standard deep-learning assumptions about domain adaptation and image synthesis rather than new physical principles; model weights constitute the primary fitted elements.

free parameters (1)
  • DDIM and U-Net training hyperparameters
    Learning rates, network depths, and adversarial loss weights are optimized on the generated data to achieve the reported image quality.
assumptions (1)
  • domain assumption Adversarial training can align feature distributions between DDIM-generated EID images and real EID images sufficiently for accurate material decomposition
    Invoked to justify applying the trained U-Net to real EID inputs.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Unpaired Deep Learning Synthesis of Photon-Counting CT Material Basis Maps from Non-contrast Energy-Integrating Abdominal CT Images." pith.science (2026). https://pith.science/paper/Y5EIVAOX

@misc{pith2026260624317,
  author       = {Pith},
  title        = {Pith review of: Unpaired Deep Learning Synthesis of Photon-Counting CT Material Basis Maps from Non-contrast Energy-Integrating Abdominal CT Images},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/Y5EIVAOX}},
  note         = {Machine review of arXiv:2606.24317}
}
read the original abstract

Photon-counting Computed Tomography (PCCT) is the most advanced Computed Tomography (CT) technology, offering significant improvements in image quality and diagnostic capabilities. However, since PCCT has only recently been adopted in the clinc, there are no publicly available PCCT image datasets for study. We therefore aim to synthesize PCCT spectral material-basis images from publicly available EID CT images. We propose a two-step deep learning model designed to synthesize photon-counting spectral material basis images from public Energy-Integrating Detector (EID) CT images. In the first step, we use a Denoising Diffusion Implicit Model (DDIM) to generate EID CT images from PCCT images. In the second step we use a U-Net with a Domain-Adversarial Neural Network to predict water and iodine maps from generated EID CT images. We also reconstruct basis images and virtual monoenergetic images (VMIs) from the predicted material-basis maps for evaluation. We evaluated the generated water and iodine maps as well as the 40 and 70 keV PCCT images in terms of Hounsfield Unit accuracy, modulation transfer function and noise power spectrum as well as qualitative image appearance. The reconstructed 40 and 70 keV PCCT images exhibit higher spatial resolution while preserving the anatomical structures and textures of the original EID CT images, thereby demonstrating the feasibility of the proposed approach. The proposed framework provides a feasible approach for synthesizing PCCT spectral material-basis images from conventional EID CT without requiring paired images. This method has the potential to provide large sets of synthetic training and evaluation data for PCCT algorithm development in data-limited environments.

Figures

Figures reproduced from arXiv: 2606.24317 by the authors.

Figure 1
Figure 1. Example of generated EID CT images by Step 1. The images at the first row are generated [PITH_FULL_IMAGE:figures/full_fig_p016_1.png] view at source ↗
Figure 2
Figure 2. Comparison of the mean NPS measured in matched ROIs for EID CT, PCCT, and the [PITH_FULL_IMAGE:figures/full_fig_p017_2.png] view at source ↗
Figure 3
Figure 3. Example results from the U-Net model. The left column shows predictions obtained from [PITH_FULL_IMAGE:figures/full_fig_p020_3.png] view at source ↗
Figures from the paper (5 more)
Figure 4
Figure 4. Figure 4: Example of 40 and 70 keV VMIs from validation dataset. (a) 40 and 70 keV VMIs of [PITH_FULL_IMAGE:figures/full_fig_p022_4.png]
Figure 5
Figure 5. Figure 5: Representative image showing the three ROIs used for HU bias assessment. [PITH_FULL_IMAGE:figures/full_fig_p022_5.png]
Figure 6
Figure 6. Figure 6: Examples of results on real EID CT images. (a) Predicted 40 and 70 keV VMIs and original [PITH_FULL_IMAGE:figures/full_fig_p024_6.png]
Figure 7
Figure 7. Figure 7: Representative image showing the three ROIs used for HU bias assessment. [PITH_FULL_IMAGE:figures/full_fig_p025_7.png]
Figure 8
Figure 8. Figure 8: MTF analysis based on three representative regions of interest (ROIs). ROIs are shown in [PITH_FULL_IMAGE:figures/full_fig_p025_8.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

37 extracted references · 3 canonical work pages

  1. [1]

    Initial clinical images from a second-generation prototype silicon-based photon-counting computed tomography system,

    H. Almqvist, D. Crotty, S. Nyren,et al., “Initial clinical images from a second-generation prototype silicon-based photon-counting computed tomography system,”Academic Radiol- ogy31, 572–581 (2024). Epub 2023-08-08

    2024

  2. [2]

    Photon-counting x-ray detectors for ct,

    M. Danielsson, M. Persson, and M. Sj ¨olin, “Photon-counting x-ray detectors for ct,”Physics in Medicine and Biology66, 03TR01 (2021)

    2021

  3. [3]

    Catsim: a new computer assisted tomography simulation environment,

    B. De Man, S. Basu, N. Chandra,et al., “Catsim: a new computer assisted tomography simulation environment,” inMedical Imaging 2007: Physics of Medical Imaging, J. Hsieh and M. J. Flynn, Eds.,Proceedings of SPIE6510, 65102G, SPIE (2007)

    2007

  4. [4]

    Deep-learning-based spectral motion artifact correc- tion on photon-counting cardiac ct images,

    R. Huang, K. Larsson, D. Hein,et al., “Deep-learning-based spectral motion artifact correc- tion on photon-counting cardiac ct images,”Physics in Medicine & Biology(2026). Epub ahead of print. PMID: 41604762. 28

    2026

  5. [5]

    Systematic review of synthetic computed tomography generation methodologies for use in magnetic resonance imaging-only radia- tion therapy,

    E. Johnstone, J. J. Wyatt, A. M. Henry,et al., “Systematic review of synthetic computed tomography generation methodologies for use in magnetic resonance imaging-only radia- tion therapy,”International Journal of Radiation Oncology, Biology, Physics100, 199–217 (2018). Epub 2017-09-08

    2018

  6. [6]

    Medical image synthesis with context-aware generative adversarial networks,

    D. Nie, R. Trullo, J. Lian,et al., “Medical image synthesis with context-aware generative adversarial networks,” inMedical Image Computing and Computer Assisted Intervention – MICCAI 2017, M. Descoteaux, L. Maier-Hein, A. Franz,et al., Eds.,Lecture Notes in Com- puter Science10435, 417–425, Springer, Cham (2017)

    2017

  7. [7]

    Deep mr to ct synthesis using unpaired data,

    J. M. Wolterink, A. M. Dinkla, M. H. F. Savenije,et al., “Deep mr to ct synthesis using unpaired data,” inSimulation and Synthesis in Medical Imaging, S. A. Tsaftaris, A. Gooya, A. F. Frangi,et al., Eds.,Lecture Notes in Computer Science10557, 14–23, Springer, Cham (2017)

    2017

  8. [8]

    Conversion between ct and mri images using diffusion and score- matching models,

    Q. Lyu and G. Wang, “Conversion between ct and mri images using diffusion and score- matching models,”arXiv preprint arXiv:2209.12104(2022)

    work page arXiv 2022

Show all 37 references
  1. [9]

    Generative adversarial network in medical imaging: A re- view,

    X. Yi, E. Walia, and P. Babyn, “Generative adversarial network in medical imaging: A re- view,”Medical Image Analysis58, 101552 (2019). Epub 2019-08-31

    2019

  2. [10]

    Synthetic ct generation from mri using 3d transformer- based denoising diffusion model,

    S. Pan, E. Abouei, J. Wynne,et al., “Synthetic ct generation from mri using 3d transformer- based denoising diffusion model,”Medical Physics51, 2538–2548 (2024). Epub 2023-11-27

    2024

  3. [11]

    Synthesis of virtual monoenergetic images from kilovoltage peak images using wavelet loss enhanced cyclegan for improving radiomics features reproducibil- ity,

    Z. Xu, M. Li, B. Li,et al., “Synthesis of virtual monoenergetic images from kilovoltage peak images using wavelet loss enhanced cyclegan for improving radiomics features reproducibil- ity,”Quantitative Imaging in Medicine and Surgery14(3), 2370–2390 (2024). 29

    2024

  4. [12]

    Virtual monoenergetic ct imaging via deep learning,

    W. Cong, Y . Xi, P. Fitzgerald,et al., “Virtual monoenergetic ct imaging via deep learning,” Patterns1, 100128 (2020)

    2020

  5. [13]

    Image synthesis of monoenergetic ct image in dual-energy ct using kilovoltage ct with deep convolutional generative adversarial networks,

    D. Kawahara, S. Ozawa, T. Kimura,et al., “Image synthesis of monoenergetic ct image in dual-energy ct using kilovoltage ct with deep convolutional generative adversarial networks,” Journal of Applied Clinical Medical Physics22, 184–192 (2021). Epub 2021-02-18

    2021

  6. [14]

    Estimating dual-energy ct imaging from single-energy ct data with material decomposition convolutional neural network,

    T. Lyu, W. Zhao, Y . Zhu,et al., “Estimating dual-energy ct imaging from single-energy ct data with material decomposition convolutional neural network,”Medical Image Analysis 70, 102001 (2021)

    2021

  7. [15]

    A quality-checked and physics-constrained deep learning method to estimate material basis images from single-kv contrast-enhanced chest ct scans,

    Y . Li, X. Tie, K. Li,et al., “A quality-checked and physics-constrained deep learning method to estimate material basis images from single-kv contrast-enhanced chest ct scans,”Medical Physics50(6), 3368–3388 (2023)

    2023

  8. [16]

    Image quality assessment of deep learning-based vir- tual monoenergetic images from single-energy ct pulmonary angiography,

    K. Li, P. Nagpal, B. F. Mullan,et al., “Image quality assessment of deep learning-based vir- tual monoenergetic images from single-energy ct pulmonary angiography,”Journal of Com- puter Assisted Tomography50(2), 263–271 (2026)

    2026

  9. [17]

    Distilling photon-counting ct into routine chest ct through clinically validated degradation modeling,

    J. Liu, X. Zhou, W. Li,et al., “Distilling photon-counting ct into routine chest ct through clinically validated degradation modeling,”arXiv preprint arXiv:2604.07329(2026)

    2026 arXiv

  10. [18]

    Data from lung ct segmentation challenge 2017 (lctsc)

    J. Yang, G. Sharp, H. Veeraraghavan,et al., “Data from lung ct segmentation challenge 2017 (lctsc).” The Cancer Imaging Archive (TCIA) Dataset (2017). Accessed May 22, 2026

    2017

  11. [19]

    Initial clinical images from a second-generation prototype silicon-based photon-counting computed tomography system,

    H. Almqvist, D. Crotty, S. Nyren,et al., “Initial clinical images from a second-generation prototype silicon-based photon-counting computed tomography system,”Academic Radiol- ogy31, 572–581 (2024). Epub 2023-08-08. 30

    2024

  12. [20]

    MATLAB version R2023b

    The MathWorks Inc., “MATLAB version R2023b.”https://www.mathworks.com/ products/matlab.html(2023). Natick, Massachusetts, United States

    2023

  13. [21]

    imregtform: Estimate geometric transformation that aligns two 2-d or 3-d images

    The MathWorks, Inc., “imregtform: Estimate geometric transformation that aligns two 2-d or 3-d images.” MATLAB Help Center, Image Processing Toolbox documentation page (2026). Accessed February 27, 2026

    2026

  14. [22]

    Denoising diffusion implicit models

    J. Song, C. Meng, and S. Ermon, “Denoising diffusion implicit models.” https://arxiv.org/abs/2010.02502 (2022)

    2010 arXiv

  15. [23]

    Denoising diffusion probabilistic models,

    J. Ho, A. Jain, and P. Abbeel, “Denoising diffusion probabilistic models,” inAdvances in Neural Information Processing Systems, H. Larochelle, M. Ranzato, R. Hadsell,et al., Eds., 33, 6840–6851, Curran Associates, Inc. (2020)

    2020

  16. [24]

    U-Net: Convolutional Networks for Biomedical Image Segmentation,

    O. Ronneberger, P. Fischer, and T. Brox, “U-Net: Convolutional Networks for Biomedical Image Segmentation,” inMedical Image Computing and Computer-Assisted Intervention – MICCAI 2015,Lecture Notes in Computer Science9351, 234–241, Springer (2015)

    2015

  17. [25]

    Domain-adversarial training of neural networks,

    Y . Ganin, E. Ustinova, H. Ajakan,et al., “Domain-adversarial training of neural networks,” Journal of machine learning research17(59), 1–35 (2016)

    2016

  18. [26]

    Very deep convolutional networks for large-scale image recognition,

    K. Simonyan and A. Zisserman, “Very deep convolutional networks for large-scale image recognition,” inInternational Conference on Learning Representations, International Con- ference on Learning Representations (2015)

    2015

  19. [27]

    Imagenet: A large-scale hierarchical image database,

    J. Deng, W. Dong, R. Socher,et al., “Imagenet: A large-scale hierarchical image database,” in2009 IEEE Computer Society Conference on Computer Vision and Pattern Recognition 31 (CVPR 2009), 20–25 June 2009, Miami, Florida, USA, 248–255, IEEE Computer Society (2009)

    2009

  20. [28]

    Goodfellow, Y

    I. Goodfellow, Y . Bengio, and A. Courville,Deep Learning, MIT Press (2016)

    2016

  21. [29]

    Pytorch: An imperative style, high-performance deep learning library,

    A. Paszke, S. Gross, F. Massa,et al., “Pytorch: An imperative style, high-performance deep learning library,” inAdvances in Neural Information Processing Systems,32, 8024–8035 (2019)

    2019

  22. [30]

    Adam: A method for stochastic optimization,

    D. P. Kingma and J. Ba, “Adam: A method for stochastic optimization,” inInternational Conference on Learning Representations (ICLR), (2015)

    2015

  23. [31]

    The noise power spectrum in computed x-ray tomography,

    S. J. Riederer, N. J. Pelc, and D. A. Chesler, “The noise power spectrum in computed x-ray tomography,”Physics in Medicine and Biology23(3), 446–454 (1978)

    1978

  24. [32]

    Scope of validity of psnr in image/video quality assess- ment,

    Q. Huynh-Thu and M. Ghanbari, “Scope of validity of psnr in image/video quality assess- ment,”Electronics Letters44(13), 800–801 (2008)

    2008

  25. [33]

    Image quality assessment: From error visibility to structural similarity,

    Z. Wang, A. C. Bovik, H. R. Sheikh,et al., “Image quality assessment: From error visibility to structural similarity,”IEEE Transactions on Image Processing13(4), 600–612 (2004)

    2004

  26. [34]

    The unreasonable effectiveness of deep features as a perceptual metric,

    R. Zhang, P. Isola, A. A. Efros,et al., “The unreasonable effectiveness of deep features as a perceptual metric,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 586–595 (2018)

    2018

  27. [35]

    Imagenet classification with deep convolu- tional neural networks,

    A. Krizhevsky, I. Sutskever, and G. E. Hinton, “Imagenet classification with deep convolu- tional neural networks,” inAdvances in Neural Information Processing Systems 25 (NeurIPS 2012), P. L. Bartlett, F. C. N. Pereira, C. J. C. Burges,et al., Eds., 1106–1114 (2012). 32

    2012

  28. [36]

    Mind: Modality independent neigh- bourhood descriptor for multi-modal deformable registration,

    M. P. Heinrich, M. Jenkinson, M. Bhushan,et al., “Mind: Modality independent neigh- bourhood descriptor for multi-modal deformable registration,”Medical Image Analysis16, 1423–1435 (2012). Epub 2012-05-31

    2012

  29. [37]

    A method for measuring the presampled mtf of digital radiographic systems using an edge test device,

    E. Samei, M. J. Flynn, and D. A. Reimann, “A method for measuring the presampled mtf of digital radiographic systems using an edge test device,”Medical Physics25(1), 102–113 (1998). List of Figures 1 Example of generated EID CT images by Step 1. The images at the first row are...

    1998

Pith tools

Reviewed June 25, 2026 · model on record in the stance chip above.