Pith. sign in

REVIEW 3 major objections 6 minor 57 references

Multipath cycleGAN for harmonization of paired and unpaired low-dose lung computed tomography reconstruction kernels

T0 review · 3 major / 6 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 A shared-latent multipath cycleGAN harmonizes CT reconstruction kernels across vendors, reducing emphysema quantification bias.

desk verdict A genuine two-stage extension of the authors' own multipath cycleGAN with a careful held-out evaluation, but the abstract overclaims: Table 3 still shows significant kernel effects for Philips C, Philips D, and GE LUNG after harmonization. read the letter →

arxiv 2505.22568 v1 pith:SJTWDR4O submitted 2025-05-28 eess.IV cs.CV

Aravind R. Krishnan , Thomas Z. Li , Lucas W. Remedios , Michael E. Kim , Chenyu Gao , Gaurav Rudravaram , Elyssa M. McMaster , Adam M. Saunders
show 7 more authors
Shunxing Bao Kaiwen Xu Lianrui Zuo Kim L. Sandler Fabien Maldonado Yuankai Huo Bennett A. Landman
This is my paper · ORCID
classification eess.IVcs.CV
keywords multipathcycleGANsharedlatentspacepercentemphysemaeffectsizereconstructionkernelharmonizationlow-doseCTNLST
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 tries to show that one image-translation network, a multipath cycleGAN whose encoders and decoders all write into and read from a single shared latent space, can replace per-pair cycleGANs for harmonizing CT reconstruction kernels from Siemens, GE, and Philips scanners in a low-dose lung cancer screening cohort. If true, multi-centre and longitudinal studies could compare emphysema measurements across vendors without training a separate model for every kernel pair, and kernel-induced bias in percent emphysema could be largely removed. The evidence is that on paired kernels, harmonization reduced the median RMSE in percent emphysema with p<0.05, while on unpaired kernels, harmonization to a reference soft kernel made the kernel effect statistically non-significant for several kernels in a general linear model. Dice overlap of segmentations suggested muscle and fat anatomy were mostly preserved, though lung-vessel overlap was more modest.

What carries the argument

The load-bearing object is the multipath cycleGAN: for each domain, a ResNet encoder maps images into a shared latent space of size N×256×128×128, and a domain-specific decoder reads that space to produce the target kernel style, with a PatchGAN discriminator per domain deciding real-versus-synthetic. A cycle-consistency loss and an identity loss computed on downsampled images, so the loss is insensitive to kernel texture, keep content and radio-opacity stable. Two-stage training freezes the stage-one encoders and decoders when new kernels are added, so the new domains adapt to the existing shared representation rather than learning separate pairwise models.

What would settle it

Segment skeletal muscle, subcutaneous adipose tissue, and lung vessels manually, or with a kernel-agnostic tool, on the same subjects' original and harmonized GE BONE and GE STANDARD images; if the overlap is substantially lower than the reported TotalSegmentator Dice, or if chest-wall muscle appears relabelled as lung in the harmonized images, then the anatomical-consistency claim is refuted. A second check would be to recompute the general linear model on harmonized Philips D images, where the paper already finds p<0.05, and show whether the residual bias exceeds a clinically meaningful emphysema threshold such as one absolute percentage point.

Watch

Extended reading notes

Core claim

The central claim is that paired and unpaired kernel harmonization with a shared latent space multipath cycleGAN mitigates errors in emphysema quantification and maintains anatomical consistency after harmonization. The network represents every kernel as its own encoder-decoder pair, but all pairs meet in one shared, high-dimensional latent space, so a single model trained in two stages (four kernels first, then three more with frozen stage-one weights) covers all 42 kernel combinations. On 240 withheld scans per kernel, harmonizing hard kernels to a soft reference reduced Bland-Altman bias and RMSE for paired kernels, and harmonizing all kernels to the Siemens B30f reference brought most emphysema medians close to the reference; the general linear model showed several kernels losing statistical significance after harmonization. Anatomical consistency was assessed by Dice overlap between TotalSegmentator segmentations before and after harmonization, with high Dice for skeletal muscle and subcutaneous adipose tissue and reasonable overlap for lung vessels.

Load-bearing premise

The anatomical-consistency claim rests on the assumption that Dice overlap between TotalSegmentator segmentations of the original and harmonized images measures true anatomical preservation, yet the authors note that for GE BONE and GE STANDARD, skeletal muscle and subcutaneous fat are sometimes modified as lung tissue, so the metric itself can miss hallucinations.

Editorial extensions

If this is right

  • A single shared-latent model can serve many kernel pairs, so multi-centre studies do not need to train and maintain one cycleGAN per vendor pair.
  • Harmonizing all kernels to one reference soft kernel reduces kernel-related confounding in emphysema, making longitudinal and cross-vendor comparisons more trustworthy.
  • The two-stage training scheme suggests new vendor kernels can be added by freezing old weights and training only the new encoders, decoders, and discriminators.
  • Residual kernel effects remain for Philips D, Philips C, and GE LUNG, so harmonization to a single reference does not yet fully erase vendor-specific bias.

Reading between the lines

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

  • Because the identity and cycle losses are defined on texture-invariant downsampled images, the same architecture could plausibly be extended to other kernel-sensitive quantitative biomarkers such as coronary calcium scoring or body composition, provided the anatomical-consistency limitations are addressed first.
  • The reported residual bias on Philips C and D suggests that a shared latent space alone may not suffice for all vendors; a testable fix would be to add a per-domain adversarial or feature-alignment term in the latent space.
  • The anatomical-consistency claim is only as strong as TotalSegmentator's reliability on hard-kernel images; an external validation with manual or multi-tool segmentations on GE BONE and GE STANDARD harmonized images would directly test it.
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

3 major / 6 minor

Summary. The paper proposes a multipath cycleGAN with domain-specific encoders/decoders and a shared latent space, trained in two stages, to harmonize CT reconstruction kernels from Siemens, GE, and Philips in the NLST cohort. The model is evaluated on paired and unpaired kernel pairs by measuring percent emphysema before and after harmonization, by fitting a general linear model for kernel effects, and by computing TotalSegmentator Dice scores for lung vessels, muscle, and subcutaneous adipose tissue. The authors compare against per-path cycleGAN and switchable cycleGAN baselines and conclude that the shared latent space multipath cycleGAN mitigates emphysema quantification errors and preserves anatomical consistency.

Significance. If the claims were fully supported, the main contribution would be useful: a single network that harmonizes many paired and unpaired kernel combinations, avoiding per-pair models, with a large withheld test set (240 scans per kernel) and an independent statistical evaluation. The two-stage training, direct comparison to two baselines, and downstream quantification of emphysema are strengths. However, the central claim as stated in the abstract, that unpaired harmonization 'eliminates confounding differences,' is contradicted by the paper's own Table 3, and the anatomical-consistency evidence is weakened by the authors' admitted segmentation failures on GE BONE and GE STANDARD. The contribution is still defensible if reframed as partial mitigation of kernel effects, but the current overstatement is a load-bearing issue.

major comments (3)
  1. [Abstract and Section 4.5, Table 3] The abstract states that unpaired kernel harmonization 'eliminates confounding differences in emphysema (p>0.05),' but Table 3 shows that after multipath harmonization the Philips C (coefficient 2.20, p<0.05), Philips D (coefficient 2.94, p<0.05), and GE LUNG (coefficient -2.26, p<0.05) kernel effects remain statistically significant. Thus 3 of 6 kernel categories still show detectable kernel effects, and the claim of elimination is not supported. Furthermore, p>0.05 is not evidence of equivalence; no equivalence margin, confidence interval on the kernel coefficients, or two one-sided test (TOST) is reported. Please either add an equivalence analysis with a prespecified margin or revise the abstract and conclusion to claim only partial mitigation, explicitly reporting which kernels remain significant.
  2. [Section 3.4, Table 4, and Section 5] The anatomical consistency claim for GE BONE and GE STANDARD is not supported by the Dice scores as reported. The authors acknowledge in Section 5 that for these kernels 'skeletal muscle around the chest wall and SAT are modified as lung tissues' and that body composition measurements were therefore not evaluated. Since TotalSegmentator's own segmentations fail on exactly the structures being measured, high Dice overlap between segmentations of the original and harmonized images does not establish anatomical preservation for these kernels. The manuscript should either exclude GE BONE and GE STANDARD from the anatomical-consistency conclusion or provide an independent validation that the segmentation failures are not confounding the Dice comparison.
  3. [Section 3.3] The statistical methods for unpaired harmonization are underspecified. The text says a Mann-Whitney U test is used to 'assess statistical significance between emphysema measurements before and after harmonization,' but it does not state which distributions are compared (e.g., source kernel vs. reference kernel before harmonization, and harmonized kernel vs. reference after harmonization). Without a clear statement of the null hypothesis and the reported p-values, the reader cannot tell what 'p>0.05' in the abstract refers to. Please specify the exact comparisons, report the corresponding test statistics, and provide confidence intervals or effect sizes for the harmonization-induced changes.
minor comments (6)
  1. [Section 4, paragraph 1] The text refers to 'the selection criteria described in Section 2.7,' but the epoch selection procedure is actually described in Section 3.2; please correct the cross-reference.
  2. [Section 3.3, Eq. (3)] Equation (3) in Section 3.3 is numbered the same as Eq. (3) in Section 3.2; renumber the general linear model equation to avoid confusion.
  3. [Table 2 and Section 5, paired kernels paragraph] For the (C, D) pair, the multipath cycleGAN RMSE is 2.47 [2.11, 2.93], substantially larger than the cycleGAN RMSE of 1.31 [1.18, 1.49]; the claim that the proposed approach shows 'comparable performance' on paired kernels should be softened or discussed for this pair.
  4. [Section 5, paragraph on anatomical consistency] The limitation statement refers to hallucinations for GE BONE and GE STANDARD 'as evidenced in Figure 8,' but Figure 8 shows soft-to-hard kernel harmonization, not the unpaired harmonicization to Siemens B30f; please clarify which figure demonstrates the chest-wall and SAT hallucinations.
  5. [Section 2.4] The paper repeatedly distinguishes paired and unpaired data, but the described training losses (adversarial, cycle-consistency, identity) do not appear to use the one-to-one pixel correspondence of paired kernels in a pixel-wise supervision term. Please clarify whether paired correspondence is used anywhere in the loss, or whether 'paired' refers only to the evaluation setting.
  6. [Section 2.1 and Table 1] GE STANDARD appears twice in Table 1 (once paired with BONE and once with LUNG); please clarify whether the 100 training volumes for each GE STANDARD row are disjoint and whether the withheld 240 test volumes can contain the same patient from both paired sets, which would affect the independence of the test evaluation.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity; the central claim is supported by held-out empirical evaluation against external baselines, not by construction or self-citation.

full rationale

Inspection of the derivation chain finds no step where a claimed prediction is equivalent to its inputs by definition. The central claim is empirical: a multipath cycleGAN with a shared latent space is trained on 100 NLST scans per kernel, selected by validation performance using emphysema-based MSE/KL criteria (Eqs. 1-4), and then evaluated on 240 withheld test scans per kernel (Sections 2.1, 3.2, 4). The emphysema endpoint is a fixed -950 HU threshold (Section 3.1), not a parameter fitted to the model's outputs. Baselines (traditional cycleGAN and switchable cycleGAN) are independently trained and compared on the same held-out data, so the multipath model's performance is not forced by construction. Author self-citations (refs. 25 and 30) provide the prior architecture and data QA procedure, but the paper's new conclusions rest on the held-out comparisons, not on those citations alone. The statistical criticism that p>0.05 does not prove equivalence is a correctness/interpretation concern, not circularity, and the acknowledged TotalSegmentator limitations on GE BONE/STANDARD (Section 5) are validity risks, not input-output equivalence. No fitted parameter is relabeled as a prediction, and no equation reduces to its own assumption.

Assumptions & free parameters 6 free parameters · 5 assumptions · 0 invented entities

The central claims rest on CT domain assumptions (paired correspondence, -950 HU threshold, standardization of FOV), the validity of TotalSegmentator Dice as a fidelity measure, and several hand-chosen hyperparameters, including the validation-based epoch-selection weights. These are reasonable for an empirical ML paper but should be disclosed as presumptions.

free parameters (6)
  • Epoch-selection weights = 0.5/0.25/0.25 (Stage 1), 0.2 each (Stage 2)
    Hand-chosen balance between paired MSE and unpaired KL divergence in Eqs (1)-(2); directly determines which checkpoint is evaluated.
  • lambda_identity schedule = 1e6 (epoch 1), decayed by /100, fixed at 0.01 from epoch 6
    Hand-chosen annealing of the identity loss in Stage one (Section 2.4).
  • lambda_cycle = 10
    Cycle-consistency weight inherited from default cycleGAN (Section 2.4).
  • Learning rate and epochs = 0.002, 120-200 epochs Stage 1; 0.002, 200 epochs Stage 2
    Chosen by the authors; no ablation reported.
  • Identity loss downsampling factor = 256x256 bilinear
    Downsampling is said to smooth the kernel and preserve radio-opacity; the factor is hand-chosen (Section 2.3).
  • Stage-two subsampling ratio = 20% of images per epoch
    Selected to make training feasible on two A6000 GPUs (Section 2.4).
assumptions (5)
  • domain assumption Paired kernel scans from the same vendor have one-to-one pixel correspondence after manual QA of spatial alignment
    Used to define paired data and to compare emphysema measurements point-wise (Section 2.1).
  • domain assumption Percent emphysema is validly measured as the fraction of lung voxels below -950 HU
    Threshold fixed at -950 HU (Section 3.1), following Gevenois et al.; the harmonization target is defined by this biomarker.
  • domain assumption TotalSegmentator segmentations are equally accurate on original and harmonized images
    Dice overlap between segmentations is the measure of anatomical consistency (Section 3.4); the authors admit hallucinations on some GE kernels, undercutting this premise.
  • standard math Cycle-consistency and adversarial losses preserve content while changing style
    The cycleGAN framework (Section 2.3) assumes that reconstructing the source image from the translated image enforces content preservation.
  • domain assumption Circular masking standardizes the field of view without removing relevant anatomy
    Siemens and Philips images are masked to a circular 512x512 FOV to match GE (Section 2.2); anatomy outside the mask is discarded.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Multipath cycleGAN for harmonization of paired and unpaired low-dose lung computed tomography reconstruction kernels." pith.science (2026). https://pith.science/paper/SJTWDR4O

@misc{pith2026250522568,
  author       = {Pith},
  title        = {Pith review of: Multipath cycleGAN for harmonization of paired and unpaired low-dose lung computed tomography reconstruction kernels},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/SJTWDR4O}},
  note         = {Machine review of arXiv:2505.22568}
}
read the original abstract

Reconstruction kernels in computed tomography (CT) affect spatial resolution and noise characteristics, introducing systematic variability in quantitative imaging measurements such as emphysema quantification. Choosing an appropriate kernel is therefore essential for consistent quantitative analysis. We propose a multipath cycleGAN model for CT kernel harmonization, trained on a mixture of paired and unpaired data from a low-dose lung cancer screening cohort. The model features domain-specific encoders and decoders with a shared latent space and uses discriminators tailored for each domain.We train the model on 42 kernel combinations using 100 scans each from seven representative kernels in the National Lung Screening Trial (NLST) dataset. To evaluate performance, 240 scans from each kernel are harmonized to a reference soft kernel, and emphysema is quantified before and after harmonization. A general linear model assesses the impact of age, sex, smoking status, and kernel on emphysema. We also evaluate harmonization from soft kernels to a reference hard kernel. To assess anatomical consistency, we compare segmentations of lung vessels, muscle, and subcutaneous adipose tissue generated by TotalSegmentator between harmonized and original images. Our model is benchmarked against traditional and switchable cycleGANs. For paired kernels, our approach reduces bias in emphysema scores, as seen in Bland-Altman plots (p<0.05). For unpaired kernels, harmonization eliminates confounding differences in emphysema (p>0.05). High Dice scores confirm preservation of muscle and fat anatomy, while lung vessel overlap remains reasonable. Overall, our shared latent space multipath cycleGAN enables robust harmonization across paired and unpaired CT kernels, improving emphysema quantification and preserving anatomical fidelity.

Figures

Figures reproduced from arXiv: 2505.22568 by the authors.

Figure 1
Figure 1. Reconstruction kernels influence the noise and resolution of the underlying anatomical structure in a computed tomography image. (a) Paired reconstruction kernels obtained from a given vendor exhibit a one-to￾one pixel correspondence between the scans which enables kernel harmonization. However, (b) across vendors, unpaired kernels show differences in anatomy, scan protocol, field of view and reconstruction window. … view at source ↗
Figure 2
Figure 2. We hypothesize that inter-vendor and intra-vendor harmonization across different combinations of reconstruction kernels can be performed using a multipath cycleGAN model, trained in two stages. The paired reconstruction kernels are grouped together based on the manufacturer while all other combinations denote unpaired kernels. In Stage one, we harmonize across all possible combinations of four different reconstructi… view at source ↗
Figure 3
Figure 3. For any given pair of reconstruction kernels, there exists a forward and backward path. The generator is a ResNet, formed by a source encoder and target decoder in the forward path and a target encoder and source decoder in the backward path. Each generator produces a synthetic image with the style of either the source or target kernel. A PatchGAN is used as a discriminator for the corresponding domain to distinguis… view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: Bland Altman style plots are used to represent the agreement on emphysema measurements between paired hard and soft reconstruction kernels before and after harmonization. The solid line represents the mean difference between the emphysema measurements and the dashed li…
Figure 6
Figure 6. Figure 6: Paired kernels exhibit a one-to-one pixel correspondence between the hard and soft kernel in each vendor with differences in the pixel noise. Hard kernels sharpen the image while soft kernels smoothen the image. Harmonizing to the corresponding soft kernel enforces qua…
Figure 7
Figure 7. Figure 7: Variation in scanner protocol introduces differences in the texture of the lung parenchyma, thereby introducing differences in quantitative image measures. Harmonization of all kernels to the reference B30f soft kernel enforces consistency in the texture of the lung pa…
Figure 8
Figure 8. Figure 8: Harmonizing a soft kernel to hard kernel is a challenging task since it is difficult to recover high frequency components from a blurry image. We observe that harmonizing all soft kernels to a reference hard kernel restores the high frequency details in the region of i…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

57 extracted references · 45 canonical work pages

  1. [1]

    Spatial domain filtering for fast modification of the tradeoff between image sharpness 23 and pixel noise in computed tomography

    Schaller S, Wildberger JE, Raupach R, Niethammer M, Klingenbeck-Regn K, Flohr T. Spatial domain filtering for fast modification of the tradeoff between image sharpness 23 and pixel noise in computed tomography. IEEE Trans Med Imaging. 2003;22(7):846-853. doi:10.1109/TMI.2003.815073

    work page arXiv 2003

  2. [2]

    CT Scan Transformation from a Sharp to a Soft Reconstruction Kernel Using Filtering Techniques

    Lasek J, Piórkowski A. CT Scan Transformation from a Sharp to a Soft Reconstruction Kernel Using Filtering Techniques. In: Communications in Computer and Information Science. Vol 1376 CCIS. Springer Science and Business Media Deutschland GmbH; 2021:56-65. doi:10.1007/978-981-16-1086-8_6

    work page doi:10.1007/978-981-16-1086-8_6 2021

  3. [3]

    Matching and Homogenizing Convolution Kernels for Quantitative Studies in Computed Tomography

    Mackin D, Ger R, Gay S, et al. Matching and Homogenizing Convolution Kernels for Quantitative Studies in Computed Tomography. Invest Radiol. 2019;54(5):288-295. doi:10.1097/RLI.0000000000000540

    work page doi:10.1097/rli.0000000000000540 2019

  4. [4]

    Emphysema: Effect of reconstruction algorithm on CT imaging measures

    Boedeker KL, McNitt-Gray MF , Rogers SR, et al. Emphysema: Effect of reconstruction algorithm on CT imaging measures. Radiology. 2004;232(1):295-301. doi:10.1148/radiol.2321030383

    work page doi:10.1148/radiol.2321030383 2004

  5. [5]

    Effects of CT Section Thickness and Reconstruction Kernel on Emphysema Quantification

    Gierada DS, Bierhals AJ, Choong CK, et al. Effects of CT Section Thickness and Reconstruction Kernel on Emphysema Quantification. Relationship to the Magnitude of the CT Emphysema Index. Acad Radiol. 2010;17(2):146-156. doi:10.1016/j.acra.2009.08.007

    work page doi:10.1016/j.acra.2009.08.007 2010

  6. [6]

    Computed Tomography-based Body Composition Analysis and Its Role in Lung Cancer Care

    Troschel AS, Troschel FM, Best TD, et al. Computed Tomography-based Body Composition Analysis and Its Role in Lung Cancer Care. J Thorac Imaging. 2020;35(2):91-

    work page 2020

  7. [7]

    Automatic Calcium Scoring in Low-Dose Chest CT Using Deep Neural Networks with Dilated Convolutions

    Lessmann N, Van Ginneken B, Zreik M, et al. Automatic Calcium Scoring in Low-Dose Chest CT Using Deep Neural Networks with Dilated Convolutions. IEEE Trans Med Imaging. 2018;37(2):615-625. doi:10.1109/TMI.2017.2769839

    work page arXiv 2018

  8. [8]

    Reproducibility of CT radiomic features within the same patient: Influence of radiation dose and CT reconstruction settings

    Meyer M, Ronald J, Vernuccio F , et al. Reproducibility of CT radiomic features within the same patient: Influence of radiation dose and CT reconstruction settings. Radiology. 2019;293(3):583-591. doi:10.1148/radiol.2019190928

    work page doi:10.1148/radiol.2019190928 2019

Show all 57 references
  1. [9]

    Impact of CT Convolution Kernel on Robustness of Radiomic Features for Different Lung Diseases and Tissue Types.; 2021

    Denzler S, Vuong D, Bogowicz M, et al. Impact of CT Convolution Kernel on Robustness of Radiomic Features for Different Lung Diseases and Tissue Types.; 2021

    2021

  2. [10]

    Image filtering as an alternative to the application of a different reconstruction kernel in CT imaging: Feasibility study in lung cancer screening

    Ohkubo M, Wada S, Kayugawa A, Matsumoto T, Murao K. Image filtering as an alternative to the application of a different reconstruction kernel in CT imaging: Feasibility study in lung cancer screening. Med Phys. 2011;38(7):3915-3923. doi:10.1118/1.3590363

    2011 doi

  3. [11]

    CT-HARMONICA: physics-based CT harmonization for reliable lung density quantification

    Sotoudeh-Paima S, Samei E, Abadi E. CT-HARMONICA: physics-based CT harmonization for reliable lung density quantification. In: Iftekharuddin KM, Chen W, eds. Medical Imaging 2023: Computer-Aided Diagnosis. SPIE; 2023:52. doi:10.1117/12.2654346

    2023 doi

  4. [12]

    Automated technique to measure noise in clinical CT examinations

    Christianson O, Winslow J, Frush DP , Samei E. Automated technique to measure noise in clinical CT examinations. American Journal of Roentgenology. 2015;205(1):W93-W99. doi:10.2214/AJR.14.13613

    2015 doi

  5. [13]

    Harmonizing CT images via physics-based deep neural networks

    Zarei M, Sotoudeh-Paima S, McCabe C, Abadi E, Samei E. Harmonizing CT images via physics-based deep neural networks. In: Fahrig R, Sabol JM, Yu L, eds. Medical Imaging 2023: Physics of Medical Imaging. SPIE; 2023:66. doi:10.1117/12.2654215 24

    2023 doi

  6. [14]

    A Physics-informed Deep Neural Network for Harmonization of CT Images

    Zarei M, Paima SS, McCabe C, Abadi E, Samei E. A Physics-informed Deep Neural Network for Harmonization of CT Images. IEEE Trans Biomed Eng. Published online 2024. doi:10.1109/TBME.2024.3428399

    2024

  7. [15]

    CT image conversion among different reconstruction Kernels without a sinogram by using a convolutional neural network

    Lee SM, Lee JG, Lee G, et al. CT image conversion among different reconstruction Kernels without a sinogram by using a convolutional neural network. Korean J Radiol. 2019;20(2):295-303. doi:10.3348/kjr.2018.0249

    2019

  8. [16]

    Accurate image super-resolution using very deep convolutional networks

    Kim J, Lee JK, Lee KM. Accurate image super-resolution using very deep convolutional networks. In: Proceedings of the IEEE Computer Society Conference on Computer Vision and Pattern Recognition. Vol 2016-December. IEEE Computer Society; 2016:1646-1654. doi:10.1109/CVPR.2016.182

    2016 doi

  9. [17]

    CT kernel conversions using convolutional neural net for super-resolution with simplified squeeze-and-excitation blocks and progressive learning among smooth and sharp kernels

    Eun D in, Woo I, Park B, Kim N, Lee A SM, Seo JB. CT kernel conversions using convolutional neural net for super-resolution with simplified squeeze-and-excitation blocks and progressive learning among smooth and sharp kernels. Comput Methods Programs Biomed. 2020;196. doi:10.1...

    2020

  10. [18]

    Deep Learning–based Image Conversion of CT Reconstruction Kernels Improves Radiomics Reproducibility for Pulmonary Nodules or Masses

    Choe J, Lee SM, Do KH, et al. Deep Learning–based Image Conversion of CT Reconstruction Kernels Improves Radiomics Reproducibility for Pulmonary Nodules or Masses. Radiology. 2019;292(2):365-373. doi:10.1148/radiol.2019181960

    2019 doi

  11. [19]

    Deep learning-enabled accurate normalization of reconstruction kernel effects on emphysema quantification in low-dose CT

    Jin H, Heo C, Kim JH. Deep learning-enabled accurate normalization of reconstruction kernel effects on emphysema quantification in low-dose CT. Phys Med Biol. 2019;64(13). doi:10.1088/1361-6560/ab28a1

    2019 doi

  12. [20]

    Image-to-Image Translation: Methods and Applications

    Pang Y , Lin J, Qin T, Chen Z. Image-to-Image Translation: Methods and Applications. IEEE Trans Multimedia. 2022;24:3859-3881. doi:10.1109/TMM.2021.3109419

    2022

  13. [21]

    Image-to-image translation with conditional adversarial networks

    Isola P , Zhu JY , Zhou T, Efros AA. Image-to-image translation with conditional adversarial networks. In: Proceedings - 30th IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2017. Vol 2017-January. Institute of Electrical and Electronics Engineers Inc.; 2017:5...

    2017 doi

  14. [22]

    Unpaired Image-to-Image Translation Using Cycle- Consistent Adversarial Networks

    Zhu JY , Park T, Isola P , Efros AA. Unpaired Image-to-Image Translation Using Cycle- Consistent Adversarial Networks. In: Proceedings of the IEEE International Conference on Computer Vision. Vol 2017-October. Institute of Electrical and Electronics Engineers Inc.; 2017:2242-2...

    2017 doi

  15. [23]

    CT Image Harmonization for Enhancing Radiomics Studies

    Selim M, Zhang J, Fei B, Zhang GQ, Chen J. CT Image Harmonization for Enhancing Radiomics Studies. Published online July 3, 2021. http://arxiv.org/abs/2107.01337

    2021 arXiv

  16. [24]

    Kernel Conversion for Robust Quantitative Measurements of Archived Chest Computed Tomography Using Deep Learning-Based Image-to-Image Translation

    Tanabe N, Kaji S, Shima H, et al. Kernel Conversion for Robust Quantitative Measurements of Archived Chest Computed Tomography Using Deep Learning-Based Image-to-Image Translation. Front Artif Intell. 2022;4. doi:10.3389/frai.2021.769557

    2022

  17. [25]

    Lung CT harmonization of paired reconstruction kernel images using generative adversarial networks

    Krishnan AR, Xu K, Li TZ, et al. Lung CT harmonization of paired reconstruction kernel images using generative adversarial networks. Med Phys. Published online March 26,

  18. [26]

    Continuous Conversion of CT Kernel Using Switchable CycleGAN with AdaIN

    Yang S, Kim EY , Ye JC. Continuous Conversion of CT Kernel Using Switchable CycleGAN with AdaIN. IEEE Trans Med Imaging. 2021;40(11):3015-3029. doi:10.1109/TMI.2021.3077615 25

    2021

  19. [27]

    Cross-Vendor CT Image Data Harmonization Using CVH-CT

    Selim M, Zhang J, Fei B, Zhang GQ, Ge GY , Chen J. Cross-Vendor CT Image Data Harmonization Using CVH-CT. Published online October 18, 2021. http://arxiv.org/abs/2110.09693

    2021 arXiv

  20. [28]

    Leveraging CycleGAN in Lung CT Sinogram-free Kernel Conversion

    Gravina M, Marrone S, Docimo L, et al. Leveraging CycleGAN in Lung CT Sinogram-free Kernel Conversion. In: Lecture Notes in Computer Science (Including Subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics). Vol 13231 LNCS. Springer Science and...

    2022 doi

  21. [29]

    TotalSegmentator: Robust Segmentation of 104 Anatomic Structures in CT Images • Content codes

    Wasserthal J, Breit • Hanns-Christian, Meyer MT, et al. TotalSegmentator: Robust Segmentation of 104 Anatomic Structures in CT Images • Content codes. Radiol Artif Intell. 2023;5(5). doi:10.5281/zenodo.6802613

    2023 doi

  22. [30]

    Inter-vendor harmonization of CT reconstruction kernels using unpaired image translation

    Krishnan AR, Xu K, Li TZ, et al. Inter-vendor harmonization of CT reconstruction kernels using unpaired image translation. In: SPIE-Intl Soc Optical Eng; 2024:45. doi:10.1117/12.3006608

    2024 doi

  23. [31]

    New England Journal of Medicine

    Reduced Lung-Cancer Mortality with Low-Dose Computed Tomographic Screening. New England Journal of Medicine. 2011;365(5):395-409. doi:10.1056/NEJMoa1102873

    2011 doi

  24. [32]

    Description and implementation of a quality control program in an imaging-based clinical trial

    Cagnon CH, Cody DD, McNitt-Gray MF , Seibert JA, Judy PF , Aberle DR. Description and implementation of a quality control program in an imaging-based clinical trial. Acad Radiol. 2006;13(11):1431-1441. doi:10.1016/j.acra.2006.08.015

    2006 doi

  25. [33]

    CT quality assurance in the lung screening study component of the national lung screening trial: Implications for multicenter imaging trials

    Gierada DS, Garg K, Nath H, Strollo DC, Fagerstrom RM, Ford MB. CT quality assurance in the lung screening study component of the national lung screening trial: Implications for multicenter imaging trials. American Journal of Roentgenology. 2009;193(2):419-424. doi:10.2214/AJR.08.1995

    2009 doi

  26. [34]

    The National Lung Screening Trial: Overview and Study Design

    National Lung Screening Trial Research Team. The National Lung Screening Trial: Overview and Study Design. Radiology. 2011;258 (1):243-253. doi:10.1148/radiol.10091808/-/DC1

    2011 doi

  27. [35]

    The first step for neuroimaging data analysis: DICOM to NIfTI conversion

    Li X, Morgan PS, Ashburner J, Smith J, Rorden C. The first step for neuroimaging data analysis: DICOM to NIfTI conversion. J Neurosci Methods. 2016;264:47-56. doi:10.1016/j.jneumeth.2016.03.001

    2016 doi

  28. [36]

    Perceptual Losses for Real-Time Style Transfer and Super- Resolution

    Johnson J, Alahi A, Li FF . Perceptual Losses for Real-Time Style Transfer and Super- Resolution. Vol 9906. (Leibe B, Matas J, Sebe N, Welling M, eds.). Springer International Publishing; 2016. doi:10.1007/978-3-319-46475-6

    2016 doi

  29. [37]

    Adam: A Method for Stochastic Optimization

    Kingma DP , Ba J. Adam: A Method for Stochastic Optimization. Published online December 22, 2014. http://arxiv.org/abs/1412.6980

    2014 arXiv

  30. [38]

    Least Squares Generative Adversarial Networks

    Mao X, Li Q, Xie H, Lau RYK, Wang Z, Smolley SP . Least Squares Generative Adversarial Networks

  31. [39]

    Automatic lung segmentation in routine imaging is primarily a data diversity problem, not a methodology problem

    Hofmanninger J, Prayer F , Pan J, Röhrich S, Prosch H, Langs G. Automatic lung segmentation in routine imaging is primarily a data diversity problem, not a methodology problem. Eur Radiol Exp. 2020;4(1). doi:10.1186/s41747-020-00173-2

    2020 doi

  32. [40]

    Optimal threshold in CT quantification of emphysema

    Wang Z, Gu S, Leader JK, et al. Optimal threshold in CT quantification of emphysema. Eur Radiol. 2013;23(4):975-984. doi:10.1007/s00330-012-2683-z 26

    2013 doi

  33. [41]

    Comparison of Computed Density and Microscopic Morphometry in Pulmonary Emphysema

    Gevenois PA, De Vuyst P , De Maertelaer V , et al. Comparison of Computed Density and Microscopic Morphometry in Pulmonary Emphysema. Vol 154.; 1996

    1996

  34. [42]

    Quantitative validation of the severity of emphysema by multi-detector CT

    Atta H, Seifeldein GS, Rashad A, Elmorshidy R. Quantitative validation of the severity of emphysema by multi-detector CT. Egyptian Journal of Radiology and Nuclear Medicine. 2015;46(2):355-361. doi:10.1016/j.ejrnm.2014.11.016

    2015 doi

  35. [43]

    Measurement STATISTICAL METHODS FOR ASSESSING AGREEMENT BETWEEN TWO METHODS OF CLINICAL MEASUREMENT

    Bland JM, Altman DG. Measurement STATISTICAL METHODS FOR ASSESSING AGREEMENT BETWEEN TWO METHODS OF CLINICAL MEASUREMENT

  36. [44]

    The bootstrap: A technique for data-driven statistics

    Henderson AR. The bootstrap: A technique for data-driven statistics. Using computer- intensive analyses to explore experimental data. Clinica Chimica Acta. 2005;359(1-2):1-

    2005

  37. [45]

    The Corsini Encyclopedia of Psychology - 2010 - McKnight - Mann‐Whitney U Test

    2010

  38. [46]

    Distribution matching losses can hallucinate features in medical image translation

    Cohen JP , Luck M, Honari S. Distribution matching losses can hallucinate features in medical image translation. In: Lecture Notes in Computer Science (Including Subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics). Vol 11070 LNCS. Springer V...

    2018 doi

  39. [47]

    doi:10.1016/j.cccn.2005.04.002

    2005 doi

  40. [48]

    Statistical Power Analysis for the Behavioral Sciences Second Edition

    Cohen J. Statistical Power Analysis for the Behavioral Sciences Second Edition

  41. [49]

    Emphysema quantification using low-dose computed tomography with deep learning–based kernel conversion comparison

    Bak SH, Kim JH, Jin H, et al. Emphysema quantification using low-dose computed tomography with deep learning–based kernel conversion comparison. Eur Radiol. 2020;30(12):6779-6787. doi:10.1007/s00330-020-07020-3

    2020 doi

  42. [50]

    Automated lung vessel segmentation reveals blood vessel volume redistribution in viral pneumonia

    Poletti J, Bach M, Yang S, et al. Automated lung vessel segmentation reveals blood vessel volume redistribution in viral pneumonia. Eur J Radiol. 2022;150. doi:10.1016/j.ejrad.2022.110259

    2022

  43. [51]

    Quantitative CT Analysis in Chronic Hypersensitivity Pneumonitis: A Convolutional Neural Network Approach

    Aliboni L, Dias OM, Pennati F , et al. Quantitative CT Analysis in Chronic Hypersensitivity Pneumonitis: A Convolutional Neural Network Approach. Acad Radiol. 2022;29:S31-S40. doi:10.1016/j.acra.2020.10.009

    2022 doi

  44. [52]

    Quantification of pathologic air trapping in lung transplant patients using ct density mapping: Comparison with other ct air trapping measures

    Solyanik O, Hollmann P , Dettmer S, et al. Quantification of pathologic air trapping in lung transplant patients using ct density mapping: Comparison with other ct air trapping measures. PLoS One. 2015;10(10). doi:10.1371/journal.pone.0139102

    2015 doi

  45. [53]

    CT-definable subtypes of chronic obstructive pulmonary disease: A statement of the fleischner society1

    Lynch DA, Austin JHM, Hogg JC, et al. CT-definable subtypes of chronic obstructive pulmonary disease: A statement of the fleischner society1. Radiology. 2015;277(1):192-

    2015

  46. [57]

    Body composition assessment with limited field-of-view computed tomography: A semantic image extension perspective

    Xu K, Li T, Khan MS, et al. Body composition assessment with limited field-of-view computed tomography: A semantic image extension perspective. Med Image Anal. 2023;88:102852. doi:10.1016/j.media.2023.102852 27 Tables Table 1. Reconstruction kernels from four different manufac...

    2023

  47. [100]

    doi:10.1097/RTI.0000000000000428

    doi

  48. [205]

    doi:10.1148/radiol.2015141579

    doi

  49. [2024]

    doi:10.1002/mp.17028

    doi

Pith tools

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