pith. machine review for the scientific record. sign in

arxiv: 2605.12164 · v1 · submitted 2026-05-12 · 📡 eess.IV · physics.med-ph

Recognition: no theorem link

A Comparative Analysis of CT Degradation for LDCT Nodule Classification using Radiomics

Anna Corti, Jiaying Liu, Luca Mainardi, Valentina D.A. Corino

Pith reviewed 2026-05-13 04:29 UTC · model grok-4.3

classification 📡 eess.IV physics.med-ph
keywords low-dose CTlung nodule classificationradiomicsCycleGANimage degradationsynthetic datadomain adaptationmachine learning
0
0 comments X

The pith

CycleGAN degradation of standard-dose CT scans creates synthetic low-dose images that train lung nodule classifiers with better balance than undegraded data.

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

The paper tests three methods for turning standard-dose CT scans into synthetic low-dose versions to expand training data for nodule classification. It shows that models trained on the resulting images extract radiomic features and classify better on real low-dose screening cases than a baseline trained only on original high-dose scans. CycleGAN produced the synthetic images whose statistics most closely matched actual low-dose scans and delivered the highest classifier performance. This approach lets researchers build screening models without collecting new low-dose patient data.

Core claim

Degraded images generated using CycleGAN approach led to the most balanced performance on the classification task using Adam Booster classifier, achieving an AUC of 0.861, sensitivity of 0.743 and specificity of 0.858 in the independent test. The baseline model trained on non-degraded SDCT data failed to generalize to the real LDCT set with AUC 0.789 and sensitivity 0.571. CycleGAN also achieved the best Fréchet inception distance and kernel inception distance scores among the three degradation methods.

What carries the argument

CycleGAN unpaired image-to-image translation that simulates low-dose CT noise and contrast from standard-dose scans, followed by radiomic feature extraction for machine learning nodule classification.

If this is right

  • A model trained only on standard-dose data loses sensitivity when tested on real low-dose screening scans.
  • CycleGAN degradation produces images with the closest statistical match to real low-dose CT among the three methods tested.
  • Synthetic low-dose data generated this way can substitute for real low-dose cases when training classifiers on the LIDC-IDRI dataset.
  • The three degradation approaches differ in how well their outputs support downstream radiomics-based classification.

Where Pith is reading between the lines

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

  • Researchers could generate large training sets for other low-dose imaging tasks without new patient scans.
  • The same degradation step might be applied to improve model robustness in CT tasks beyond lung nodules.
  • If the feature distributions match closely enough, this method could lower the radiation exposure needed during model development.

Load-bearing premise

Radiomic features extracted from the synthetic low-dose images behave the same way as features from real low-dose scans when used for nodule classification.

What would settle it

Train the same classifier on a held-out set of real low-dose CT cases using either the synthetic degradation pipeline or actual low-dose scans and compare the resulting AUC and sensitivity.

Figures

Figures reproduced from arXiv: 2605.12164 by Anna Corti, Jiaying Liu, Luca Mainardi, Valentina D.A. Corino.

Figure 1
Figure 1. Figure 1: Examples of SDCT (a) and LDCT (b) subjects from the LIDC-IDRI dataset and examples of the corresponding [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Example of paired SDCT and simulated LDCT slices from the Low-dose CT image and projection dataset. [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Overview of the pipeline. consistent generative adversarial network (CycleGAN), a popular unpaired image-to-image translation method [18]. A detailed description of these methods and the implemented training strategies is provided in the following sections. To compare the LDCT simulation methods, we propose a pipeline (Fig.3) that includes a radiomics-based nodule classification task. First, the LIDC-IDRI … view at source ↗
Figure 4
Figure 4. Figure 4: Pipeline to simulate low-dose computed tomography (LDCT). The Radon transformation is applied to the [PITH_FULL_IMAGE:figures/full_fig_p005_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Architecture of Pix2Pix, composed by a generator (G) and a discriminator (D), used to replicate the degradation [PITH_FULL_IMAGE:figures/full_fig_p006_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Schematic representation of the CycleGAN model. Generator G translates images from the SDCT to the [PITH_FULL_IMAGE:figures/full_fig_p007_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Pipeline for pulmonary nodule classification using radiomics. [PITH_FULL_IMAGE:figures/full_fig_p008_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Fig.8. All three methods introduce visible non-uniform noise compared to the original SD patch. Method 1 presents [PITH_FULL_IMAGE:figures/full_fig_p009_8.png] view at source ↗
Figure 8
Figure 8. Figure 8: Visual analysis of nodule patches from degradation methods. The figure compares patches from the original [PITH_FULL_IMAGE:figures/full_fig_p010_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Venn diagram representing the number of features considered stable, discriminant and non-redundant for each [PITH_FULL_IMAGE:figures/full_fig_p011_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Bootstrapped distribution violin plots of the nodule classification model on the independent real LDCT test [PITH_FULL_IMAGE:figures/full_fig_p012_10.png] view at source ↗
read the original abstract

Low-dose computed tomography (LDCT) is the standard modality for lung cancer screening, known for its low radiation dose but high noise levels. While existing literature focuses on denoising LDCT images, comparative research on simulating LDCT characteristics to directly use these images for model development is lacking. This study shifts the focus from denoising images to degrading available standard-dose CT (SDCT) data, generating synthetic images for data augmentation to train classifiers for screening-detected nodules. We compare three degradation methods: (1) a sinogram domain statistical noise insertion; (2) replicate a validated physics-based simulation using Pix2Pix; and (3) unpaired CycleGAN. The generated images were utilized to simulate LDCT screening scenario replacing 695 SDCT cases from the LIDC-IDRI dataset, from which radiomic features were extracted to train machine learning models for lung nodule classification. Regarding image quality, CycleGAN achieved the best Fr\'echet inception distance (0.1734) and kernel inception distance (0.0813; 0.1002) scores, indicating distributional alignment with the target low-dose domain. In the nodule classification task, results confirmed the necessity of domain adaptation since a baseline model trained on non-degraded SDCT data failed to generalize to the real LDCT set (AUC 0.789) with a low sensitivity (0.571). Degraded images generated using CycleGAN approach led to the most balanced performance on the classification task using Adam Booster classifier, achieving an AUC of 0.861, sensitivity of 0.743 and specificity of 0.858 in the independent test. Our findings confirm that generating synthetic LDCT data from standard-dose scans is a viable strategy for training robust nodule classifiers for screening detected nodules.

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

1 major / 2 minor

Summary. The paper claims that among three degradation methods applied to SDCT images to simulate LDCT (sinogram statistical noise insertion, Pix2Pix, and unpaired CycleGAN), the CycleGAN approach produces synthetic images with the closest distributional alignment to real LDCT (FID 0.1734, KID 0.0813/0.1002). When used to augment the LIDC-IDRI dataset for radiomics-based nodule classification, it yields superior performance with the Adam Booster classifier on an independent real LDCT test set (AUC 0.861, sensitivity 0.743, specificity 0.858), compared to baseline SDCT training (AUC 0.789, sensitivity 0.571). The authors conclude this is a viable strategy for training robust classifiers for screening-detected nodules.

Significance. If substantiated, this work provides evidence that image-to-image translation models like CycleGAN can effectively simulate LDCT degradation for data augmentation in radiomics pipelines. This has potential significance for medical imaging AI, as it leverages more readily available SDCT data to improve model performance on noisy LDCT without additional radiation exposure or data collection. The reported improvements in balanced classification metrics highlight practical benefits for lung cancer screening applications.

major comments (1)
  1. [Results] The improved classification performance with CycleGAN-degraded images (AUC 0.861) is presented as resulting from better simulation of LDCT characteristics. However, the manuscript provides no quantitative analysis comparing the distributions of radiomic features extracted from the CycleGAN synthetic LDCT images versus real LDCT images (such as per-feature statistical tests or multivariate distances). This is a load-bearing gap because the central claim that the synthetics enable generalization to real LDCT depends on the features behaving similarly, rather than the improvement arising from generic augmentation or other artifacts.
minor comments (2)
  1. [Methods] The experimental setup lacks explicit description of how the 695 SDCT cases were selected for replacement, the composition of the independent test set, and whether stratified splitting or cross-validation was employed to ensure robust evaluation.
  2. [Abstract] The KID scores are reported as 0.0813; 0.1002 without clarifying what the two values represent (e.g., different kernel sizes or train/test splits).

Simulated Author's Rebuttal

1 responses · 0 unresolved

We thank the referee for their thorough and constructive review of our manuscript. We have carefully considered the major comment and provide a point-by-point response below, including plans for revision.

read point-by-point responses

  1. Referee: [Results] The improved classification performance with CycleGAN-degraded images (AUC 0.861) is presented as resulting from better simulation of LDCT characteristics. However, the manuscript provides no quantitative analysis comparing the distributions of radiomic features extracted from the CycleGAN synthetic LDCT images versus real LDCT images (such as per-feature statistical tests or multivariate distances). This is a load-bearing gap because the central claim that the synthetics enable generalization to real LDCT depends on the features behaving similarly, rather than the improvement arising from generic augmentation or other artifacts.

    Authors: We agree that this represents a substantive gap in the current manuscript. While the CycleGAN approach demonstrated the strongest image-level distributional alignment via FID (0.1734) and KID (0.0813/0.1002), these metrics do not directly confirm similarity in the specific radiomic feature space used for nodule classification. The observed improvement in generalization to the independent real LDCT test set (AUC 0.861 vs. baseline 0.789) could indeed partly reflect generic augmentation benefits rather than targeted LDCT simulation. In the revised manuscript, we will add quantitative comparisons of radiomic feature distributions, including per-feature statistical tests (e.g., Kolmogorov-Smirnov or Wilcoxon rank-sum tests with multiple-comparison correction) and multivariate measures (e.g., Mahalanobis distance or Earth Mover's Distance on the feature vectors) between CycleGAN-degraded images and real LDCT. We will also report these results alongside the classification metrics to better substantiate the mechanism of improvement. revision: yes

Circularity Check

0 steps flagged

Empirical comparison study with no derivation chain or self-referential reductions

full rationale

The manuscript is a standard empirical ML pipeline: three image-degradation techniques (sinogram noise insertion, Pix2Pix, CycleGAN) are applied to SDCT scans from LIDC-IDRI to produce synthetic LDCT images; radiomic features are extracted; classifiers are trained and evaluated on a held-out real LDCT test set. No equations, uniqueness theorems, fitted parameters renamed as predictions, or self-citations are invoked to derive the central performance numbers (AUC 0.861 etc.). All reported metrics are direct experimental outcomes on external data, so the result does not reduce to its inputs by construction.

Axiom & Free-Parameter Ledger

2 free parameters · 1 axioms · 0 invented entities

The paper relies on standard assumptions in medical image synthesis and radiomics; no new entities postulated.

free parameters (2)
  • CycleGAN training parameters
    Hyperparameters for the generative model are fitted during training to match the low-dose domain.
  • Noise insertion parameters
    Statistical parameters for sinogram noise simulation.
axioms (1)
  • domain assumption Synthetic degraded images preserve discriminative radiomic features for nodule classification
    The approach assumes that features from degraded images are representative of real LDCT for ML training.

pith-pipeline@v0.9.0 · 5631 in / 1211 out tokens · 151381 ms · 2026-05-13T04:29:00.130228+00:00 · methodology

discussion (0)

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

Reference graph

Works this paper leans on

22 extracted references · 22 canonical work pages

  1. [1]

    Sung et al

    H. Sung et al. Global cancer statistics 2020: GLOBOCAN estimates of incidence and mortality worldwide for 36 cancers in 185 countries.CA: A Cancer Journal for Clinicians, 71(3):209–249, May 2021

    work page 2020

  2. [2]

    H. J. de Koning et al. Reduced lung-cancer mortality with volume CT screening in a randomized trial.New England Journal of Medicine, 382(6):503–513, February 2020

    work page 2020

  3. [3]

    Reduced lung-cancer mortality with low-dose computed tomographic screening.New England Journal of Medicine, 365(5):395–409, August 2011

    The National Lung Screening Trial Research Team. Reduced lung-cancer mortality with low-dose computed tomographic screening.New England Journal of Medicine, 365(5):395–409, August 2011

    work page 2011

  4. [4]

    W. Kim, S. Y . Jeon, G. Byun, H. Yoo, and J. H. Choi. A systematic review of deep learning-based denoising for low-dose computed tomography from a perceptual quality perspective.Biomedical Engineering Letters, 14(6):1153–1173, August 2024

    work page 2024

  5. [5]

    K. A. S. H. Kulathilake, N. A. Abdullah, A. Q. M. Sabri, and K. W. Lai. A review on deep learning approaches for low-dose computed tomography restoration.Complex & Intelligent Systems, 9(3):2713–2745, June 2023. 13 APREPRINT- MAY13, 2026

    work page 2023

  6. [6]

    T. Zhao, M. McNitt-Gray, and D. Ruan. A convolutional neural network for ultra-low-dose CT denoising and emphysema screening.Medical Physics, 46(9):3941–3950, September 2019

    work page 2019

  7. [7]

    Gholizadeh-Ansari, J

    M. Gholizadeh-Ansari, J. Alirezaie, and P. Babyn. Deep learning for low-dose CT denoising using perceptual loss and edge detection layer.Journal of Digital Imaging, 33(2):504–515, September 2019

    work page 2019

  8. [8]

    Chen et al

    H. Chen et al. Low-dose CT denoising with convolutional neural network. InProceedings of the International Symposium on Biomedical Imaging (ISBI), pages 143–146, June 2017

    work page 2017

  9. [9]

    Eulig, B

    E. Eulig, B. Ommer, and M. Kachelrieß. Benchmarking deep learning-based low-dose CT image denoising algorithms.Medical Physics, 51(12):8776–8788, December 2024

    work page 2024

  10. [10]

    C. H. McCollough et al. Low-dose CT for the detection and classification of metastatic liver lesions: Results of the 2016 low dose CT grand challenge.Medical Physics, 44(10):e339–e352, October 2017

    work page 2016

  11. [11]

    S. G. Armato et al. The lung image database consortium (LIDC) and image database resource initiative (IDRI): A completed reference database of lung nodules on CT scans.Medical Physics, 38(2), 2011

    work page 2011

  12. [12]

    T. R. Moen et al. Low-dose CT image and projection dataset.Medical Physics, 48(2):902–911, February 2021

    work page 2021

  13. [13]

    Snowsill et al

    T. Snowsill et al. Low-dose computed tomography for lung cancer screening in high-risk populations: a systematic review and economic evaluation.Health Technology Assessment, 22(69):1–276, November 2018

    work page 2018

  14. [14]

    Rampinelli, D

    C. Rampinelli, D. Origgi, and M. Bellomi. Low-dose CT: technique, reading methods and image interpretation. Cancer Imaging, 12(3):548, 2013

    work page 2013

  15. [15]

    L. Yu, M. Shiung, D. Jondal, and C. H. McCollough. Development and validation of a practical lower-dose- simulation tool for optimizing computed tomography scan protocols.Journal of Computer Assisted Tomography, 36(4):477–487, 2012

    work page 2012

  16. [16]

    Zeng et al

    D. Zeng et al. A simple low-dose X-ray CT simulation from high-dose scan.IEEE Transactions on Nuclear Science, 62(5):2226–2233, October 2015

    work page 2015

  17. [17]

    Isola, J

    P. Isola, J. Y . Zhu, T. Zhou, and A. A. Efros. Image-to-image translation with conditional adversarial networks. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2017

    work page 2017

  18. [18]

    J. Y . Zhu, T. Park, P. Isola, and A. A. Efros. Unpaired image-to-image translation using cycle-consistent adversarial networks. InProceedings of the IEEE International Conference on Computer Vision (ICCV), pages 2242–2251, March 2017

    work page 2017

  19. [19]

    J. Liu, A. Corti, V . D. A. Corino, and L. Mainardi. Lung nodule classification using radiomics model trained on degraded SDCT images.Computer Methods and Programs in Biomedicine, 257:108474, December 2024

    work page 2024

  20. [20]

    J. J. M. van Griethuysen et al. Computational radiomics system to decode the radiographic phenotype.Cancer Research, 77(21):e104–e107, November 2017

    work page 2017

  21. [21]

    Lo Iacono, R

    F. Lo Iacono, R. Maragna, G. Pontone, and V . D. A. Corino. A novel data augmentation method for radiomics analysis using image perturbations.Journal of Imaging Informatics in Medicine, 37(5):2401–2414, May 2024

    work page 2024

  22. [22]

    Christensen et al

    J. Christensen et al. ACR lung-RADS v2022: Assessment categories and management recommendations.Chest, 165(3):738–753, March 2024. 14

    work page 2024