pith. sign in

arxiv: 1907.00273 · v1 · pith:6FH3C46Xnew · submitted 2019-06-29 · 📡 eess.IV · cs.CV

DuDoNet: Dual Domain Network for CT Metal Artifact Reduction

Wei-An Lin , Haofu Liao , Cheng Peng , Xiaohang Sun , Jingdan Zhang , Jiebo Luo , Rama Chellappa , Shaohua Kevin Zhou This is my paper

Pith reviewed 2026-05-25 12:23 UTC · model grok-4.3

classification 📡 eess.IV cs.CV
keywords metal artifact reductionCT imagingdual domain networksinogram consistencyRadon inversion layerdeep learningimage restorationend-to-end training
0
0 comments X

The pith

A dual-domain network with a Radon inversion layer corrects both sinograms and CT images to reduce metal artifacts.

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

The paper proposes DuDoNet, an end-to-end trainable network that restores consistency in the X-ray projection domain while enhancing the reconstructed CT images. Existing single-domain methods either miss the structured non-local nature of metal artifacts or create new inconsistencies when altering projections alone. The network links the two domains through a differentiable Radon inversion layer so that corrections in one domain can influence the other during training. A reader would care because metallic implants are common in patients, and current CT images often remain unusable for diagnosis without better artifact removal.

Core claim

The authors claim that simultaneously operating in the sinogram and image domains via an end-to-end network, connected by a novel Radon inversion layer that permits gradient flow from image to sinogram, produces superior metal artifact reduction compared with prior single-domain approaches and constitutes the first such dual-domain architecture for this task.

What carries the argument

The Radon inversion layer, which models the forward and backward projection processes to link the sinogram domain and image domain while allowing end-to-end gradient back-propagation.

If this is right

  • Joint optimization across domains removes both the original metal streaks and the secondary artifacts that arise from inconsistent sinogram edits.
  • Gradient signals from the final image quality can directly improve the sinogram restoration step.
  • The same architecture can in principle be retrained on different scanner geometries or implant materials.
  • Clinical CT workflows could incorporate the network as a post-processing step without separate sinogram and image pipelines.

Where Pith is reading between the lines

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

  • The dual-domain linkage may extend to other tomography problems where projection data and reconstructed images must stay consistent, such as limited-angle reconstruction.
  • Replacing the analytic Radon layer with a learned projection operator could test whether the performance gain comes from the domain coupling itself or from the specific inversion formula.
  • If the method scales to cone-beam CT, it would directly affect dental and interventional imaging where metal artifacts are frequent.

Load-bearing premise

The Radon inversion layer must accurately represent the actual CT projection mathematics so that gradients can flow usefully between domains without creating new inconsistencies.

What would settle it

Training the network on a dataset of real CT scans containing known metallic implants and then measuring whether residual artifacts remain larger than those produced by the best single-domain baseline on the same scans.

Figures

Figures reproduced from arXiv: 1907.00273 by Cheng Peng, Haofu Liao, Jiebo Luo, Jingdan Zhang, Rama Chellappa, Shaohua Kevin Zhou, Wei-An Lin, Xiaohang Sun.

Figure 1
Figure 1. Figure 1: (a) Sample MAR results for a CT image with [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: The proposed Dual Domain Network (DuDoNet) for MAR. Given a degraded sinogram [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Sample simulated metal artifact on patient CT. The X-ray spectrum is shown in the upper-left corner. Metallic [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Visual comparisons between models without RC [PITH_FULL_IMAGE:figures/full_fig_p006_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Visual comparisons between models without MP [PITH_FULL_IMAGE:figures/full_fig_p006_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Visual comparisons between models without SE [PITH_FULL_IMAGE:figures/full_fig_p006_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Visual comparisons on MAR for different types of metallic implants. [PITH_FULL_IMAGE:figures/full_fig_p007_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: illustrates the general fanbeam CT geometry. The X-ray source and the arc detector rotate with respect to the origin. The distance between the X-ray source and the origin is D. For each projection angle β, the arc detector receives the X-rays transmitted from the object. The inten￾sity values received by the detector is represented as a 1D signal with independent variable γ. As shown in the top of [PITH_F… view at source ↗
Figure 9
Figure 9. Figure 9: Parallel-beam CT geometry. RIL consists of three modules: (1) parallel-beam conver￾sion module, (2) Ram-Lak filtering module and (3) back￾projection module. Given a fanbeam sinogram Yfan(β, γ), we first convert it to a parallel beam sinogram Ypara(t, θ) using the fan-to-parallel beam conversion. Then, we can reconstruct the CT image X(u, v) using the Ram-Lak fil- [PITH_FULL_IMAGE:figures/full_fig_p009_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Evaluations on real data. All models are exactly [PITH_FULL_IMAGE:figures/full_fig_p011_10.png] view at source ↗
read the original abstract

Computed tomography (CT) is an imaging modality widely used for medical diagnosis and treatment. CT images are often corrupted by undesirable artifacts when metallic implants are carried by patients, which creates the problem of metal artifact reduction (MAR). Existing methods for reducing the artifacts due to metallic implants are inadequate for two main reasons. First, metal artifacts are structured and non-local so that simple image domain enhancement approaches would not suffice. Second, the MAR approaches which attempt to reduce metal artifacts in the X-ray projection (sinogram) domain inevitably lead to severe secondary artifact due to sinogram inconsistency. To overcome these difficulties, we propose an end-to-end trainable Dual Domain Network (DuDoNet) to simultaneously restore sinogram consistency and enhance CT images. The linkage between the sigogram and image domains is a novel Radon inversion layer that allows the gradients to back-propagate from the image domain to the sinogram domain during training. Extensive experiments show that our method achieves significant improvements over other single domain MAR approaches. To the best of our knowledge, it is the first end-to-end dual-domain network for MAR.

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 / 1 minor

Summary. The paper proposes DuDoNet, an end-to-end trainable dual-domain neural network for metal artifact reduction (MAR) in CT. It uses a novel Radon inversion layer to link the sinogram and image domains, enabling simultaneous sinogram consistency restoration and image enhancement via gradient back-propagation. The central claim is that this yields significant improvements over single-domain MAR methods and is the first such end-to-end dual-domain network.

Significance. If the Radon inversion layer provides a consistent differentiable linkage, the work would represent a meaningful advance in MAR by moving beyond single-domain limitations to joint optimization. The introduction of the layer itself could have broader utility in differentiable tomography pipelines. The paper explicitly positions the contribution as the first end-to-end dual-domain approach.

major comments (1)
  1. [Abstract] Abstract (Radon inversion layer paragraph): The central claim that the layer 'allows the gradients to back-propagate from the image domain to the sinogram domain during training' without introducing new inconsistencies rests on an unverified assumption that the discrete implementation exactly matches the physical projection geometry and forms a consistent (adjoint) pair. No explicit verification (e.g., adjoint test, gradient consistency ablation, or comparison to the data-generation forward model) is referenced; if the pair is approximate, dual-domain training optimizes an inconsistent objective, which could undermine the reported gains over single-domain baselines.
minor comments (1)
  1. [Abstract] Abstract: 'sigogram' is a typo and should read 'sinogram'.

Simulated Author's Rebuttal

1 responses · 0 unresolved

We thank the referee for the constructive comment on the Radon inversion layer. We address the concern point-by-point below and agree that additional verification will strengthen the manuscript.

read point-by-point responses

  1. Referee: [Abstract] Abstract (Radon inversion layer paragraph): The central claim that the layer 'allows the gradients to back-propagate from the image domain to the sinogram domain during training' without introducing new inconsistencies rests on an unverified assumption that the discrete implementation exactly matches the physical projection geometry and forms a consistent (adjoint) pair. No explicit verification (e.g., adjoint test, gradient consistency ablation, or comparison to the data-generation forward model) is referenced; if the pair is approximate, dual-domain training optimizes an inconsistent objective, which could undermine the reported gains over single-domain baselines.

    Authors: We agree that the manuscript does not reference explicit verification (adjoint test, gradient consistency ablation, or direct comparison to the forward model used for data generation). The layer is implemented as a differentiable approximation to the inverse Radon transform to enable gradient flow, but without the requested checks it is not demonstrated that the discrete pair is exactly consistent. In the revision we will add (i) an adjoint test for the layer, (ii) a gradient-consistency ablation, and (iii) a comparison against the simulation forward projector, together with a brief discussion of any residual approximation error. These additions will be placed in the Methods section and referenced from the abstract. revision: yes

Circularity Check

0 steps flagged

No circularity: data-driven network with experimental validation

full rationale

The paper introduces DuDoNet as an end-to-end trainable dual-domain CNN for metal artifact reduction, with a novel but explicitly constructed Radon inversion layer to enable gradient flow between sinogram and image domains. All performance claims rest on comparative experiments against single-domain baselines rather than any self-referential definition, fitted parameter renamed as prediction, or load-bearing self-citation. The linkage between domains is presented as an engineering choice whose correctness is tested empirically, not derived tautologically from the inputs. No equations or steps reduce by construction to the training data or prior outputs of the same model.

Axiom & Free-Parameter Ledger

1 free parameters · 1 axioms · 1 invented entities

The central claim depends on the effectiveness of the novel layer and the assumption that dual-domain training can resolve inconsistencies better than single-domain methods, with network parameters fitted during training.

free parameters (1)
  • network hyperparameters and loss weights
    The specific architecture details, training parameters, and balancing of dual domain losses are chosen to achieve the reported performance.
axioms (1)
  • domain assumption The CT imaging process can be accurately modeled by the Radon transform for the purpose of inversion and gradient flow.
    Invoked in the description of the Radon inversion layer linking the domains.
invented entities (1)
  • Radon inversion layer no independent evidence
    purpose: To enable differentiable mapping between sinogram and image domains for end-to-end training.
    New component proposed in the paper to connect the two domains.

pith-pipeline@v0.9.0 · 5744 in / 1402 out tokens · 33206 ms · 2026-05-25T12:23:40.959709+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

36 extracted references · 36 canonical work pages · 3 internal anchors

  1. [1]

    Adler and O

    J. Adler and O. ¨Oktem. Learned primal-dual reconstruction. IEEE Transactions on Medical Imaging , 37(6):1322–1332, June 2018. 2, 4 4The domain of CT images with simulated metal artifacts. Raw CT LI NMAR RDN-CT CNNMAR DuDoNet Raw CT LI NMAR RDN-CT CNNMAR DuDoNet Figure 10: Evaluations on real data. All models are exactly the same as in the main paper (no ...

    work page 2018

  2. [2]

    Bestimmung der absorption des rothen lichts in farbigen fl ¨ussigkeiten

    Beer. Bestimmung der absorption des rothen lichts in farbigen fl ¨ussigkeiten. Annalen der Physik und Chemie , 162(5):78–88, 1852. 3

    work page

  3. [3]

    Dabov, A

    K. Dabov, A. Foi, V . Katkovnik, and K. Egiazarian. Image denoising by sparse 3-d transform-domain collaborative fil- tering. IEEE Transactions on image processing, 16(8):2080– 2095, 2007. 1

    work page 2080

  4. [4]

    X. Duan, L. Zhang, Y . Xiao, J. Cheng, Z. Chen, and Y . Xing. Metal artifact reduction in ct images by sinogram tv inpaint- ing. In Nuclear Science Symposium Conference Record,

    work page

  5. [5]

    IEEE, pages 4175–4177

    NSS’08. IEEE, pages 4175–4177. IEEE, 2008. 1

    work page 2008

  6. [6]

    Gjesteby, Q

    L. Gjesteby, Q. Yang, Y . Xi, B. Claus, Y . Jin, B. De Man, and G. Wang. Reducing metal streak artifacts in ct images via deep learning: Pilot results. In The 14th International Meeting on Fully Three-Dimensional Image Reconstruction in Radiology and Nuclear Medicine , pages 611–614, 2017. 3

    work page 2017

  7. [7]

    Gjesteby, Q

    L. Gjesteby, Q. Yang, Y . Xi, Y . Zhou, J. Zhang, and G. Wang. Deep learning methods to guide ct image reconstruction and reduce metal artifacts. In SPIE Medical Imaging, 2017. 3

    work page 2017

  8. [8]

    Guo and H

    J. Guo and H. Chao. Building dual-domain representations for compression artifacts reduction. In European Conference on Computer Vision, pages 628–644. Springer, 2016. 1

    work page 2016

  9. [9]

    Gupta, K

    H. Gupta, K. H. Jin, H. Q. Nguyen, M. T. McCann, and M. Unser. Iterative metal artifact reduction for x-ray com- puted tomography using unmatched projector/backprojector pairs. IEEE Transactions on Medical Imaging, 43(6):3019– 3033, 2016. 3

    work page 2016

  10. [10]

    Isola, J.-Y

    P. Isola, J.-Y . Zhu, T. Zhou, and A. A. Efros. Image-to-image translation with conditional adversarial networks. In The IEEE Conference on Computer Vision and Pattern Recog- nition (CVPR), pages 1125–1134, 2017. 2

    work page 2017

  11. [11]

    K. H. Jin, M. T. McCann, E. Froustey, and M. Unser. Deep convolutional neural network for inverse problems in imag- ing. IEEE Transactions on Image Processing , 26(9):4509– 4522, Sept 2017. 2, 4

    work page 2017

  12. [12]

    A. C. Kak and M. Slaney. Principles of computerized tomo- graphic imaging. Society for Industrial and Applied Mathe- matics, 2001. 3, 4

    work page 2001

  13. [13]

    W. A. Kalender, R. Hebel, and J. Ebersberger. Reduc- tion of ct artifacts caused by metallic implants. Radiology, 164(2):576–577, 1987. 1, 3, 4, 8, 10, 13, 14

    work page 1987

  14. [14]

    Adam: A Method for Stochastic Optimization

    D. Kingma and J. Ba. Adam: A method for stochastic opti- mization. arXiv preprint arXiv:1412.6980, 2014. 5

    work page internal anchor Pith review Pith/arXiv arXiv 2014

  15. [15]

    Ledig, L

    C. Ledig, L. Theis, F. Husz ´ar, J. Caballero, A. Cunning- ham, A. Acosta, A. P. Aitken, A. Tejani, J. Totz, Z. Wang, et al. Photo-realistic single image super-resolution using a generative adversarial network. In The IEEE Conference on Computer Vision and Pattern Recognition (CVPR) , vol- ume 2, page 4, 2017. 1

    work page 2017

  16. [16]

    Lehtinen, J

    J. Lehtinen, J. Munkberg, J. Hasselgren, S. Laine, T. Kar- ras, M. Aittala, and T. Aila. Noise2Noise: Learning image restoration without clean data. In International Conference on Machine Learning (ICML), volume 80, pages 2965–2974,

    work page

  17. [17]

    Makitalo and A

    M. Makitalo and A. Foi. Optimal inversion of the anscombe transformation in low-count poisson image denoising. IEEE transactions on Image Processing, 20(1):99–109, 2011. 1

    work page 2011

  18. [18]

    Mehranian, M

    A. Mehranian, M. R. Ay, A. Rahmim, and H. Zaidi. X-ray ct metal artifact reduction using wavelet domain l0 sparse regularization. IEEE Transactions on Medical Imaging , 32:1707–1722, 2013. 1, 3

    work page 2013

  19. [19]

    Meyer, R

    E. Meyer, R. Raupach, M. Lell, B. Schmidt, and M. Kachel- rieß. Normalized metal artifact reduction (nmar) in com- puted tomography. Medical physics , 37(10):5482–5493,

    work page

  20. [20]

    J. Pan, W. Ren, Z. Hu, and M. Yang. Learning to deblur im- ages with exemplars. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2018. 2

    work page 2018

  21. [21]

    H. S. Park, Y . E. Chung, S. M. Lee, H. P. Kim, and J. K. Seo. Sinogram-consistency learning in ct for metal artifact reduction. arXiv preprint arXiv:1708.00607, 2017. 2, 3

    work page internal anchor Pith review Pith/arXiv arXiv 2017

  22. [22]

    Paszke, S

    A. Paszke, S. Gross, S. Chintala, G. Chanan, E. Yang, Z. De- Vito, Z. Lin, A. Desmaison, L. Antiga, and A. Lerer. Auto- matic differentiation in pytorch. In NIPS-W, 2017. 5, 9

    work page 2017

  23. [23]

    Ronneberger, P

    O. Ronneberger, P. Fischer, and T. Brox. U-net: Convolu- tional networks for biomedical image segmentation. InMed- ical Image Computing and Computer Assisted Intervention (MICCAI), pages 234–241. Springer, 2015. 2, 9

    work page 2015

  24. [24]

    Ulyanov, A

    D. Ulyanov, A. Vedaldi, and V . Lempitsky. Deep image prior. In The IEEE Conference on Computer Vision and Pattern Recognition (CVPR), June 2018. 2

    work page 2018

  25. [25]

    J. Wang, Y . Zhao, J. H. Noble, and B. M. Dawant. Condi- tional generative adversarial networks for metal artifact re- duction in ct images of the ear. InMedical Image Computing and Computer Assisted Intervention (MICCAI) , 2018. 2, 3, 8, 10, 13, 14

    work page 2018

  26. [26]

    X. Wang, K. Yu, S. Wu, J. Gu, Y . Liu, C. Dong, Y . Qiao, and C. C. Loy. Esrgan: Enhanced super-resolution genera- tive adversarial networks. In The European Conference on Computer Vision Workshops (ECCVW), September 2018. 1, 2, 8

    work page 2018

  27. [27]

    W ¨urfl, F

    T. W ¨urfl, F. C. Ghesu, V . Christlein, and A. Maier. Deep learning computed tomography. In Medical Image Com- puting and Computer Assisted Intervention (MICCAI), pages 432–440. Springer International Publishing, 2016. 2, 4

    work page 2016

  28. [28]

    Xu and H

    S. Xu and H. Dang. Deep residual learning enabled metal artifact reduction in ct. In Medical Imaging 2018: Physics of Medical Imaging, volume 10573, page 105733O. Interna- tional Society for Optics and Photonics, 2018. 3

    work page 2018

  29. [29]

    K. Yan, X. Wang, L. Lu, L. Zhang, A. P. Harrison, M. Bagheri, and R. M. Summers. Deep lesion graphs in the wild: Relationship learning and organization of signifi- cant radiology image findings in a diverse large-scale lesion database. In The IEEE Conference on Computer Vision and Pattern Recognition (CVPR), June 2018. 5, 10

    work page 2018

  30. [30]

    Zhang, B

    H. Zhang, B. Dong, and B. Liu. A reweighted joint spatial- radon domain ct image reconstruction model for metal ar- tifact reduction. SIAM J. Imaging Sciences , 11:707–733,

    work page

  31. [31]

    Zhang and V

    H. Zhang and V . M. Patel. Densely connected pyramid de- hazing network. In The IEEE Conference on Computer Vi- sion and Pattern Recognition (CVPR), 2018. 2

    work page 2018

  32. [32]

    Zhang, W

    X. Zhang, W. Yang, Y . Hu, and J. Liu. Dmcnn: Dual-domain multi-scale convolutional neural network for compression ar- tifacts removal. In IEEE International Conference on Image Processing (ICIP), 2018. 1

    work page 2018

  33. [33]

    Zhang, Y

    Y . Zhang, Y . Tian, Y . Kong, B. Zhong, and Y . Fu. Resid- ual dense network for image super-resolution. In The IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018. 1, 2, 8, 13, 14

    work page 2018

  34. [34]

    Zhang and H

    Y . Zhang and H. Yu. Convolutional neural network based metal artifact reduction in x-ray computed tomography. IEEE Transactions on Medical Imaging , 2018. 1, 2, 3, 5, 8, 10, 13, 14

    work page 2018

  35. [35]

    Zhang, X

    Z. Zhang, X. Liang, X. Dong, Y . Xie, and G. Cao. A sparse-view ct reconstruction method based on combination of densenet and deconvolution. IEEE Transactions on Med- ical Imaging, 37(6):1407–1417, June 2018. 2, 4

    work page 2018

  36. [36]

    Joint Sub-bands Learning with Clique Structures for Wavelet Domain Super-Resolution

    Z. Zhong, T. Shen, Y . Yang, Z. Lin, and C. Zhang. Joint sub-bands learning with clique structures for wavelet domain super-resolution. arXiv preprint arXiv:1809.04508, 2018. 1 Ground Truth With Metal Artifact LI [12] NMAR [18] cGAN-CT [24] RDN-CT [32] CNNMAR [33] DuDoNet Ground Truth With Metal Artifact LI [12] NMAR [18] cGAN-CT [24] RDN-CT [32] CNNMAR [...

    work page internal anchor Pith review Pith/arXiv arXiv 2018