Respiratory Motion Correction in Abdominal MRI using a Densely Connected U-Net with GAN-guided Training

Hing-Chiu Chang; Ka-Wai Kwok; Kit-Hang Lee; Qi Dou; Shihui Chen; Wenhao Jiang; Yui-Lun Ng; Zhiyu Liu

arxiv: 1906.09745 · v1 · pith:CQGGERUQnew · submitted 2019-06-24 · 📡 eess.IV · cs.AI

Respiratory Motion Correction in Abdominal MRI using a Densely Connected U-Net with GAN-guided Training

Wenhao Jiang , Zhiyu Liu , Kit-Hang Lee , Shihui Chen , Yui-Lun Ng , Qi Dou , Hing-Chiu Chang , Ka-Wai Kwok This is my paper

Pith reviewed 2026-05-25 17:19 UTC · model grok-4.3

classification 📡 eess.IV cs.AI

keywords abdominal MRIrespiratory motion correctionmotion artifactsU-NetGANdeep learningimage restorationperceptual loss

0 comments

The pith

A densely connected U-Net with GAN-guided training recovers abdominal MRI images from respiratory motion artifacts.

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

The paper aims to show that a deep learning model can restore clear abdominal MRI scans degraded by breathing motion without extra scan time or patient interventions like breath-holding. The model combines a densely connected U-Net architecture with adversarial training from a GAN and a perceptual loss to map corrupted images back to high-quality ones. It tests this on MRI data containing both simulated motion and real patient breathing artifacts. If successful, the approach offers a post-processing solution that avoids the discomfort and time costs of conventional motion mitigation methods. Results indicate the model produces effective motion removal across the tested cases.

Core claim

A densely connected U-Net with generative adversarial network-guided training and a perceptual loss function recovers MR images from respiratory motion artifacts, demonstrated on a diverse collection of MRI data affected by both synthetic and authentic respiration artifacts.

What carries the argument

Densely connected U-Net with GAN-guided training and perceptual loss, which learns to invert the effects of respiratory motion on image quality.

If this is right

Abdominal MRI can yield usable diagnostic images without requiring breath holds, sedation, or respiratory gating.
The same model handles both artificially added motion and naturally occurring patient breathing artifacts.
Post-acquisition correction becomes feasible, avoiding increases in total scan duration.
Improved image clarity supports better tissue characterization and lesion localization in standard abdominal protocols.

Where Pith is reading between the lines

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

The architecture could be tested on motion correction tasks in other body regions or imaging modalities such as cardiac MRI.
Deployment in scanner software might reduce the rate of repeat scans caused by motion.
Performance on pediatric or obese patient cohorts, where motion patterns differ, would test broader applicability.

Load-bearing premise

The training examples of synthetic and authentic motion artifacts capture the range of patterns that appear in new clinical abdominal MRI scans.

What would settle it

A quantitative comparison on a large set of previously unseen real clinical abdominal MRI scans where the corrected images show no measurable gain in sharpness, artifact reduction, or diagnostic readability over the original motion-affected images.

Figures

Figures reproduced from arXiv: 1906.09745 by Hing-Chiu Chang, Ka-Wai Kwok, Kit-Hang Lee, Qi Dou, Shihui Chen, Wenhao Jiang, Yui-Lun Ng, Zhiyu Liu.

**Figure 3.** Figure 3: Comparison of our proposed model with other three well-known deep learning models in motion correction. Residual motion images were calculated by subtracting the target patches from the motion-affected or model output pictures. Darker residual motion images indicate better correction results [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗

**Figure 4.** Figure 4: Quantitative comparisons among our proposed model and other deep learning models, regarding to FSIM (Upper left), SSIM (Upper right), IQI (Lower left) and PSNR (Lower right). Such artifacts were induced under the FSE, GRE sequences, and the scans with simulated motion. 3.3 Experimental Results The results of the proposed model and other deep-learning-based methods are compared qualitatively (in [PITH_FUL… view at source ↗

read the original abstract

Abdominal magnetic resonance imaging (MRI) provides a straightforward way of characterizing tissue and locating lesions of patients as in standard diagnosis. However, abdominal MRI often suffers from respiratory motion artifacts, which leads to blurring and ghosting that significantly deteriorate the imaging quality. Conventional methods to reduce or eliminate these motion artifacts include breath holding, patient sedation, respiratory gating, and image post-processing, but these strategies inevitably involve extra scanning time and patient discomfort. In this paper, we propose a novel deep-learning-based model to recover MR images from respiratory motion artifacts. The proposed model comprises a densely connected U-net with generative adversarial network (GAN)-guided training and a perceptual loss function. We validate the model using a diverse collection of MRI data that are adversely affected by both synthetic and authentic respiration artifacts. Effective outcomes of motion removal are demonstrated. Our experimental results show the great potential of utilizing deep-learning-based methods in respiratory motion correction for abdominal MRI.

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.

Desk Editor's Note private letter to a colleague

Straight application of dense U-Net plus GAN to MRI motion correction, but the abstract gives no numbers so the performance claim stays untestable.

read the letter

The core of this paper is taking a densely connected U-Net, training it with a GAN and perceptual loss, and pointing it at respiratory motion artifacts in abdominal MRI. The components are all established, so the work is an application rather than a new framework. They do test on both synthetic motion and real patient scans, which is a reasonable step beyond pure simulation and shows they tried to address the practical setting. That part is straightforward and worth noting. The main weakness is that the abstract supplies no quantitative results at all—no PSNR, SSIM, Dice scores, no baseline comparisons, no ablation on the GAN term or the dense connections, and no mention of statistical tests. Without those, it is impossible to judge whether the method actually removes artifacts better than simpler registration or filtering approaches or whether it generalizes beyond the training distribution. The central assumption that synthetic-plus-authentic data covers the range of clinical breathing patterns is left unexamined in the provided material. The paper is aimed at the small group of researchers already working on deep-learning fixes for MRI artifacts. A reader in that niche might pick up implementation details from the full text, but anyone outside it or anyone needing reproducible evidence will find little to use. The citation pattern looks standard for the subfield and does not appear to hide prior work. Because the clinical problem is real and the method is at least coherent, this deserves a serious referee rather than an immediate desk reject, though the review will almost certainly focus on demanding the missing metrics and controls.

Referee Report

2 major / 1 minor

Summary. The manuscript proposes a novel deep-learning model comprising a densely connected U-Net, GAN-guided training, and a perceptual loss function to recover abdominal MR images from respiratory motion artifacts. It validates the approach on a diverse collection of MRI data affected by both synthetic and authentic respiration artifacts and claims that effective motion removal is demonstrated, highlighting the potential of deep-learning methods for this task without requiring additional scan time or patient discomfort.

Significance. If supported by rigorous quantitative evaluation, the work could be significant for clinical abdominal MRI by offering a post-processing alternative to breath-holding, sedation, or gating. The architectural choices (dense U-Net + GAN + perceptual loss) build on established image-restoration techniques and address a practical problem in a high-volume imaging modality.

major comments (2)

[Abstract] Abstract: The central claim that 'effective outcomes of motion removal are demonstrated' is unsupported by any quantitative metrics (e.g., PSNR, SSIM, MAE), error bars, ablation studies, baseline comparisons, or statistical tests. Without these, the data-to-claim link cannot be evaluated and the soundness of the empirical contribution remains unclear.
[Abstract / Validation] Validation description (abstract and results): The training distribution (synthetic plus authentic artifacts) is asserted to enable generalization, but no details on data composition, train/test split, or out-of-distribution testing are provided to substantiate that the model recovers unseen clinical respiratory patterns rather than memorizing the training artifacts.

minor comments (1)

[Abstract] The abstract would benefit from explicit mention of the loss-function weighting coefficients and any hyperparameter choices that are treated as free parameters.

Simulated Author's Rebuttal

2 responses · 0 unresolved

Thank you for the constructive feedback on our manuscript. We appreciate the emphasis on rigorous quantitative evaluation and validation details. We address the major comments point by point below and will incorporate revisions to strengthen the empirical support.

read point-by-point responses

Referee: [Abstract] Abstract: The central claim that 'effective outcomes of motion removal are demonstrated' is unsupported by any quantitative metrics (e.g., PSNR, SSIM, MAE), error bars, ablation studies, baseline comparisons, or statistical tests. Without these, the data-to-claim link cannot be evaluated and the soundness of the empirical contribution remains unclear.

Authors: We agree that the abstract's qualitative claim would benefit from quantitative backing to allow proper evaluation. The manuscript demonstrates outcomes primarily through visual comparisons on synthetic and authentic artifacts. To address this, we will revise the manuscript to add quantitative metrics including PSNR, SSIM, MAE, error bars, ablation studies, baseline comparisons, and statistical tests in the results section. revision: yes
Referee: [Abstract / Validation] Validation description (abstract and results): The training distribution (synthetic plus authentic artifacts) is asserted to enable generalization, but no details on data composition, train/test split, or out-of-distribution testing are provided to substantiate that the model recovers unseen clinical respiratory patterns rather than memorizing the training artifacts.

Authors: The abstract provides a high-level summary of the validation data. We acknowledge that additional specifics are needed to support claims of generalization. In the revision, we will expand the methods and results sections with details on data composition, train/test splits, and out-of-distribution testing to clarify how the model handles unseen clinical patterns. revision: yes

Circularity Check

0 steps flagged

No derivation chain; empirical ML model shows no circularity

full rationale

The paper proposes and empirically validates a Dense U-Net + GAN model for respiratory motion correction in abdominal MRI. No equations, derivations, or first-principles predictions are present in the abstract or described claims. The central result is experimental performance on held-out synthetic and authentic artifact data, which is standard non-circular validation for ML methods and does not reduce to self-definition, fitted inputs renamed as predictions, or self-citation chains. No load-bearing self-citations or ansatzes are invoked. The derivation chain is absent, so the paper is self-contained against its benchmarks with score 0.

Axiom & Free-Parameter Ledger

2 free parameters · 1 axioms · 0 invented entities

The claim rests on empirical success of a neural network whose weights are learned from data. This introduces many free parameters in architecture and loss weighting. The key domain assumption is that the collected training artifacts represent real clinical variability.

free parameters (2)

U-Net and dense connection hyperparameters
Layer counts, growth rates, and skip-connection densities are chosen or tuned during model development.
GAN and perceptual loss weighting coefficients
Relative weights between adversarial, perceptual, and reconstruction losses are adjusted to achieve reported performance.

axioms (1)

domain assumption Training distribution of synthetic and authentic artifacts matches real clinical respiratory motion variability
Validation on the described collection is taken to support generalization to unseen scans.

pith-pipeline@v0.9.0 · 5717 in / 1191 out tokens · 35939 ms · 2026-05-25T17:19:00.896608+00:00 · methodology

discussion (0)

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

Removing Motion Artifact in MRI by Using a Perceptual Loss Driven Deep Learning Framework
cs.CV 2026-04 unverdicted novelty 5.0

PERCEPT-Net uses motion perceptual loss in a residual U-Net with attention and multi-scale modules to remove MRI motion artifacts more effectively than prior methods on clinical data.

Reference graph

Works this paper leans on

16 extracted references · 16 canonical work pages · cited by 1 Pith paper · 1 internal anchor

[1]

Journal of Magnetic Resonance Imaging 42(4), 887-901 (2015)

Zaitsev, M., Maclaren, J., Herbst, M.: Motion artifacts in MRI: a complex problem with many partial solutions. Journal of Magnetic Resonance Imaging 42(4), 887-901 (2015)

work page 2015
[2]

P., Rockwell, D

Malviya, S., Voepel-Lewis, T., Eldevik, O. P., Rockwell, D. T., Wong, J. H., Tait, A. R.: Sedation and general anaesthesia in children undergoing MRI and CT: adverse events and outcomes. British journal of anaesthesia 84(6), 743-748 (2000)

work page 2000
[3]

Journal of Physics C: Solid State Physics 10(3), L55 (1977)

Mansfield, P.: Multi-planar image formation using NMR spin echoes. Journal of Physics C: Solid State Physics 10(3), L55 (1977)

work page 1977
[4]

P.: Sparse MRI: The application of compressed sensing for rapid MR imaging

Lustig, M., David D., John M. P.: Sparse MRI: The application of compressed sensing for rapid MR imaging. Magnetic Resonance in Medicine: An Official Journal of the Interna-tional Society for Magnetic Resonance in Medicine 58(6), 1182-1195 (2007)

work page 2007
[5]

Journal of Magnetic Resonance Imaging 40(1), 13-25 (2014)

Zhang, T., Chowdhury, S., Lustig, M., et al.: Clinical performance of contrast enhanced abdominal pediatric MRI with fast combined parallel imaging compressed sensing recon-struction. Journal of Magnetic Resonance Imaging 40(1), 13-25 (2014)

work page 2014
[6]

L., McNamara, M

Ehman, R. L., McNamara, M. T., Pallack, M., Hricak, H., Higgins, C. B.: Magnetic reso-nance imaging with respiratory gating: techniques and advantages. American journal of Roentgenology 143(6), 1175-1182 (1984)

work page 1984
[7]

Magnetic resonance in medicine 70(6), 1608-1618 (2013)

Loktyushin, A., Nickisch, H., Pohmann, R., Schölkopf, B.: Blind retrospective motion cor-rection of MR images. Magnetic resonance in medicine 70(6), 1608-1618 (2013)

work page 2013
[8]

L., Onishi, H., Motosugi, U.: Method for motion artifact reduction using a convolutional neural network for dynamic contrast enhanced MRI of the liver

Tamada, D., Kromrey, M. L., Onishi, H., Motosugi, U.: Method for motion artifact reduction using a convolutional neural network for dynamic contrast enhanced MRI of the liver. arXiv preprint arXiv:1807.06956 (2018)

work page arXiv 2018
[9]

arXiv preprint arXiv:1809.06276 (2018)

Armanious, K., Küstner, T., Nikolaou, K., Gatidis, S., Yang, B.: Retrospective correction of Rigid and Non-Rigid MR motion artifacts using GANs. arXiv preprint arXiv:1809.06276 (2018)

work page arXiv 2018
[10]

J., Kirk, S., Lee, Y., et al.: Radiology Data from The Cancer Genome Atlas Liver Hepatocellular Carcinoma [TCGA-LIHC] collection

Erickson, B. J., Kirk, S., Lee, Y., et al.: Radiology Data from The Cancer Genome Atlas Liver Hepatocellular Carcinoma [TCGA-LIHC] collection. The Cancer Imaging Ar-chive (2016). http://doi.org/10.7937/K9/TCIA.2016.IMMQW8UQ

work page doi:10.7937/k9/tcia.2016.immqw8uq 2016
[11]

IEEE transactions on medical imaging 37(12), 2663-2674 (2018)

Li, X., Chen, H., Qi, X., Dou, Q., Fu, C.W., Heng, P.A.: H-denseunet: Hybrid densely con-nected unet for liver and tumor segmentation from ct volumes. IEEE transactions on medical imaging 37(12), 2663-2674 (2018)

work page 2018
[12]

In: Proceedings of the IEEE conference on computer vision and pattern recognition, pp

Ledig, C., Theis, L., Huszár, F., et al.: Photo-realistic single image super-resolution using a generative adversarial network. In: Proceedings of the IEEE conference on computer vision and pattern recognition, pp. 4681-4690 (2017)

work page 2017
[13]

Keras, https://github.com/fchollet/keras, last accessed 2019/04/02

work page 2019
[14]

Adam: A Method for Stochastic Optimization

Kingma, Diederik P., Jimmy Ba.: Adam: A method for stochastic optimization. arXiv pre-print arXiv:1412.6980 (2014)

work page internal anchor Pith review Pith/arXiv arXiv 2014
[15]

9351, pp

LNCS, vol. 9351, pp. 234-241. Springer, Cham (2015). https://doi.org/10.1007/978-3-319-24574-4_28

work page doi:10.1007/978-3-319-24574-4_28 2015
[16]

In: Proceedings of the IEEE conference on computer vision and pattern recognition, pp

Isola, P., Zhu, J.Y., Zhou, T., Efros, A.A.: Image-to-image translation with conditional ad-versarial networks. In: Proceedings of the IEEE conference on computer vision and pattern recognition, pp. 1125-1134 (2017)

work page 2017

[1] [1]

Journal of Magnetic Resonance Imaging 42(4), 887-901 (2015)

Zaitsev, M., Maclaren, J., Herbst, M.: Motion artifacts in MRI: a complex problem with many partial solutions. Journal of Magnetic Resonance Imaging 42(4), 887-901 (2015)

work page 2015

[2] [2]

P., Rockwell, D

Malviya, S., Voepel-Lewis, T., Eldevik, O. P., Rockwell, D. T., Wong, J. H., Tait, A. R.: Sedation and general anaesthesia in children undergoing MRI and CT: adverse events and outcomes. British journal of anaesthesia 84(6), 743-748 (2000)

work page 2000

[3] [3]

Journal of Physics C: Solid State Physics 10(3), L55 (1977)

Mansfield, P.: Multi-planar image formation using NMR spin echoes. Journal of Physics C: Solid State Physics 10(3), L55 (1977)

work page 1977

[4] [4]

P.: Sparse MRI: The application of compressed sensing for rapid MR imaging

Lustig, M., David D., John M. P.: Sparse MRI: The application of compressed sensing for rapid MR imaging. Magnetic Resonance in Medicine: An Official Journal of the Interna-tional Society for Magnetic Resonance in Medicine 58(6), 1182-1195 (2007)

work page 2007

[5] [5]

Journal of Magnetic Resonance Imaging 40(1), 13-25 (2014)

Zhang, T., Chowdhury, S., Lustig, M., et al.: Clinical performance of contrast enhanced abdominal pediatric MRI with fast combined parallel imaging compressed sensing recon-struction. Journal of Magnetic Resonance Imaging 40(1), 13-25 (2014)

work page 2014

[6] [6]

L., McNamara, M

Ehman, R. L., McNamara, M. T., Pallack, M., Hricak, H., Higgins, C. B.: Magnetic reso-nance imaging with respiratory gating: techniques and advantages. American journal of Roentgenology 143(6), 1175-1182 (1984)

work page 1984

[7] [7]

Magnetic resonance in medicine 70(6), 1608-1618 (2013)

Loktyushin, A., Nickisch, H., Pohmann, R., Schölkopf, B.: Blind retrospective motion cor-rection of MR images. Magnetic resonance in medicine 70(6), 1608-1618 (2013)

work page 2013

[8] [8]

L., Onishi, H., Motosugi, U.: Method for motion artifact reduction using a convolutional neural network for dynamic contrast enhanced MRI of the liver

Tamada, D., Kromrey, M. L., Onishi, H., Motosugi, U.: Method for motion artifact reduction using a convolutional neural network for dynamic contrast enhanced MRI of the liver. arXiv preprint arXiv:1807.06956 (2018)

work page arXiv 2018

[9] [9]

arXiv preprint arXiv:1809.06276 (2018)

Armanious, K., Küstner, T., Nikolaou, K., Gatidis, S., Yang, B.: Retrospective correction of Rigid and Non-Rigid MR motion artifacts using GANs. arXiv preprint arXiv:1809.06276 (2018)

work page arXiv 2018

[10] [10]

J., Kirk, S., Lee, Y., et al.: Radiology Data from The Cancer Genome Atlas Liver Hepatocellular Carcinoma [TCGA-LIHC] collection

Erickson, B. J., Kirk, S., Lee, Y., et al.: Radiology Data from The Cancer Genome Atlas Liver Hepatocellular Carcinoma [TCGA-LIHC] collection. The Cancer Imaging Ar-chive (2016). http://doi.org/10.7937/K9/TCIA.2016.IMMQW8UQ

work page doi:10.7937/k9/tcia.2016.immqw8uq 2016

[11] [11]

IEEE transactions on medical imaging 37(12), 2663-2674 (2018)

Li, X., Chen, H., Qi, X., Dou, Q., Fu, C.W., Heng, P.A.: H-denseunet: Hybrid densely con-nected unet for liver and tumor segmentation from ct volumes. IEEE transactions on medical imaging 37(12), 2663-2674 (2018)

work page 2018

[12] [12]

In: Proceedings of the IEEE conference on computer vision and pattern recognition, pp

Ledig, C., Theis, L., Huszár, F., et al.: Photo-realistic single image super-resolution using a generative adversarial network. In: Proceedings of the IEEE conference on computer vision and pattern recognition, pp. 4681-4690 (2017)

work page 2017

[13] [13]

Keras, https://github.com/fchollet/keras, last accessed 2019/04/02

work page 2019

[14] [14]

Adam: A Method for Stochastic Optimization

Kingma, Diederik P., Jimmy Ba.: Adam: A method for stochastic optimization. arXiv pre-print arXiv:1412.6980 (2014)

work page internal anchor Pith review Pith/arXiv arXiv 2014

[15] [15]

9351, pp

LNCS, vol. 9351, pp. 234-241. Springer, Cham (2015). https://doi.org/10.1007/978-3-319-24574-4_28

work page doi:10.1007/978-3-319-24574-4_28 2015

[16] [16]

In: Proceedings of the IEEE conference on computer vision and pattern recognition, pp

Isola, P., Zhu, J.Y., Zhou, T., Efros, A.A.: Image-to-image translation with conditional ad-versarial networks. In: Proceedings of the IEEE conference on computer vision and pattern recognition, pp. 1125-1134 (2017)

work page 2017