review: noise and artifact reduction for mri using deep learningreview: noise and artifact reduction...

Review: Noise andartifact reduction forMRI using deeplearningDaiki Tamada, Department of Radiology, University of Yamanashi, Chuo, Yamanashi, 409-3898, Japan

Submitted to Magnetic Resonance in Medical Sciences on 2/27/2020

For several years, numerous attemptshave been made to reduce noise andartifacts in MRI. Although there have

been many successful methods to addressthese problems, practical implementation forclinical images is still challenging becauseof its complicated mechanism. Recently,deep learning received considerable attention,emerging as a machine learning approach indelivering robust MR image processing. Thepurpose here is therefore to explore furtherand review noise and artifact reduction usingdeep learning for MRI.

Introduction

MRI suffers from various kinds of noise and artifactsbecause of the nature of the NMR signal detectionand spatial encoding scheme1. There are severalreasons that MRI is sensitive to these undesirableerrors compared to other modalities. For instance,hardware-induced errors often stem from a compli-cated acquisition scheme that depends on a reliablespectrometer, RF coils, and magnetic fields. In ad-

dition, a long acquisition time is also an essentialreason. Body motion, including respiratory and car-diac motion and B0 drift during the scan, can degradethe image quality as well. Notwithstanding the in-tense dedication of MR vendors to improve thesefundamental problems, noise and artifact are stillnon-negligible degradations to MR images.

Generally, thermal noise is derived from the hu-man body and MRI system itself, such as RF coils,transmission lines, receiver circuits, and so on. More-over, external RF noise can be a source of the noise.To make matters worse, some sequences such asdiffusion-weighted imaging suffer from low signal-to-noise ratio (SNR) because of their weak signalintensity. To improve the SNR of MR images, thesignal averaging approach is widely used, althoughit requires a longer scan time as the number of ex-citation (NEX) increases. To address these hitches,many advanced filtering methods have been devel-oped using the underlying assumption of MR imagessuch as noise distribution and structures 2. How-ever, the violation of the assumption often inducesthe undesirable change of texture. In addition tothis, some algorithms utilize iterative reconstruction,which yield long computation time.

HowToTeX.com • Two column article template page 1 of 9

arX

iv:2

002.

1288

9v1

[ee

ss.I

V]

28

Feb

2020

Artifacts in MRI are caused by various reasonssuch as external errors and inappropriate spatialencoding. For example, it is widely known that un-dersampling in Cartesian imaging results in aliasingartifacts. In this case, compressed sensing (CS) recon-struction is generally used to remove such artifacts.Streak artifact is caused by undersampling, motion,and temporal intensity change owing to contrastagent in radial sampling. Advanced imaging tech-niques, combined with golden angle trajectory andCS reconstruction, have shown excellent performanceagainst these problems3-7. Lack of high-frequencysignal leads to Gibbs artifact, which is also calledtruncation, or ringing artifact. Simple k-space do-main filtering is often used although it may causeblurring. Body imaging suffers from motion arti-facts that arise from respiratory and cardiac motion.Respiratory triggering8, fast acquisition9, and ra-dial sampling3 are clinically available to suppressartifacts.

Further, retrospective correction10 approacheshave been proposed, although it is still in the researchstage. Inhomogeneous B0 field or off-resonance effectresults in banding artifacts, which are often seenin balanced steady-state free precession. The de-velopment of robust and practical methods is stillchallenging despite the aforementioned efforts to re-move artifacts.

Recently, a deep learning approach, which enablesfeature extraction and complicated nonlinear imageprocessing, is gaining traction to reduce noise andartifacts in MRI11. Deep learning (DL) is a success-ful machine learning technique based on the neuralnetwork used for segmentation, lesion detection, andreconstruction for MRI. In this study, we review thedeep learning techniques and applications for reduc-ing noise and artifacts in MRI. First, we discussthe deep learning architecture used for noise andartifact reduction. The following section describesstate-of-the-art applications.

1 Deep Learning Architecture

Many DL networks have been proposed for noise andartifact reduction. Convolutional neural networks(CNN), which has a small receptive field, is widelyused for noise reduction because Gaussian noise isincoherent and position-independent. However, onlythe networks with a larger receptive field are oftenadopted to improve artifacts because most of theartifacts are distributed globally. In this section, weprovide an overview of deep learning architecture

commonly used for noise and artifact reduction.Single-scale CNN (SCNN), which is widely used

in the various fields shown in Fig. 1, is efficientin improving noisy images12, 13. SCNN consists ofmore than one convolution layer without a down- orup-sampling structure, such as pooling and stride.In most cases, residual components such as noise orartifact images are used as the output because it iswell-known that the manifold of the residual com-ponents has a simpler structure compared to thatof the clean image components14, resulting in fasterconversion and better generalization of the network.Denoising convolutional neural networks (DnCNN)is also a thriving network used for image denoisingthrough a process of learning residual componentsof images, which is noise. Besides these, batch nor-malization is used to improve training time for thenetwork. ResNet is commonly used to achieve a com-plex and deeper structure of the network to overcomethe vanishing gradient problem. ResNet is also usedfor artifact reduction because it has a large receptivefield that covers whole input images owing to a largenumber of layers 15.

Autoencoders can extract clean features fromnoisy images, similar to the principal componentanalysis procedure, which is used to extract signifi-cant information from high-dimensional datasets 16.Given that Gondara first developed an autoencoder-based denoising filter for X-ray images17, many stud-ies have been proposed for CT, MRI, and others,recently18, 19. Autoencoder, which has a differentmechanism compared to SCNN, compresses the en-tire image information into a low-dimensional struc-ture to effectively learn underlying manifold. Unfor-tunately, autoencoders may lose important informa-tion such as edge or fine structure, if the input imageis not a redundant representation.

U-net, often used for image segmentation, has alarge receptive field by utilizing multi-scale features20.This network has asymmetric up- and down-samplingstructures with a skip connection, which concatenateseach layer, as shown in Fig. 2. The skip connectioncontributes to capturing localized features such asthe fine structure of images. For the artifact reduc-tion, U-net is the most popular network because thelarge receptive field is generally required to extractthe artifact that is commonly observed globally. Thestudy by Lee et al. implied that U-net gave a bet-ter de-aliasing performance as compared to SCNN14,although a relatively large number of trainable pa-rameters is required for U-net.

Generative adversarial networks (GAN) approachis a promising learning technique for removing noise


Figure 1: Example of a single-scale CNN structure consisting of multiple convolution layers. Each convolution layerhas a set of filters, which is followed by activation function such as ReLU, to extract features of inputimages. The residual learning approach is usually adopted for noise and artifact reduction. Noise orartifact reduced images are calculated by subtracting the output of the network from input images.

and artifacts21. GAN consists of two separate net-works: generator and discriminator. Generator isthe network that creates the output images, noise-or artifact-free images, whereas discriminator actsas a classifier to determine whether the generatedoutput is real or fake. Feedback from the discrim-inator contributes to updating the network of thegenerator to create convincing fake images to fool thediscriminator. U-net is widely used for the generator.

2 Applications

Many applications have been proposed using theabove successful DL networks for noise and artifactreduction. Although most parts of them are stillresearch stage, promising results were demonstratedwith a limited size of datasets. In this section, weintroduce state-of-art applications shown in Tab. 1.

Some DL networks focus on denoising for brain MRimaging. Bermudez et al. proposed an autoencoderwith skip connections for T1-weighted (T1w) imagingof the brain19. The developed network, trained with528 T1w images, significantly improved the imagequality based on PSNR analysis. DnCNN is alsowidely used for denoising of brain MR images22, 23.Denoising using DL is used not only for the im-provement of image quality but also for reducingscan time. Currently, MR images with higher res-olutions are required to achieve better diagnosticperformance. However, improvement of patient ex-perience, particularly scan time, in the MR scanneris also an urgent priority. To overcome these conflict-ing problems, many DL-based denoising approaches

have been developed. Kawamura et al. proposedthe DnCNN-based approach for denoising multi-shotdiffusion-weighted imaging for the brain. The studyshowed that highly swift DWI could be achieved byreducing the number of excitations (NEX), as shownin Fig. 3. Kidoh et al. proposed the SCNN network,which can be used for multiple sequences such asT1w, T2w, FLAIR, and MPRAGE24. The proposednetwork utilizes a discrete cosine transform (DCT)layer to separate zero and high-frequency compo-nents. The separated high-frequency componentsgo through 22 convolution layers, whereas the zero-frequency component is connected to the last layer.This structure may contribute to robust denoisingperformance because it doesnt use contrast informa-tion of images. Imaging can be achieved at a pace ofmore than twofold by using the proposed method.

Moreover, Arterial spin labeling (ASL) perfusionimaging, which generally suffers from low SNR, withDL denoising is also a promising application25, 26.Xie et al. developed the DL denoising approach forASL using a wide interface network (WIN)25. TheWIN is a residual learning CNN with a large size ofconvolution kernel and filters. Quantitative analysisbased on SNR suggested that the proposed methodcan reduce the scan time for ASL by reducing NEX.Kim et al. proposed a network consisting of two sep-arate pathways for extracting multi-scale features26.The pathway with convolution kernel, batch normal-ization, and ReLU is used for the local structure ofimages. Meanwhile, the other pathway uses a dilatedpathway to increase the receptive field in order toextract large-scale information. Evaluation by a ra-diologist implied that the ASL images with low NEX


Figure 2: Overview of U-net, which has up- and down-sampling structure - the same as autoencoder - used for noiseand artifact reduction. Skip connections by copying features from earlier layers into later layers are usedto preserve high-frequency information, which is often lost in the down-sampling step. This network iscommonly used for removing artifact reduction, and requires a larger receptive field compared to noisereduction.

were significantly improved by using this proposedCNN method.

Further, Streak artifact reduction is one of themost successful applications using DL27. Thereare many studies for reducing streak artifact forcomputed tomography (CT) images because streakartifact on CT is a very common undesirable im-age degradation28-30. Recently, radial samplingfor abdominal MR imaging is getting attention be-cause of its robustness against respiratory inducedmotion3, 5, 6. In the case of radial sampling, streakartifact can be observed in MR images as well as CT.Han et al. developed U-net to remove streak artifactsfor the brain and abdominal MR images; a networkadopted residual learning with pre-training using CTand MRI datasets. This study indicated that theDL-based method could be superior to conventionalcompressed sensing algorithms.

Several studies of retrospective and blinded mo-tion artifact correction techniques for the brain31-33,c-spine34, liver35, 36, upper abdomen37, and cardiac38

imaging have been proposed. Pawar et al. proposedthe motion artifact reduction technique using theU-net for MPRAGE images of the brain. Evalua-tion based on naturalness image quality evaluator(NIQE)39, which is reference-free image quality met-rics, revealed the proposed method improved themotion artifact. Tamada et al. also proposed amotion artifact reduction for DCE-MRI of the liverusing DnCNN-based network with multi-channel in-put and output. A training dataset was generatedby simulating the respiratory motion. The results

indicated the DL approach possessed the ability toremove artifacts, as show in Fig. 4. It is a challengeto acquire artifact-free images under the influenceof cardiac motion. Oksuz et al. proposed the recon-struction technique for cardiac imaging by combiningthe 3D CNN and the recurrent convolutional neuralnetwork (RCNN)40 used for 2D spatiotemporal imag-ing. The 3D CNN is adopted to detect the corruptedk-space lines by the motion whereas RCNN removesthe motion artifact coming from the corrupted lines.

Besides, some applications are used for re-moving aliasing artifact alternatives to CSreconstruction41-43. CS reconstruction, which is afast MRI technique using random undersamplingand redundancy of MR images, generally requireslong computation time. DL approaches can con-tribute to the reduction of computation cost for theundersampled datasets. A de-aliasing method usingthe U-net for undersampled k-space was proposed toachieve efficient reconstruction by Hyun et al. 41.Yang et al. proposed a GAN-based reconstructionmethod for undersampled images42. Recently,a model-based DL network, which has unrolledarchitecture, was proposed to reduce trainableparameters of the network43.

The DL-based approach is also useful in solv-ing other complicated problems such as removingGibbs44-47, metal48-50, and banding artifact51. Zhanget al. proposed the CNN-based Gibbs artifact, whichcan be an alternative to conventional k-space domainfiltering methods45. The results indicated that theproposed method removed artifacts without addi-


Figure 3: Specimen images of denoising brain DWI using DnCNN, the Noise2Noise approach. This network includes17 convolution layers with batch normalization and ReLU. Multi-shot DWI images (a) acquired with NEXof 1 were denoised using the network. PSNR analysis between the denoised and high-NEX images indicatedthat the network successfully reduced the noise component.

tional blurring, which is often observed in traditionalmethods. Kwon proposed metal artifact reductionusing the U-net and bipolar readout imaging to im-prove blurring and distortion caused by a strongoff-resonance effect48. Seo et al. developed the CNNto correct metal artifact of MR images acquired usingslice encoding for metal artifact correction49.

3 Discussion

In this paper, we summarized DL networks and itsapplications used for noise and artifact reduction.Typical networks including SCNN, U-net, autoen-coder, GAN architectures, which have characteristicsof structure different from each other, were explained.Further, we reviewed applications using these net-works. The studies reviewed here demonstrate theexcellent performance of noise and artifact reductionfor MRI. These results revealed that DL-base ap-proaches can be used to remove complicated artifactssuch as aliasing, streaking, and so on, which are stillchallenging problems using conventional methods.

Although many promising techniques have beenproposed, challenges of removing noise and artifactusing DL exist. To begin with, it is challenging toprepare training datasets for some cases. For noisereduction, images with a large number of NEX areoften used as clean datasets. However, motions, suchas respiratory, cardiac, and CSF pulsation, duringa scan can lead to blurring and undesirable imagedegradation. Noise2Noise approach could be one ofthe practical solutions for noise reduction52. Nonethe-

less, datasets generated by simulation are generallyused for artifact reduction. In this case, the discrep-ancy between the simulated and acquired datasetsleads to poor filtering performance.

Furthermore, it is difficult to conduct appropri-ate image quality evaluation for noise and artifactreduction. MSE, PSNR, and SSIM are widely usedto evaluate the image quality, particularly for noisereduction networks. Conversely, there is difficultyin using these conventional metrics for artifact re-duction because it is challenging to prepare groundtruth datasets that have no artifact images. Toevaluate the image quality without ground truth,reference-free metrics such as Blind/ReferencelessImage Spatial Quality Evaluator53 and NIQE39 are areasonable alternative to conventional metrics. How-ever, clinical evaluation is essential for practical usebecause these metrics offer excellent performance fornatural images. It is essential to pay attention to thecontrast, texture, and fine structure after DL-basedfiltering because inappropriate training, such as theuse of a small number of datasets and wrong noiseor artifact modeling, could lead to a critical changeof contexture of images.

4 Conclusion and Outlook

In this study, we reviewed DL-based methods toreduce noise and artifact for MR images. Because ofthe recent advancement of deep learning technology, awide variety of techniques are proposed to serve thesepurposes. However, even though DL-based methods


Figure 4: Sample depictions of aliasing reduction of the liver using multi-channel DnCNN, which have seven layersof convolution layers. Multi-contrast images obtained from DCE-MRI were used for the input of thenetwork. MR images with respiratory artifact and blurring were significantly improved by via this network.The acquisition was implemented using the multi-phase 3D T1-weighted spoiled gradient-echo sequencewith a dual-echo bipolar readout.

are getting popular in the MR community owingto its excellent performance and low computationalcost, there are still issues to be solved for practicaluse. Nevertheless, some applications are overcomingthese setbacks and are about to come onto the clinicalsite as products from MR vendors. The DL-basedmethod will therefore be widely used in a range ofMR applications shortly.

5 References

1. Bellon EM, Haacke EM, Coleman PE, SaccoDC, Steiger DA, Gangarosa RE. MR artifacts: areview. American Journal of Roentgenology 1986;147:1271-1281.

2. Balafar M. Review of noise reducing algorithmsfor brain MRI images. methods 2012; 10:11.

3. Chandarana H, Feng L, Block TK, et al. Free-breathing contrast-enhanced multiphase MRI of theliver using a combination of compressed sensing, par-allel imaging, and golden-angle radial sampling 2013;48.

4. Chandarana H, Block KT, Winfeld MJ, etal. Free-breathing contrast-enhanced T1-weightedgradient-echo imaging with radial k-space samplingfor paediatric abdominopelvic MRI 2014; 24:320-326.

5. Chandarana H, Block TK, Rosenkrantz AB,et al. Free-breathing radial 3D fat-suppressed T1-weighted gradient echo sequence: a viable alternativefor contrast-enhanced liver imaging in patients un-able to suspend respiration 2011; 46:648-653.

6. Feng L, Grimm R, Block KT, et al. Golde-

nangle radial sparse parallel MRI: combination ofcompressed sensing, parallel imaging, and goldenan-gle radial sampling for fast and flexible dynamicvolumetric MRI 2014; 72:707-717.

7. Feng L, Axel L, Chandarana H, Block KT, Sod-ickson DK, Otazo RJMrim. XDGRASP: goldenangleradial MRI with reconstruction of extra motionstatedimensions using compressed sensing 2016; 75:775-788.

8. Chavhan GB, Babyn PS, Vasanawala SSJR.Abdominal MR imaging in children: motion compen-sation, sequence optimization, and protocol organi-zation 2013; 33:703-719.

9. Zhang T, Chowdhury S, Lustig M, et al. Clin-ical performance of contrast enhanced abdominalpediatric MRI with fast combined parallel imagingcompressed sensing reconstruction 2014; 40:13-25.

10. Cheng JY, Alley MT, Cunningham CH,Vasanawala SS, Pauly JM, Lustig MJMrim. Non-rigid motion correction in 3D using autofocusingwithlocalized linear translations 2012; 68:1785-1797.

11. Lundervold AS, Lundervold A. An overview ofdeep learning in medical imaging focusing on MRI.Zeitschrift fr Medizinische Physik 2019; 29:102-127.

12. Kim J, Kwon Lee J, Mu Lee K. Accurateimage super-resolution using very deep convolutionalnetworks. Proceedings of the IEEE conference oncomputer vision and pattern recognition, 2016; 1646-1654.

13. Zhang K, Zuo W, Chen Y, Meng D, Zhang LJI-ToIP. Beyond a gaussian denoiser: Residual learningof deep cnn for image denoising 2017; 26:3142-3155.


14. Lee D, Yoo J, Ye JC. Deep residual learning forcompressed sensing MRI. Biomedical Imaging (ISBI2017), 2017 IEEE 14th International Symposium on,2017; 15-18.15. Ghodrati V, Shao J, Bydder M, et al. MRimage reconstruction using deep learning: evaluationof network structure and loss functions. Quantitativeimaging in medicine and surgery 2019; 9:1516.16. Vincent P, Larochelle H, Bengio Y, ManzagolP-A. Extracting and composing robust features withdenoising autoencoders. Proceedings of the 25thinternational conference on Machine learning, 2008;1096-1103.17. Gondara L. Medical image denoising using con-volutional denoising autoencoders. 2016 IEEE 16thInternational Conference on Data Mining Workshops(ICDMW), 2016; 241-246.18. Nishio M, Nagashima C, Hirabayashi S, et al.Convolutional auto-encoder for image denoising ofultra-low-dose CT. Heliyon 2017; 3:e00393.19. Bermudez C, Plassard AJ, Davis LT, NewtonAT, Resnick SM, Landman BA. Learning implicitbrain MRI manifolds with deep learning. MedicalImaging 2018: Image Processing, 2018; 105741L.20. Ronneberger O, Fischer P, Brox T. U-net: Con-volutional networks for biomedical image segmenta-tion. International Conference on Medical imagecomputing and computer-assisted intervention, 2015;234-241.21. Goodfellow I, Pouget-Abadie J, Mirza M, etal. Generative adversarial nets. Advances in neuralinformation processing systems, 2014; 2672-2680.22. Jiang D, Dou W, Vosters L, Xu X, Sun Y,Tan T. Denoising of 3D magnetic resonance imageswith multi-channel residual learning of convolutionalneural network. Japanese journal of radiology 2018;36:566-574.23. Rao GS, Srinivas B. De-noising of MRI BrainTumor image using Deep Convolutional Neural Net-work. Available at SSRN 3357284 2019.24. Kidoh M, Shinoda K, Kitajima M, et al. DeepLearning Based Noise Reduction for Brain MR Imag-ing: Tests on Phantoms and Healthy Volunteers.Magnetic Resonance in Medical Sciences 2019:mp.2019-0018.25. Xie D, Bai L, Wang Z. Denoising arterialspin labeling cerebral blood flow images using deeplearning. arXiv preprint arXiv:180109672 2018.26. Kim KH, Choi SH, Park S-H. Improving arte-rial spin labeling by using deep learning. Radiology2018; 287:658-666.27. Han Y, Yoo J, Kim HH, Shin HJ, Sung K,Ye JC. Deep learning with domain adaptation for

accelerated projectionreconstruction MR. Magneticresonance in medicine 2018; 80:1189-1205.28. Gjesteby L, Yang Q, Xi Y, et al. Reducingmetal streak artifacts in CT images via deep learning:Pilot results. The 14th international meeting on fullythree-dimensional image reconstruction in radiologyand nuclear medicine, 2017; 611-614.29. Han YS, Yoo J, Ye JC. Deep residuallearning for compressed sensing CT reconstructionvia persistent homology analysis. arXiv preprintarXiv:161106391 2016.30. Han Y, Ye JC. Framing U-Net via deep con-volutional framelets: Application to sparse-view CT.Ieee T Med Imaging 2018; 37:1418-1429.31. Pawar K, Chen Z, Shah NJ, Egan GF. Mo-conet: Motion correction in 3D MPRAGE images us-ing a convolutional neural network approach. arXivpreprint arXiv:180710831 2018.32. Duffy BA, Zhang W, Tang H, et al. Retrospec-tive correction of motion artifact affected structuralMRI images using deep learning of simulated motion2018.33. Johnson PM, Drangova M. Motion correctionin MRI using deep learning. ISMRM Scientific Meet-ing & Exhibition, Honolulu, 2018; 4098.34. Lee H, Ryu K, Nam Y, Lee J, Kim D-H. Reduc-tion of respiratory motion artifact in c-spine imagingusing deep learning: Is substitution of navigator pos-sible? ISMRM Scientific Meeting & Exhibition, 2018;2660.35. Tamada D, Kromrey M-L, Ichikawa S, OnishiH, Motosugi U. Motion Artifact Reduction Usinga Convolutional Neural Network for Dynamic Con-trast Enhanced MR Imaging of the Liver. MagneticResonance in Medical Sciences 2020; 19:64-76.36. Tamada D, Onishi H, Motosugi U. MotionArtifact Reduction in Abdominal MR Imaging usingthe U-NET Network. ICMRM and Scientific Meetingof KSMRM, Seoul, Korea, 2018; PP0311.37. Jiang W, Liu Z, Lee K-H, et al. Respiratorymotion correction in abdominal MRI using a denselyconnected U-Net with GAN-guided training. arXivpreprint arXiv:190609745 2019.38. Oksuz I, Clough J, Ruijsink B, et al. Detec-tion and Correction of Cardiac MRI Motion Arte-facts During Reconstruction from k-space. Interna-tional Conference on Medical Image Computing andComputer-Assisted Intervention, 2019; 695-703.39. Mittal A, Soundararajan R, Bovik AC. Makinga “completely blind” image quality analyzer. IEEESignal Processing Letters 2012; 20:209-212.40. Qin C, Schlemper J, Caballero J, Price AN,Hajnal JV, Rueckert D. Convolutional recurrent neu-


ral networks for dynamic MR image reconstruction.Ieee T Med Imaging 2018; 38:280-290.41. Hyun CM, Kim HP, Lee SM, Lee S, SeoJKJPim, biology. Deep learning for undersampledMRI reconstruction 2018.42. Yang G, Yu S, Dong H, et al. DAGAN: Deepde-aliasing generative adversarial networks for fastcompressed sensing MRI reconstruction. Ieee T MedImaging 2017; 37:1310-1321.43. Aggarwal HK, Mani MP, Jacob M. MoDL:Model-based deep learning architecture for inverseproblems. Ieee T Med Imaging 2018; 38:394-405.44. Zhang Q, Ruan G, Yang W, Zhao K, Wu EX,Feng Y. Gibbs-Ringing Artifact Reduction in MRIvia Machine Learning Using Convolutional NeuralNetwork. ISMRM 2018, 2018; 0429.45. Zhang Q, Ruan G, Yang W, et al. MRI Gibb-sringing artifact reduction by means of machine learn-ing using convolutional neural networks. Magneticresonance in medicine 2019; 82:2133-2145.46. Muckley MJ, Ades-Aron B, Papaioannou A,et al. Training a Neural Network for Gibbs andNoise Removal in Diffusion MRI. arXiv preprintarXiv:190504176 2019.47. Zhao X, Zhang H, Zhou Y, Bian W, ZhangT, Zou X. Gibbs-ringing artifact suppression withknowledge transfer from natural images to MR im-ages. Multimedia Tools and Applications 2019:1-23.48. Kwon K, Kim D, Park H. A Learning-BasedMetal Artifacts Correction Method for MRI Us-ing Dual-Polarity Readout Gradients and SimulatedData. International Conference on Medical Im-age Computing and Computer-Assisted Intervention,2018; 189-197.49. Seo S, Do WJ, Luu HM, Kim KH, Choi SH,Park SH. Artificial neural network for Slice Encod-ing for Metal Artifact Correction (SEMAC) MRI.Magnetic Resonance in Medicine 2019.50. Kim JW, Kwon K, Kim B, Park H. AttentionGuided Metal Artifact Correction in MRI using DeepNeural Networks. arXiv preprint arXiv:1910087052019.51. Kim KH, Park S-H. Artificial neural net-work for suppression of banding artifacts in balancedsteady-state free precession MRI. Magnetic resonanceimaging 2017; 37:139-146.52. Lehtinen J, Munkberg J, Hasselgren J, etal. Noise2noise: Learning image restoration withoutclean data. arXiv preprint arXiv:180304189 2018.53. Mittal A, Moorthy AK, Bovik AC. No-reference image quality assessment in the spatialdomain. IEEE Transactions on image processing2012; 21:4695-4708.


Table 1: Overview of DL studies for noise and artifact reduction

Purpose Authors Network

Denoising for T1-weightedbrain images

Camilo Bermudez, et al19 Autoencoder with skip connec-tions

Denoising for brain MRI Dongsheng Jiang, et al22 Multi-channel DnCNNDenoising for ASL Ki Hwan Kim et al26 Parallel two CNN with differ-

ent scale of resolutionDenoising for ASL Danfeng Xie, et al25 Wide Inference NetworkDenoising for T1-, T2-weighted, and FLAIR brainimages

Masafumi Kidoh, et al24 Single-scale CNN with DCT

Denoising for multi-shot DWIof the brain

Motohide Kawamura, et al54 DnCNN with Noise2Noise ap-proach

Streak artifact reduction forradial MRI

Yoseob Han, et al27 U-net

Motion artifact reduction forbrain MRI

Patricia Johnson, et al33 Single-scale CNN

Motion artifact reduction forDCE-MRI of the liver

Daiki Tamada, et al35 Multi-channel DnCNN

Motion artifact reduction forabdomen MRI

Wenhao Jiang, et al37 GAN using U-net as generatornetwork

Motion artifact reduction forcardiac MRI

Ilkay Oksuz, et al38 3DCNN and RCNN

Motion artifact reduction forbrain MRI

Ben A Duffy, et al32 GAN using HighRes3dNet asa generator

Motion artifact reduction forcervical spine MRI

Hongpyo Lee, et al34 U-net

Aliasing artifact reduction forundersampled MRI

Chang Min Hyun, et al41 U-net


Guang Yang, et al42 GAN using U-net as generatornetwork


Hemant Kumar Aggarwal, etal43

Model-based DL

Gibbs artifact reduction Qianqian Zhang, et al45 Single-scale CNNGibbs artifact reduction Xiaole Zhao, et al47 Enhanced Deep Super-

Resolution NetworkMetal artifact reduction Kinam Kwon, et al U-netMetal artifact reduction Jee Won Kim, et al50 U-netMetal artifact reduction forSEMAC

Sunghun Seo, et al49 U-net

Banding artifact reduction forbSSFP

Ki Hwan Kim, et al51 Multilayer perceptron


review: noise and artifact reduction for mri using deep learningreview: noise and artifact reduction...

Documents