Journal of Physics:

Conference Series

PAPER • OPEN ACCESS You may also like

- An adaptive parameter decoupling
A review on generative based methods for MRI algorithm-based image reconstruction
model (ADAIR) for rapid golden-angle
reconstruction radial DCE-MRI
Zhifeng Chen, Zhenguo Yuan, Junying
Cheng et al.
To cite this article: Xiang Zhao et al 2022 J. Phys.: Conf. Ser. 2330 012002 - Assessment of data consistency through
cascades of independently recurrent
inference machines for fast and robust
accelerated MRI reconstruction
D Karkalousos, S Noteboom, H E Hulst et
al.
View the article online for updates and enhancements.
- Simultaneous tumor and surrogate motion
tracking with dynamic MRI for radiation
therapy planning
Seyoun Park, Rana Farah, Steven M Shea
et al.

This content was downloaded from IP address 39.45.23.113 on 12/11/2024 at 19:10

SIUSAI-2022 IOP Publishing
Journal of Physics: Conference Series 2330 (2022) 012002 doi:10.1088/1742-6596/2330/1/012002

A review on generative based methods for MRI reconstruction

Xiang Zhao1,*, Tiejun Yang2 and Bingjie Li1

1
School of Information Science and Engineering, Henan University of Technology,
Zhengzhou, China
2
School of Artificial Intelligence and Big Data, Henan University of Technology,
Zhengzhou, China

learnerzx@gmail.com

Abstract. Magnetic resonance imaging (MRI) is one of the most important methods for
clinical diagnosis. However, the main drawback of MRI is the long imaging time, which will
cause the moving artifact by patient movements. With the rapid development of the computing
power of computer, deep learning is widely used in computer vision, natural language
processing, visual recognition and so on. Meanwhile, a large number of reconstruction methods
based on deep learning have also emerged. Recently, many generative models have been
proposed to solve the perception quality problem that existed in fast MRI images. In this paper,
we manage to survey the motivations and reconstruction strategies of generative-based
methods published in journals and conferences over the past five years. First, the background
and theoretical basis of MRI reconstruction are introduced. Secondly, the application of
generative-based methods in MRI reconstruction field is comprehensively summarized and
analyzed, including Generative Adversarial Network (GAN), Variational Autoencoder (VAE)
and VAE-GAN. Then the advantages and disadvantages of the existing generative-based MRI
reconstruction methods are discussed. Finally, several publicly available MR image datasets
and evaluation metrics are presented, which can provide a reference for researchers and
practitioners working in related domains. The conclusions and challenges are also given.

1. Introduction
MRI plays an important role in clinical diagnosis because of being friendly to the human body and its
clear imaging, but it will take a long time to obtain high resolution MR image. The traditional methods
based on compressed sensing (CS) have made great progress in MRI reconstruction, however,
accelerated MRI acquisition has still been a significant challenge for MRI applications due to the long
iteration time and large aliasing artifact in the traditional algorithms.
Recently, deep learning has been widely used in many fields, and CNN-based methods have also
been used for undersampled MRI reconstruction [1][2]. In 2016, Wang et al. [3] initially applied CNN
for MRI reconstruction, and their results showed significant improvement compared with traditional
reconstruction methods. Subsequently, many variant networks based on CNN emerged, including the
network cascaded multiple CNN structures for reconstruction [2], the network combined residual
block for reconstruction. All of these were end-to-end methods that directly established the mapping
between undersampled image and full-sampled real image. With the deep research of these methods,
another unrolled optimization method has been proposed [4][5][6][7][8]. The unrolled optimization
method iteratively expanded the traditional physically-driven algorithm into a multi-layer data-driven

Content from this work may be used under the terms of the Creative Commons Attribution 3.0 licence. Any further distribution
of this work must maintain attribution to the author(s) and the title of the work, journal citation and DOI.
Published under licence by IOP Publishing Ltd 1
SIUSAI-2022 IOP Publishing
Journal of Physics: Conference Series 2330 (2022) 012002 doi:10.1088/1742-6596/2330/1/012002

deep neural network, in which some parameters in the traditional iterative algorithm can be learned by
the neural network. It can reduce the data required in general deep learning methods and speed up the
optimization of traditional iterative algorithms. Traditional optimization algorithms used in MRI
reconstruction included gradient descent (GD) [9], Proximal Gradient Descent (PGD) [2], Iterative
Shrinkage-Thresholding Algorithm (ISTA) [10], Alternating Direction Method of Multipliers
(ADMM) [11], Primal Dual Hybrid Gradient (PDHG) [12], etc.
Although the unrolled methods have interpretability and require less training data, the effect of
MRI reconstruction is not ideal. Due to the excellent performance of GAN in generative tasks, there
exist many methods that use GAN and other generative models to accelerate MRI reconstruction,
which can be seen in Figure 1. In this paper, we focus on the models based on generative modeling for
MRI reconstruction, including GAN, VAE and VAE-GAN, and make a summarization about their
advantages and disadvantages.
Given the rapid development of MRI reconstruction, it is necessary to summarize the large number
of generative methods reported in the literature and analyze the challenges in MRI reconstruction. The
paper presents a comprehensive overview about MRI reconstruction on the basis of generative models.

Number of publications
400
39
53
200 12 39
2 20 267 45
0 16
4 166 166
0 67 92
2018 2019 2020 2021 2022

GAN VAE GAN-VAE

Figure 1. Number of papers of generative-based methods for MRI reconstruction published since
2018 according to the Web of Science and Medical Image Computing and Computer Assisted
Intervention (MICCAI) databases.

2. Generative MRI reconstruction methods

In this section, we review the generative networks in MRI reconstruction, which used different
training strategies, including GAN, VAE and VAE-GAN. The advantages and disadvantages of the
generative-based methods for MRI reconstruction were discussed. Typical methods of generative MRI
reconstruction in recent years were summarized in Table 1.

2.1. GAN based methods

GAN was originally proposed in 2014 [14]. As shown in Figure 2, GAN contains two networks,
generator 𝐺 and discriminator 𝐷. GAN trains the generator to make the data distribute as realistically
as possible to the real distribution~𝑝𝑑𝑎𝑡𝑎 (𝑥) and distinguishes the real data 𝑥 between the false data
𝐺𝑧 based on the discriminator 𝐷.

Figure 2. The vanilla GAN architecture. It consists of a Generator and a Discriminator for adversarial
training.

2
SIUSAI-2022 IOP Publishing
Journal of Physics: Conference Series 2330 (2022) 012002 doi:10.1088/1742-6596/2330/1/012002

Table 1. Summarize the generative-based methods for MRI reconstruction.

Selected
Categories Year Models Description
works
Mardani et Reconstruction high-quality image under sampled k-
2018 GANCS
al. [23] space.
Yang et Frequency loss, image loss, adversarial loss and
2018 DAGAN perception loss were combined as training loss.
al. [15]
Jiang et al. Used the Wasserstein distance to alleviate the gradient
2019 WGAN
[19] vanishing problem.
GAN Shaul et The KIGAN cascaded an image space generator GIM
2020 KIGAN
al. [21] and a k-space generator GK .
Lv, J et al. Combined PI with GAN architecture for accelerated
2021 PIC-GAN
[16] multi-channel MR imaging.
Coupled two discriminators in which one was used for
Huang et PIDD-
2022 enhancing edge information, the other one was for
al. [13] GAN
holistic image reconstruction.
Proposed an effective constraint that helped map the
Chen et al.
2018 AAE lesion-bearing image to its corresponding healthy image
[28]
in the latent space.
An alternative approach was proposed to use
VAE Tezcan et
2019 VAE unsupervised deep learning to learn the probability
al. [29]
distribution of fully sampled MR images.
Zhang et conditional The network was trained using the conditional VAE
2021
al. [39] VAE loss, which incorporated the data formation process.
The network relied on a high-level similarity metric and
Mostapha VAE- invariant representations that were learned by a semi-
2019
et al. [36] GAN supervised discriminator to evaluate the generated
VAE-GAN images.
Liu et al. Proposed a dual-cycle constrained bijective framework
2021 VAE-GAN
[37] that used an unpaired cycle reconstruction constraint.
With the continuous development of deep learning, plenty of variations of GAN-based MRI
reconstruction models have been proposed, such as DAGAN [15], PIC-GAN [16], DCGAN [17],
LAPGAN [18], WGAN [19], RefineGAN [20], KIGAN [21], etc. Shitrit et al. [22] proposed a GAN-
based reconstruction framework, which estimated the missing information in k-space through a
generator, and then distinguished the generated image from the real image through a discriminator.
Yang et al. [15] replaced the CNN module in the generator with a U-Net network in DAGAN,
combined image loss and frequency loss to preserve image details during reconstruction. Mardani et
al. [23] proposed a framework integrating CS with GAN, which used GAN to map undersampled
zero-fill MR images into fully sampled high-quality MR images. Quan et al. [20] introduced a fully
residual U-shaped network in GAN, which consisted of two continuous frameworks, one for
reconstruction (RecoGAN) and the other for refining the results (RefineGAN). Jiang et al. [24] added
Wasserstein distance into GAN to reduce the distance between the reconstructed image and the ground
truth value, which made the training process more stable and alleviated the gradient vanishing. Shaul
et al. [21] proposed a two-stage GAN network, namely KIGAN, to estimate missing k-space data and
restore aliasing artifacts in image space. Li et al. [25] proposed a Structure-Enhanced GAN, namely
SEGAN, which was used to recover the structural information of MRI images in local and global
domains. Murugesan et al. [26] employed a novel GAN-based architecture for reconstructing MR
image. The model consisted of a U-Net generator and a context discriminator that combined global
and local context information in the image. Deora et al. [27] proposed a new GAN-based framework
that utilized the patch-based discriminator and the SSIM loss.

3
SIUSAI-2022 IOP Publishing
Journal of Physics: Conference Series 2330 (2022) 012002 doi:10.1088/1742-6596/2330/1/012002

Generative adversarial network makes the GAN-based MRI reconstruction model have better
performance in reconstructing fine textures in data by using the minimum adversarial and maximum
training strategy. Therefore, many GAN-based accelerated reconstruction methods have emerged in
MRI reconstruction fields in recent years, which focus on preserving high-frequency content and fine
texture details in reconstructed MR images.

2.2. VAE based methods

Actually, VAE is a variant of the Autoencoder (AE) network, which contains an encoder and decoder.
The encoder projects the input image into a low-dimensional vector, and then the features are restored
to the original shape by the decoder.
In fact, VAE derives conditional probabilities of latent vectors from fully sampled data with good
interpretability and image reconstruction quality. In 2015, Makhzani et al. [28] proposed adversarial
Autoencoders (AAE), which introduced adversariality into AE to enhance the generation effect. In
2019, Tezcan et al. [29] used the VAE to learn the non-deterministic information of undersampled k-
space. So high-quality reconstruction image was produced by using the VAE prior information. Zou et
al. [30] proposed the Gen-SToRM model in dynamic imaging, such as Cine MRI, which used the
Gaussian prior probability to deduce the evidence-based lower bound (ELBO) potential variables. But
it resulted in the poor slice alignment and the reconstructed image quality without any constraint
conditions on latent variables. In order to solve this problem, Higgins et al. [31] proposed the
constraint variational framework (cVAE) that extended the original VAE framework with a single
hyperparameter β. It was used to regulate the capacity of the channel, driving the model to learn a
more efficient potential representation of the data. Subsequently, Ahmed et al. [32] replaced
smoothness prior information with potential variables, whose parameters were directly learned from
undersampled data by backpropagation, which further improved the reconstruction performance.
In summary, the major differences between VAE and GAN lies in the use of variational inference
to approximate data distribution. Encoders in VAE are used to approach posterior distributions in the
latent space, and decoders are used to model original information according to possibilities. The
methods mentioned above apply VAE variants for MRI reconstruction to achieve better reconstruction
quality and interpretable latent space.

2.3. Combined GAN and VAE methods

GAN-based methods have the advantage of rich details and clear contour of reconstructed image
texture through the use of adversarial training strategy. However, it has the instability of the training
process and the phenomenon of gradient vanishing [33]. Since the VAE methods are based on
pixelwise, it is unable to pay more attention to the global information, which results in the generated
images blurring in MRI reconstruction. Therefore, the complementary advantages of GAN and VAE
can be used to further produce more regular and clear images.
Recently, a variety of MRI reconstruction networks combining the two approaches have attracted
more attention. Kwon et al. [34] solved the problem of mode collapse and unstable training by
introducing additional automatic encoder structures (VAE-GAN) on the basis of GAN, in which
Wasserstein GAN (WGAN-GP) [35] loss was used to avoid unstable training process. Mostapha et al.
[36] proposed the VAE-GAN-QC framework, which extended the VAE-GAN model to learn feature
information in unbalanced and heterogeneous datasets, and adopted a semi-supervised discriminator to
stabilize the network training process. In 2021, Liu et al. [37] proposed a new VAE-GAN with double
cyclic constraint, in which discriminator with cyclic reconstruction constraint was introduced to
generate accurate and realistic CINE MR images.

3. Discussion
In accelerated MR imaging, texture blurriness and aliasing artifacts often appear in reconstructed
images due to uncontrollable movement of patients during the scanning process. Although researchers

4
SIUSAI-2022 IOP Publishing
Journal of Physics: Conference Series 2330 (2022) 012002 doi:10.1088/1742-6596/2330/1/012002

have proposed methods such as CS, sparse dictionary learning and CNN in MRI reconstruction, these
methods failed to completely solve the above problem.
It is proved that the generative-based method has been applied in generative tasks successfully,
such as style transfer, image transformation, image reconstruction and so on [38]. Therefore, a large
number of generative models have been used in MRI reconstruction domains. It is found that the GAN
generation model is prone to solve the problems of image unrealistic and gradient vanishing, and the
generated images are clear, but the texture details are generated randomly in a large scale. The VAE
method has blurred reconstruction image, however, because it directly simulates potential variables
and the generated image distribution is similar to the real image, and the gradient descent gets stable in
the training process. Therefore, combining the advantages of both GAN and VAE not only
reconstructs clear images, but also better recovers real MR images. Thus, we hope that our method
summarization about accelerated MR imaging is enlightened and promising for clinical diagnosis in
the future.
Since the deep learning methods need amount of data for training to obtain available performance,
for MRI reconstruction, one of the biggest challenges of current generative-based methods is the
scarcity of dataset. For solving this problem, recently, some training strategies emerged such as
transfer learning, few-shot learning, unsupervised learning etc. It can get better results by using these
strategies in MRI reconstruction, so it is a promising direction for further in-depth research.

4. Datasets and evaluation metrics

4.1. Common datasets

4.1.1. BraTS 2021 Dataset1

The BraTS is the dataset of brain glioblastoma segmentation competition. The training set in BraTS
2021 has 2000 cases, and each case has four modes, including native (T1), T1-weighted, T2-weighted
and T2 Fluid Attenuated Inversion Recovery (T2-FLAIR) volumes. which are obtained through
different clinical schemes and scanners of different institutions. All image datasets were manually
segmented by one to four raters according to the same annotation rules.

4.1.2. Calgary-Campinas Dataset2

The Calgary-Campinas dataset consists of healthy brain images and "silver standards" generated using
the STAPLE algorithm and supervised classification. The dataset includes 35 raw data obtained from
clinical 3T MRI scanners. Which uses the 12-channel imaging coil to obtain data and uses the
Orchestra toolbox. The matrix size is 256 × 256.

4.1.3. FastMRI Dataset3

The fastMRI dataset includes knee MRI and brain MRI, which are provided by New York University
School of Medicine. Knee MRI consists of more than 1500 knee MRIs data obtained on 1.5 and 3
Tesla magnets and DICOM images of 10000 clinical knee MRIs. Brain MRI consists of 6970 brain
MRIs obtained on 1.5 and 3 Tesla magnets. The dataset includes three modes: axial T1 weighted, T2
weighted and FLAIR.

1: http://www.braintumorsegmentation.org/
2: https://sites.google.com/view/calgary-campinas-dataset/home
3: https://fastmri.org/dataset/

5
SIUSAI-2022 IOP Publishing
Journal of Physics: Conference Series 2330 (2022) 012002 doi:10.1088/1742-6596/2330/1/012002

4.2. MRI reconstruction evaluation metrics

4.2.1. Peak Signal-to-Noise Ratio (PSNR)

The PSNR is an important evaluation metric to MRI reconstruction results, and it approximately
represents the evaluation of the reconstruction quality. Generally, the higher PSNR shows the better
reconstruction quality. The concrete calculation of PSNR is as follows:
𝐿2
𝑃𝑆𝑁𝑅(𝐼𝑟𝑒𝑐 , 𝐼𝑔𝑡 ) = 10 · log10( 1 ) (1)
∑𝑁
𝐼=1(𝐼𝑔𝑡 (𝑖)−𝐼𝑟𝑒𝑐 (𝑖))
2
𝑁
where 𝐼𝑔𝑡 denotes the real image; 𝐼𝑟𝑒𝑐 stands for the predicted image; 𝐿 represents the data range, and
𝑁 is all pixels of images 𝐼𝑔𝑡 and 𝐼𝑟𝑒𝑐 .

4.2.2. Structural Similarity (SSIM)

The SSIM is a comparative measure of the brightness, contrast and structure between two image
samples. The SSIM of the whole image is the average SSIM of each image patch and is usually used
as an evaluation metric in image reconstruction, and the quality of MRI image reconstruction can be
evaluated by comparing it with full-sampled images. The specific calculation of SSIM is as follows:
(2𝜇𝐼𝑔𝑡 𝜇𝐼𝑟𝑒𝑐 +𝐶𝑙 )(𝜎𝐼𝑔𝑡 𝐼𝑟𝑒𝑐 +𝐶2 )
𝑆𝑆𝐼𝑀(𝐼𝑟𝑒𝑐 , 𝐼𝑔𝑡 ) = (2)
(𝜇𝐼2 +𝜇𝐼2𝑟𝑒𝑐 +𝐶1 )(𝜎𝐼2 +𝜎𝐼2𝑟𝑒𝑐 +𝐶2 )
𝑔𝑡 𝑔𝑡
where 𝐼𝑔𝑡 , 𝐼𝑟𝑒𝑐 , 𝜇 and 𝜎 2 denote the true, predicted images, the mean and variance respectively;
𝜎𝐼𝑔𝑡 𝐼𝑟𝑒𝑐 is the error between 𝐼𝑔𝑡 and 𝐼𝑟𝑒𝑐 ; 𝐶1 and 𝐶2 are constant relaxation terms.

4.2.3. Normalized Mean Square Error (NMSE)

The NMSE is used to evaluate the deviation between the reconstructed image and the fully sampled
image in image reconstruction. In general, the smaller the value is, the better the reconstruction results
are. The formula of NMSE is as follows:
∑𝐾
𝑘=1(𝐼𝑔𝑡 −𝐼𝑟𝑒𝑐 )
2
𝑁𝑀𝑆𝐸(𝐼𝑟𝑒𝑐 , 𝐼𝑔𝑡 ) = ∑𝐾 2 (3)
𝑘=1(𝐼𝑔𝑡 )

5. Conclusion
In this review, we have summarized some challenges in the MRI reconstruction field and some
solutions to address them. The deep learning methods have been proved to show great potential for
MR image reconstruction. In the past five years, there has been explosive growth in this research field.
In this paper, the current reconstruction methods based on generative modeling are divided into three
categories, including GAN, VAE and GAN-VAE. We have comprehensively summarized these
methods and discussed their advantages and disadvantages. Moreover, various public MR image
datasets and several evaluation metrics with details are given. In conclusion, we presume that even
though there are many promising results reported, the utilization of generative-based methods in
clinical imaging is still in its beginning phases and further exploration in this field is expected to
accomplish the degree of progress vital for the dependable clinical use of generative-based imaging
methods. It is emphasized that due to the limited space, we have to limit the number of references.
Therefore, we apologize to the authors for those articles that have not been quoted.

Acknowledgments
This paper was supported by the Innovative Funds Plan of Henan University of Technology (Grant
No. 2021ZKCJ14) and the Open Fund Project of Key Laboratory of Grain Information Processing &
Control (Grant No. KFJJ2021101).

References
[1] Zhu B, Liu J Z, Cauley S F, Rosen B R and Rosen M S 2018 Image reconstruction by domain-
transform manifold learning Nature 555(7697) 487-92.

6
SIUSAI-2022 IOP Publishing
Journal of Physics: Conference Series 2330 (2022) 012002 doi:10.1088/1742-6596/2330/1/012002

[2] Schlemper J, Caballero J, Hajnal J V, Price A N and Rueckert D 2017 A deep cascade of
convolutional neural networks for dynamic MR image reconstruction IEEE transactions on
Medical Imaging 37(2) 491-503.
[3] Wang S, Su Z, Ying L, Peng X, Zhu S and Liang F 2016 Accelerating magnetic resonance
imaging via deep learning 2016 IEEE 13th international symposium on biomedical imaging
(Prague: IEEE) pp 514-7.
[4] Aggarwal H K, Mani M P and Jacob M 2019 MoDL: Model-based deep learning architecture
for inverse problems IEEE transactions on medical imaging 38(2) 394-405.
[5] Cheng J, Wang H, Ying L and Liang D 2019 Model learning: primal dual network-s for fast MR
imaging International Conference on Medical Image Computing and Computer-Assisted
Intervention (Shenzhen: Springer) pp 21-9.
[6] Li Z, Zhang T, Wan P and Zhang D 2019 SEGAN: structure-enhanced generative adversarial
network for compressed sensing MRI reconstruction Proc. of the AAAI Conf. on Artificial
Intelligence vol 33(Menlo Park, CA: AAAI) pp 1012-19.
[7] Murugesan B, Raghavan SV, Sarveswaran K, Ram K and Sivaprakasam M 2019 Recon-
GLGAN: a global-local context based generative adversarial network for MRI reconstruction
Int. Workshop on Machine Learning for Medical Image Reconstruction (Berlin: Springer) pp 3-
15.
[8] Yutong C, David F and Guang Y 2021 Wavelet improved GAN for MRI reconstruction Medical
Imaging 2021: Physics of Medical Imaging vol 11595 (California: SPIE) p 1159513.
[9] Hammernik K, Klatzer T, Kobler E, Recht M P, Sodickson D K, Pock T and Knoll F 2018
Learning a variational network for reconstruction of accelerated MRI data Magnetic resonance
in medicine 79(6) 3055-71.
[10] Zhang J and Ghanem B 2018 ISTA-Net: Interpretable optimization-inspired deep network for
image compressive sensing Proceedings of the IEEE conference on computer vision and pattern
recognition (Utah: IEEE) pp 1828-37.
[11] Yang Y, Sun J, Li H and Xu Z 2017 ADMM-Net: A deep learning approach for compressive
sensing MRI Preprint arXiv/1705.06869.
[12] Zhang X, Lian Q, Yang Y and Su Y 2020 A deep unrolling network inspired by total variation
for compressed sensing MRI Digital Signal Processing 107 102856.
[13] Huang J, et al. 2022 Edge-enhanced dual discriminator generative adversarial network for fast
MRI with parallel imaging using multi-view information Applied Intelligence 1-18.
[14] Goodfellow I, et al. 2014 Generative adversarial nets Advances in neural information processing
systems 27.
[15] Yang G, et al. 2017 DAGAN: deep de-aliasing generative adversarial networks for fast
compressed sensing MRI reconstruction IEEE transactions on medical imaging 37(6) 1310-21.
[16] Lv J, Wang C and Yang G 2021 PIC-GAN: A parallel imaging coupled generative adversarial
network for accelerated multi-channel MRI reconstruction Diagnostics 11(1) 61.
[17] Chuquicusma M, Hussein S, Burt J and Bagci U 2018 How to fool radiologists with generative
adversarial networks? A visual turing test for lung cancer diagnosis IEEE 15th international
symposium on biomedical imaging (Washington: IEEE) pp 240-4.
[18] Baur C, Albarqouni S and Navab N 2018 MelanoGANs: high resolution skin lesion synthesis
with GANs Preprint arXiv/1804.04338.
[19] Jiang M, Yuan Z, Yang X, Zhang J, Gong Y, Xia L and Li T 2019 Accelerating CS-MRI
reconstruction with fine-tuning Wasserstein generative adversarial network IEEE Access 7
152347-57.
[20] Quan T M, Nguyen-Duc T and Jeong W K 2018 Compressed sensing MRI reconstruction using
a generative adversarial network with a cyclic loss IEEE transactions on medical imaging 37(6)
1488-97.
[21] Shaul R, David I, Shitrit O and Raviv T R 2020 Subsampled brain MRI reconstruction by
generative adversarial neural networks Medical Image Analysis 65 101747.

7
SIUSAI-2022 IOP Publishing
Journal of Physics: Conference Series 2330 (2022) 012002 doi:10.1088/1742-6596/2330/1/012002

[22] Shitrit O, Riklin T 2017 Accelerated magnetic resonance imaging by adversarial neural network
Deep learning in medical image analysis and multimodal learning for clinical decision support
(Québec City: Springer) pp 30-8.
[23] Mardani M, Gong E, Cheng J Y, Vasanawala S S, Zaharchuk G, Xing L and Pauly J M 2018
Deep generative adversarial neural networks for compressive sensing MRI IEEE transactions on
medical imaging 38(1) 167-79.
[24] Jiang M, Yuan Z, Yang X, Zhang J, Gong Y, Xia L and Li T 2019 Accelerating CS-MRI
reconstruction with fine-tuning Wasserstein generative adversarial network IEEE Access 7
152347-57.
[25] Li Z, Zhang T, Wan P and Zhang D 2019 SEGAN: structure-enhanced generative adversarial
network for compressed sensing MRI reconstruction Proceedings of the AAAI Conference on
Artificial Intelligence vol 33 pp1012-19.
[26] Murugesan B, Vijaya S, Sarveswaran K, Ram K and Sivaprakasam M 2019 Recon-GLGAN: A
global-local context based generative adversarial network for MRI reconstruction International
Workshop on Machine Learning for Medical Image Reconstruction (Shenzhen: Springer) pp 3-
15.
[27] Deora P, Vasudeva B, Bhattacharya S and Pradhan M 2020 Structure preserving compressive
sensing MRI reconstruction using generative adversarial networks Proceedings of the
IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (Seattle:
IEEE) pp 522-3.
[28] Makhzani A, Shlens J, Jaitly N, Goodfellow I and Frey B 2015 Adversarial autoencoders
Preprint arXiv/1511.05644.
[29] Tezcan K C, Baumgartner C F, Luechinger R, Pruessmann K P and Konukoglu E 2018 MR
image reconstruction using deep density priors IEEE transactions on medical imaging 38(7)
1633-42.
[30] Zou Q, Ahmed A H, Nagpal P, Kruger S and Jacob M 2021 Dynamic imaging using a deep
generative SToRM (Gen-SToRM) model IEEE transactions on medical imaging 40(11) 3102-
12.
[31] Higgins I, Matthey L, Pal A, Burgess P, Glorot X, Botvinick M, Mohamed S and Lerchner A
2017 beta-VAE: Learning Basic Visual Concepts with a Constrained Variational Framework
International Conference on Learning Representations (Toulon: OpenReview) pp 1-22.
[32] Zou Q, Ahmed H, Nagpal P, Kruger S and Jacob M 2021 Deep generative SToRM model for
dynamic imaging 2021 IEEE 18th International Symposium on Biomedical Imaging (Virtual:
IEEE) pp 114-7.
[33] Arjovsky M and Bottou L 2017 Towards principled methods for training generative adversarial
networks Preprint arXiv/1701.04862.
[34] Kwon G, Han C and Kim S 2019 Generation of 3D brain MRI using auto-encoding generative
adversarial networks International Conference on Medical Image Computing and Computer-
Assisted Intervention (Shenzhen: Springer) pp 118-26.
[35] Gulrajani I, Ahmed F, Arjovsky M, Dumoulin V and Courville A C 2017 Improved training of
wasserstein gans Advances in neural information processing systems 30.
[36] Mostapha M, Prieto J, Murphy V, Girault J, Foster M and Rumple A 2019 Semi-supervised
VAE-GAN for out-of-sample detection applied to MRI quality control International Conference
on Medical Image Computing and Computer-Assisted Intervention (Shenzhen: Springer) pp
127-36.
[37] Liu X, Xing F, Prince L, Carass A, Stone M, El G and Woo J 2021 Dual-cycle constrained
bijective VAE-GAN for tagged-to-cine magnetic resonance image synthesis 2021 IEEE 18th
International Symposium on Biomedical Imaging (Virtual: IEEE) pp 1448-52.
[38] Singh N K and Raza K 2021 Medical image generation using generative adversarial networks: a
review Health Informatics: A Computational Perspective in Healthcare 77-96.
[39] Zhang C, Barbano R and Jin B 2021 Conditional variational autoencoder for learned image
reconstruction Computation 9(11) 114.