ArticlePDF Available

Abstract and Figures

Generative adversarial networks have gained a lot of attention in the computer vision community due to their capability of data generation without explicitly modelling the probability density function. The adversarial loss brought by the discriminator provides a clever way of incorporating unlabeled samples into training and imposing higher order consistency. This has proven to be useful in many cases, such as domain adaptation, data augmentation, and image-to-image translation. These properties have attracted researchers in the medical imaging community, and we have seen rapid adoption in many traditional and novel applications, such as image reconstruction, segmentation, detection, classification, and cross-modality synthesis. Based on our observations, this trend will continue and we therefore conducted a review of recent advances in medical imaging using the adversarial training scheme with the hope of benefiting researchers interested in this technique.
Content may be subject to copyright.
Generative Adversarial Network in Medical Imaging: A Review
Xin Yia,, Ekta Waliaa,b, Paul Babyna
aDepartment of Medical Imaging, University of Saskatchewan, 103 Hospital Dr, Saskatoon, SK, S7N 0W8 Canada
bPhilips Canada, 281 Hillmount Road, Markham, Ontario, ON L6C 2S3, Canada
Generative adversarial networks have gained a lot of attention in the computer vision community due to their capability of data
generation without explicitly modelling the probability density function. The adversarial loss brought by the discriminator provides
a clever way of incorporating unlabeled samples into training and imposing higher order consistency. This has proven to be useful
in many cases, such as domain adaptation, data augmentation, and image-to-image translation. These properties have attracted
researchers in the medical imaging community, and we have seen rapid adoption in many traditional and novel applications, such as
image reconstruction, segmentation, detection, classification, and cross-modality synthesis. Based on our observations, this trend
will continue and we therefore conducted a review of recent advances in medical imaging using the adversarial training scheme
with the hope of benefiting researchers interested in this technique.
Keywords: Deep learning, Generative adversarial network, Generative model, Medical imaging, Review
1. Introduction
With the resurgence of deep learning in computer vision
starting from 2012 (Krizhevsky et al.,2012), the adoption of
deep learning methods in medical imaging has increased dra-
matically. It is estimated that there were over 400 papers pub-
lished in 2016 and 2017 in major medical imaging related con-
ference venues and journals (Litjens et al.,2017). The wide
adoption of deep learning in the medical imaging community
is due to its demonstrated potential to complement image inter-
pretation and augment image representation and classification.
In this article, we focus on one of the most interesting recent
breakthroughs in the field of deep learning - generative adver-
sarial networks (GANs) - and their potential applications in the
field of medical imaging.
GANs are a special type of neural network model where
two networks are trained simultaneously, with one focused on
image generation and the other centered on discrimination. The
adversarial training scheme has gained attention in both academia
and industry due to its usefulness in counteracting domain shift,
and eectiveness in generating new image samples. This model
has achieved state-of-the-art performance in many image gener-
ation tasks, including text-to-image synthesis (Xu et al.,2017),
super-resolution (Ledig et al.,2017), and image-to-image trans-
lation (Zhu et al.,2017a).
Unlike deep learning which has its roots traced back to the
1980s (Fukushima and Miyake,1982), the concept of adversar-
ial training is relatively new with significant recent progress (Good-
fellow et al.,2014). This paper presents a general overview of
Corresponding author
Email addresses: (Xin Yi), (Ekta Walia), Paul.Babyn (Paul Babyn)
GANs, describes their promising applications in medical imag-
ing, and identifies some remaining challenges that need to be
solved to enable their successful application in other medical
imaging related tasks.
To present a comprehensive overview of all relevant works
on GANs in medical imaging, we searched databases including
PubMed, arXiv, proceedings of the International Conference on
Medical Image Computing and Computer Assisted Intervention
(MICCAI), SPIE Medical Imaging, IEEE International Sympo-
sium on Biomedical Imaging (ISBI), and International confer-
ence on Medical Imaging with Deep Learning (MIDL). We also
incorporated cross referenced works not identified in the above
search process. Since there are research publications coming
out every month, without losing generality, we set the cut o
time of the search as January 1st, 2019. Works on arXiv that
report only preliminary results are excluded from this review.
Descriptive statistics of these papers based on task, imaging
modality and year can be found in Figure 1.
The remainder of the paper is structured as follows. We
begin with a brief introduction of the principles of GANs and
some of its structural variants in Section 2. It is followed by
a comprehensive review of medical image analysis tasks using
GANs in Section 3including but not limited to the fields of ra-
diology, histopathology and dermatology. We categorize all the
works according to canonical tasks: reconstruction, image syn-
thesis, segmentation, classification, detection, registration, and
others. Section 4summarizes the review and discusses prospec-
tive applications and identifies open challenges.
Preprint submitted to Journal of L
X Templates September 4, 2019
arXiv:submit/2829516 [cs.CV] 4 Sep 2019
50 46
3 3
Proportion of publications (%)
Retinal Fundus Imaging
8 8
Proportion of publications (%)
Number of publications
Figure 1: (a) Categorization of GAN related papers according to canonical tasks. (b) Categorization of GAN related papers according to imaging modality. (c)
Number of GAN related papers published from 2014. Note that some works performed various tasks and conducted evaluation on datasets with dierent modalities.
We counted these works multiple times in plotting these graphs. Works related to cross domain image transfer were counted based on the source domain. The
statistics presented in figure (a) and (b) are based on papers published on or before January 1st, 2019.
real or fake
Figure 2: Schematic view of the vanilla GAN for synthesis of lung nodule on
CT images. Top of the figure shows the network configuration. The part below
shows the input, output and the internal feature representations of the generator
G and discriminator D. G transforms a sample zfrom p(z) into a generated
nodule xg. D is a binary classifier that dierentiates the generated and real
images of lung nodule formed by xgand xrrespectively.
2. Background
2.1. Vanilla GAN
The vanilla GAN (Goodfellow et al.,2014) is a generative
model that was designed for directly drawing samples from the
desired data distribution without the need to explicitly model
the underlying probability density function. It consists of two
neural networks: the generator G and the discriminator D. The
input to G, zis pure random noise sampled from a prior distri-
bution p(z), which is commonly chosen to be a Gaussian or a
uniform distribution for simplicity. The output of G, xgis ex-
pected to have visual similarity with the real sample xrthat is
drawn from the real data distribution pr(x). We denote the non-
linear mapping function learned by G parametrized by θgas
xg=G(z;θg). The input to D is either a real or generated sam-
ple. The output of D, y1is a single value indicating the prob-
ability of the input being a real or fake sample. The mapping
learned by D parametrized by θdis denoted as y1=D(x;θd).
The generated samples form a distribution pg(x) which is de-
sired to be an approximation of pr(x) after successful training.
The top of Figure 2shows an illustration of a vanilla GAN’s
configuration. G in this example is generating a 2D CT slice
depicting a lung nodule.
D’s objective is to dierentiate these two groups of images
whereas the generator G is trained to confuse the discrimina-
tor D as much as possible. Intuitively, G could be viewed as a
forger trying to produce some quality counterfeit material, and
D could be regarded as the police ocer trying to detect the
forged items. In an alternative view, we can perceive G as re-
ceiving a reward signal from D depending upon whether the
generated data is accurate or not. The gradient information is
back propagated from D to G, so G adapts its parameters in or-
der to produce an output image that can fool D. The training
objectives of D and G can be expressed mathematically as:
Exrpr(x)log D(xr)+Exgpg(x)log(1 D(xg)),
Exgpg(x)log(1 D(xg)).
As can be seen, D is simply a binary classifier with a maxi-
mum log likelihood objective. If the discriminator D is trained
to optimality before the next generator G updates, then mini-
mizing LGAN
Gis proven to be equivalent to minimizing the Jensen–
Shannon (JS) divergence between pr(x) and pg(x) (Goodfellow
et al.,2014). The desired outcome after training is that sam-
ples formed by xgshould approximate the real data distribution
2.2. Challenges in optimizing GANs
The above GAN training objective is regarded as a saddle
point optimization problem (Yadav et al.,2018) and the train-
ing is often accomplished by gradient-based methods. G and
D are trained alternately from scratch so that they may evolve
together. However, there is no guarantee of balance between
the training of G and D with the JS divergence. As a conse-
quence, one network may inevitably be more powerful than the
other, which in most cases is D. When D becomes too strong
as opposed to G, the generated samples become too easy to be
separated from real ones, thus reaching a stage where gradients
from D approach zero, providing no guidance for further train-
ing of G. This happens more frequently when generating high
resolution images due to the diculty of generating meaningful
high frequency details.
Another problem commonly faced in training GANs is mode
collapse, which, as the name indicates, is a case when the dis-
tribution pg(x) learned by G focuses on a few limited modes of
the data distribution pr(x). Hence instead of producing diverse
images, it generates a limited set of samples.
2.3. Variants of GANs
2.3.1. Varying objective of D
In order to stabilize training and also to avoid mode col-
lapse, dierent losses for D have been proposed, such as f-
divergence (f-GAN) (Nowozin et al.,2016), least-square (LS-
GAN) (Mao et al.,2016), hinge loss (Miyato et al.,2018), and
Wasserstein distance (WGAN, WGAN-GP) (Arjovsky et al.,
2017;Gulrajani et al.,2017). Among these, Wasserstein dis-
tance is arguably the most popular metric. As an alternative to
the real/fake discrimination scheme, Springenberg (2015) pro-
posed an entropy based objective where real data is encour-
aged to make confident class predictions (CatGAN, Figure 3
b). In EBGAN (Zhao et al.,2016) and BEGAN (Berthelot et al.,
2017) (Figure 3c), the commonly used encoder architecture for
discriminator is replaced with an autoencoder architecture. D’s
objective then becomes matching autoencoder loss distribution
rather than data distribution.
GANs themselves lack the mechanism of inferencing the
underlying latent vector that is likely to encode the input. There-
fore, in ALI (Dumoulin et al.,2016) and BiGAN (Donahue
et al.,2016) (Figure 3d), a separate encoder network is incor-
porated. D’s objective then becomes separating joint samples
(xg,zg) and (xr,zr). In InfoGAN (Figure 3e), the discrimina-
tor outputs the latent vector that encodes part of the seman-
tic features of the generated image. The discriminator maxi-
mizes the mutual information between the generated image and
the latent attribute vector the generated image is conditioned
upon. After successful training, InfoGAN can explore inherent
data attributes and perform conditional data generation based
on these attributes. The use of class labels has been shown to
further improve generated image’s quality and this information
can be easily incorporated into D by enforcing D to provide
class probabilities and use cross entropy loss for optimization
such as used in ACGAN (Odena et al.,2016) (Figure 3f).
2.3.2. Varying objective of G
In the vanilla GAN, G transforms noise zto sample xg=
G(z). This is usually accomplished by using a decoder network
to progressively increase the spatial size of the output until the
desired resolution is achieved as shown in Figure 2.Larsen
et al. (2015) proposed a variational autoencoder network (VAE)
as the underlying architecture of G (VAEGAN, Figure 3g),
where it can use pixel-wise reconstruction loss to enforce the
decoder part of VAE to generate structures to match the real
The original setup of a GAN does not have any restrictions
on the modes of data it can generate. However, if auxiliary
information were provided during the generation, the GAN can
be driven to output images with desired properties. A GAN in
this scenario is usually referred as a conditional GAN (cGAN)
and the generation process can be expressed as xg=G(z,c).
One of the most common conditional inputs cis an image.
pix2pix, the first general purpose GAN based image-to-image
translation framework was proposed by Isola et al. (2016) (Fig-
ure 4a). Further, task related supervision was introduced to the
generator. For example, reconstruction loss for image restora-
tion and Dice loss (Milletari et al.,2016) for segmentation. This
form of supervision requires aligned training pairs. Zhu et al.
(2017a); Kim et al. (2017) relaxed this constraint by stitching
two generators together head to toe so that images can be trans-
lated between two sets of unpaired samples (Figure 4b). For
the sake of simplicity, we chose CycleGAN to represent this
idea in the rest of this paper. Another model named UNIT (Fig-
ure 4c) can also perform unpaired image-to-image transform
by combining two VAEGANs together with each one respon-
sible for one modality but sharing the same latent space (Liu
et al.,2017a). These image-to-image translation frameworks
are very popular in the medical imaging community due to their
general applicability.
Other than image, the conditional input can be class la-
bels (CGAN, Figure 3h) (Mirza and Osindero,2014), text de-
scriptions (Zhang et al.,2017a), object locations (Reed et al.,
2016a,b), surrounding image context (Pathak et al.,2016), or
sketches (Sangkloy et al.,2016). Note that ACGAN mentioned
in the previous section also has a class conditional generator.
2.3.3. Varying architecture
Fully connected layers were used as the building block in
vanilla GAN but later on, were replaced by fully convolutional
downsampling/upsampling layers in DCGAN (Radford et al.,
2015). DCGAN demonstrated better training stability hence
quickly populated the literature. As shown in Figure 2, the gen-
erator in DCGAN architecture works on random input noise
vector by successive upsampling operations eventually gener-
ating an image from it. Two of its important ingredients are
BatchNorm (Ioe and Szegedy,2015) for regulating the ex-
tracted feature scale, and LeakyRelu (Maas et al.,2013) for pre-
venting dead gradients. Very recently, Miyato et al. (2018) pro-
posed a spectral normalization layer that normalized weights
in the discriminator to regulate the scale of feature response
values. With the training stability improved, some works have
(a) Vanilla GAN
(b) CatGAN
(e) InfoGAN
(h) CGAN
(cascade or stack of GANs)
y1real or fake sample
y2certain or uncertain class prediction
y3real or fake reconstruction loss
Figure 3: A schematic view of variants of GAN. crepresents the conditional vector. In CGAN and ACGAN, cis the discrete categorical code (e.g. one hot vector)
that encodes class labels and in InfoGAN it can also be continuous code that encodes attributes. xggenerally refers to the generated image but can also be internal
representations as in SGAN.
(a) pix2pix
(b) CycleGAN
(c) UNIT
domain A domain B real image generated fake image
rAligned training sample
r||p: target consistency
rUnaligned training sample
r)) xa
r)) xb
r||p: cycle consistency
Figure 4: cGAN frameworks for image-to-image translation. pix2pix requires aligned training data whereas this constraint is relaxed in CycleGAN but usually
suers from performance loss. Note that in (a), we chose reconstruction loss as an example of target consistency. This supervision is task related and can take many
other dierent forms. (c) It consists of two VAEGANs with shared latent vector in the VAE part.
also incorporated residual connections into both the genera-
tor and discriminator and experimented with much deeper net-
works (Gulrajani et al.,2017;Miyato et al.,2018). The work
in Miyato and Koyama (2018) proposed a projection based way
to incorporate the conditional information instead of direct con-
catenation and found it to be beneficial in improving the gener-
ated image’s quality.
Directly generating high resolution images from a noise vec-
tor is hard, therefore some works have proposed tackling it in
a progressive manner. In LAPGAN (Figure 3i), Denton et al.
(2015) proposed a stack of GANs, each of which adds higher
frequency details into the generated image. In SGAN, a cas-
cade of GANs is also used but each GAN generates increasingly
lower level representations (Huang et al.,2017), which are
compared with the hierarchical representations extracted from a
discriminatively trained model. Karras et al. (2017) adopted an
alternate way where they progressively grow the generator and
discriminator by adding new layers to them rather than stack-
ing another GAN on top of the preceding one (PGGAN). This
progressive idea was also explored in conditional setting (Wang
et al.,2017b). More recently, Karras et al. (2018) proposed a
style-based generator architecture (styleGAN) where instead of
directly feeding the latent code zto the input of the genera-
tor, they transformed this code first to an intermediate latent
space and then use it to scale and shift the normalized image
feature responses computed from each convolution layer. Sim-
ilarly, Park et al. (2019) proposed SPADE where the segmenta-
tion mask was injected to the generator via a spatially adaptive
normalization layer. This conditional setup was found to better
preserve the semantic layout of the mask than directly feeding
the mask to the generator.
Schematic illustrations of the most representative GANs are
shown in Figure 3. They are GAN, CatGAN, EBGAN/BEGAN,
SGAN. Three popular image-to-image translation cGANs (pix2pix,
CycleGAN, and UNIT) are shown in Figure 4. For a more in-
depth review and empirical evaluation of these dierent variants
of GAN, we refer the reader to (Huang et al.,2018;Creswell
et al.,2018;Kurach et al.,2018).
3. Applications in Medical Imaging
There are generally two ways GANs are used in medical
imaging. The first is focused on the generative aspect, which
can help in exploring and discovering the underlying structure
of training data and learning to generate new images. This prop-
erty makes GANs very promising in coping with data scarcity
and patient privacy. The second focuses on the discriminative
aspect, where the discriminator D can be regarded as a learned
prior for normal images so that it can be used as regularizer or
detector when presented with abnormal images. Figure 5pro-
vides examples of GAN related applications, with examples (a),
(b), (c), (d), (e), (f) that focus on the generative aspect and ex-
ample (g) that exploits the discriminative aspect. In the follow-
ing subsections, in order to help the readers find applications
of their interest, we categorized all the reviewed articles into
canonical tasks: reconstruction, image synthesis, segmentation,
classification, detection, registration, and others.
3.1. Reconstruction
Due to constraints in clinical settings, such as radiation dose
and patient comfort, the diagnostic quality of acquired medical
images may be limited by noise and artifacts. In the last decade,
we have seen a paradigm shift in reconstruction methods chang-
ing from analytic to iterative and now to machine learning based
methods. These data-driven learning based methods either learn
to transfer raw sensory inputs directly to output images or serve
as a post processing step for reducing image noise and remov-
ing artifacts. Most of the methods reviewed in this section are
borrowed directly from the computer vision literature that for-
mulate post-processing as an image-to-image translation prob-
lem where the conditioned inputs of cGANs are compromised
in certain forms, such as low spatial resolution, noise contam-
ination, under-sampling, or aliasing. One exception is for MR
images where the Fourier transform is used to incorporate the
raw K-space data into the reconstruction.
The basic pix2pix framework has been used for low dose
CT denoising (Wolterink et al.,2017b), MR reconstruction (Chen
et al.,2018b;Kim et al.,2018;Dar et al.,2018b;Shitrit and
Raviv,2017), and PET denoising (Wang et al.,2018b). A pre-
trained VGG-net (Simonyan and Zisserman,2014) was further
incorporated into the optimization framework to ensure per-
ceptual similarity (Yang et al.,2017b;Yu et al.,2017;Yang
et al.,2018a;Armanious et al.,2018c;Mahapatra,2017). Yi
and Babyn (2018) introduced a pretrained sharpness detection
network to explicitly constrain the sharpness of the denoised
CT especially for low contrast regions. Mahapatra (2017) com-
puted a local saliency map to highlight blood vessels in super-
resolution process of retinal fundus imaging. A similar idea was
explored by Liao et al. (2018) in sparse view CT reconstruc-
tion. They compute a focus map to modulate the reconstructed
output to ensure that the network focused on important regions.
Besides ensuring image domain data fidelity, frequency domain
data fidelity is also imposed when raw K-space data is available
in MR reconstruction (Quan et al.,2018;Mardani et al.,2017;
Yang et al.,2018a).
Losses of other kinds have been used to highlight local im-
age structures in the reconstruction, such as the saliency loss to
reweight each pixel’s importance based on its perceptual rel-
evance (Mahapatra,2017) and the style-content loss in PET
denoising (Armanious et al.,2018c). In image reconstruction
of moving organs, paired training samples are hard to obtain.
Therefore, Rav`
ı et al. (2018) proposed a physical acquisition
based loss to regulate the generated image structure for endomi-
croscopy super resolution and Kang et al. (2018) proposed to
use CycleGAN together with an identity loss in the denoising
of cardiac CT. Wolterink et al. (2017b) found that in low dose
CT denoising, meaningful results can still be achieved when re-
moving the image domain fidelity loss from the pix2pix frame-
work, but the local image structure can be altered. Papers relat-
ing to medical image reconstruction are summarized in Table 1.
It can be noticed that the underlying methods are almost the
same for all the reconstruction tasks. MR is special case as it
has a well defined forward and backward operation, i.e. Fourier
transform, so that raw K-space data can be incorporated. The
same methodology can potentially be applied to incorporate the
sinogram data in the CT reconstruction process but we have not
seen any research using this idea as yet probably because the
sinogram data is hard to access. The more data used, either
raw K-space or image from other sequence, the better are the
reconstructed results. In general, using adversarial loss pro-
duces more visually appealing results than using pixel-wise re-
construction loss alone. But using adversarial loss to match the
generated and real data distribution may make the model hal-
lucinate unseen structures. Pixel-wise reconstruction loss helps
to combat this problem if paired samples are available, and if
the model was trained on all healthy images but employed to
reconstruct images with pathologies, the hallucination problem
will still exist due to domain mismatch. Cohen et al. (2018)
have conducted extensive experiments to investigate this prob-
lem and suggest that reconstructed images should not be used
for direct diagnosis by radiologists unless the model has been
properly verified.
However, even though the dataset is carefully curated to
match the training and testing distribution, there are other prob-
lems in further boosting performance. We have seen various
dierent losses introduced to the pix2pix framework as shown
in Table 2to improve the reconstructed fidelity of local struc-
tures. There is, however, no reliable way of comparing their
eectiveness except for relying on human observer or down-
stream image analysis tasks. Large scale statistical analysis
by human observer is currently lacking for GAN based recon-
struction methods. Furthermore, public datasets used for im-
age reconstruction are not tailored towards further medical im-
age analysis, which leaves a gap between upstream reconstruc-
tion and downstream analysis tasks. New reference standard
datasets should be created for better comparison of these GAN-
based methods.
(a) low dose CT denoising (b) Cross modality transfer (MRCT) (c) Vessel to fundus image
(d) Skin lesion synthesis (e) Organ segmentation (f) Domain adaptation
(g) Abnormality Detection
Figure 5: Example applications using GANs. Figures are directly cropped from the corresponding papers. (a) Left side shows the noise contaminated low dose CT
and right side shows the denoised CT that well preserved the low contrast regions in the liver (Yi and Babyn,2018). (b) Left side shows the MR image and right side
shows the synthesized corresponding CT. Bone structures were well delineated in the generated CT image (Wolterink et al.,2017a). (c) The generated retinal fundus
image have the exact vessel structures as depicted in the left vessel map (Costa et al.,2017b). (d) Randomly generated skin lesion from random noise (a mixture of
malignant and benign) (Yi et al.,2018). (e) An organ (lung and heart) segmentation example on adult chest X-ray. The shapes of lung and heart are regulated by
the adversarial loss (Dai et al.,2017b). (f) The third column shows the domain adapted brain lesion segmentation result on SWI sequence without training with the
corresponding manual annotation (Kamnitsas et al.,2017). (g) Abnormality detection of optical coherence tomography images of the retina (Schlegl et al.,2017).
3.2. Medical Image Synthesis
Depending on institutional protocols, patient consent may
be required if diagnostic images are intended to be used in a
publication or released into the public domain (Clinical Prac-
tice Committee,2000). GANs are widely for medical image
synthesis. This helps overcome the privacy issues related to di-
agnostic medical image data and tackle the insucient number
of positive cases of each pathology. Lack of experts annotat-
ing medical images poses another challenge for the adoption
of supervised training methods. Although there are ongoing
collaborative eorts across multiple healthcare agencies aim-
ing to build large open access datasets, e.g. Biobank, the Na-
tional Biomedical Imaging Archive (NBIA), The Cancer Imag-
ing Archive (TCIA) and Radiologist Society of North America
(RSNA), this issue remains and constrains the number of im-
ages researchers might have access to.
Traditional ways to augment training sample include scal-
ing, rotation, flipping, translation, and elastic deformation (Simard
et al.,2003). However, these transformations do not account
for variations resulting from dierent imaging protocols or se-
quences, not to mention variations in the size, shape, location
and appearance of specific pathology. GANs provide a more
generic solution and have been used in numerous works for
augmenting training images with promising results.
3.2.1. Unconditional Synthesis
Unconditional synthesis refers to image generation from ran-
dom noise without any other conditional information. Tech-
niques commonly adopted in the medical imaging community
include DCGAN, WGAN, and PGGAN due to their good train-
ing stability. The first two methods can handle an image reso-
lution of up to 256 ×256 but if higher resolution images are
desired, the progressive technique proposed in PGGAN is a
choice. Realistic images can be generated by directly using
the author released code base as long as the variations between
images are not too large, for example, lung nodules and liver
lesions. To make the generated images useful for downstream
tasks, most studies trained a separate generator for each individ-
ual class; for example, Frid-Adar et al. (2018) used three DC-
GANs to generate synthetic samples for three classes of liver
lesions (cysts, metastases, and hemangiomas); generated sam-
ples were found to be beneficial to the lesion classification task
Table 1: Medical image reconstruction publications. In the second column, * following the method denotes some modifications on the basic framework either on
the network architecture or on the employed losses. A brief description of the losses, quantitative measures and datasets can be found in Table 2,3and 7. In the
last column, symbol Ëand édenotes whether the corresponding literature used paired training data or not. All studies were performed in 2D unless otherwise
Publications Method Losses Dataset Quantitative Measure Remarks
Wolterink et al. (2017b) pix2pix* L1, 2 M29 [é] [3D] Denoising
Yi and Babyn (2018) pix2pix* L1, 2, 6 D1 M12, 13, 24, 25 [Ë]Denoising
Yanget al. (2017b) pix2pix* L1, 2, 8 D2 M12, 13, 25 [Ë] [3D] [Abdomen] Denoising
Kang et al. (2018) CycleGAN* L1, 3, 19 M12, 25 [é] [Coronary] Denoising CT
Youet al. (2018a) pix2pix* L1, 2, 9 D2 M1, 11, 12, 13, 25 [Ë] [3D] Denoising
Tang et al. (2018) SGAN L1, 2, 8 M32 [Ë] Denoising, contrast enhance
Shan et al. (2018) pix2pix* L1, 8 D2 M9, 10, 12, 13 [é] [3D] Denoising transfer from 2D
Liu et al. (2019) pix2pix* L1, 2, 8 M13, 24 [Ë] Denoising, using adjacent slice
Liao et al. (2018) pix2pix* L1, 2, 8 M11,12, 13 [Ë] Sparse view CT reconstruction
Wang et al. (2018a) pix2pix L1, 2 M32 [Ë]Metal artefact reduction cochlear implants
Youet al. (2018b) CycleGAN* L1,2, 12 D2 M12, 13, 16 [Ë] Superresolution, denoising
GANs (2018) pix2pix* L1, 2 M11, 12 [Ë] Sparse view CT reconstruction
Armanious et al. (2018b) pix2pix* L1, 2, 8, 11 M11, 12, 13, 15, [Ë] Inpainting
Quan et al. (2018) pix2pix* L1, 2, 15 D11, 12, 13 M11, 12, 13 [Ë] Under-sampled K-space
Mardani et al. (2017) pix2pix* L1, 2, 15 M1 [Ë] Under-sampled K-space
Yuet al. (2017) pix2pix* L1, 2, 8 D11, 3 M11, 12, 13, 24 [Ë]Under-sampled K-space
Yanget al. (2018a) pix2pix* L1, 2, 8, 15 D3, 15 M11, 12, 13, 24 [Ë] Under-sampled K-space
Sanchez and Vilaplana (2018) pix2pix* L1, 2, 4 D16 M12, 13 [Ë] [3D] Superresolution
Chen et al. (2018b) pix2pix* L1, 2 M11, 12, 13 [Ë] [3D] Superresolution
Kim et al. (2018) pix2pix* L1, 2 D19 M1, 11, 13, 26 [Ë] Superresolution
Dar et al. (2018b) pix2pix* L1, 2 D11, 19, 22 M12, 13 [Ë] Under-sampled K-space
Shitrit and Raviv (2017) pix2pix* L1, 2 M12 [Ë] Under-sampled K-space
Ran et al. (2018) pix2pix* L1, 2, 8 D11 M12,13 [Ë] [3D] Denoising
Seitzer et al. (2018) pix2pix* L1, 2, 8 M1, 12, 23 [Ë] Two stage
Abramian and Eklund (2018) CycleGAN L1, 2, 3 D11 M13, 21 [Ë] Facial anonymization problem
Armanious et al. (2018b) pix2pix* L1, 2, 8, 11 M11, 12, 13, 15 [Ë] Inpainting
Oksuz et al. (2018) pix2pix* L1, 2 D26 M11, 12, 13 [Ë] Motion correction
Zhang et al. (2018a) pix2pix* L1, 2, 8, 12 M12, 13 [Ë] Directly in complex-valued k-space data
Armanious et al. (2018a) pix2pix* L1, 2, 8, 11 M13, 14, 15, 18 [Ë]Motion correction
Wang et al. (2018b) cascadecGAN L1, 2 M11, 12, 27 [Ë] [3D]
Armanious et al. (2018c) pix2pix* L1, 2, 8, 11 M1, 11, 12, 13, 14, 15, 18 [Ë]
Retinal fundus imaging
Mahapatra (2017) pix2pix* L1, 2, 8, 17 M11, 12, 13 [Ë] Superresolution
ı et al. (2018) pix2pix* L1, 18, 19 M6, 13 [é] Superresolution
Table 2: A brief summary of dierent losses used in the reviewed publications in Tables 1and 5. The third column specifies conditions to be fulfilled in order to use
the corresponding loss. L in the first column stands for loss.
Abbr. Losses Requirement Remarks
L1 Ladversarial Adversarial loss introduced by the discriminator, can take the form of cross entropy loss, hinge loss, least square loss etc. as discussed in Section 2.3.1
L2 Limage Aligned training pair Element-wise data fidelity loss in image domain to ensure structure similarity to the target when aligned training pair is provided
L3 Lcycle Element-wise loss to ensure self-similarity during cycled transformation when unaligned training pair is provided
L4 Lgradient Aligned training pair Element-wise loss in the gradient domain to emphasize edges
L5 Ledge Aligned training pair Similar to Lgradient but using gradient feature map as a weight to image pixels
L6 Lsharp Aligned training pair Element-wise loss in a feature domain computed from a pre-trained network, which is expected to be the image sharpness with focus on low contrast regions
L7 Lshape,Lseg Annotated pixel-wise label Loss introduced by a segmentor to ensure faithful reconstruction of anatomic regions
L8 Lperceptual Aligned training pair Element-wise loss in a feature domain computed from a pre-trained network which expected to conform to visual perception
L9 Lstructure Aligned training pair Patch-wise loss in the image domain computed with SSIM which claims to better conform to human visual system
L10 Lstructure2 Aligned pair MIND (Heinrich et al.,2012) as used in image registration for two images with the same content from dierent modality
L11 Lstyle-content Aligned training pair Style and content loss to ensure similarity of image style and content. Style is defined as the Gram matrix which is basically the correlation of low-level
L12 Lself-reg Element-wise loss in image domain to ensure structure similarity to the input. Useful in denoising since the two have similar underlying structure
L13 Lsteer Aligned training pair Element-wise loss in a feature domain which is computed from steerable filters with focus on vessel-like structures
L14 Lclassify Aligned image-wise label Lossintroduced by a classifier to get semantic information
L15 Lfrequency Aligned training pair Element-wise loss in frequency domain (K-space) used in MR image reconstruction
L16 LKL Kullback–Leibler divergence which is commonly seen in variational inference to ensure closer approximation to the posterior distribution
L17 Lsaliency Aligned training pair Element-wise loss in a feature domain which is expected to be the saliency map
L18 Lphysical Physical model Loss introduced by a physical image acquisition model
L19 Lregulation Regulate the generated image contrast by keeping the mean value across row and column unchanged
with both improved sensitivity and specificity when combined
with real training data. Bermudez et al. (2018) claimed that
neuroradiologists found generated MR images to be of com-
parable quality to real ones, however, there were discrepancies
in anatomic accuracy. Papers related to unconditional medical
image synthesis are summarized in Table 4.
3.2.2. Cross modality synthesis
Cross modality synthesis (such as generating CT-like im-
ages based on MR images) is deemed to be useful for multiple
reasons, one of which is to reduce the extra acquisition time
and cost. Another reason is to generate new training samples
with the appearance being constrained by the anatomical struc-
tures delineated in the available modality. Most of the methods
reviewed in this section share many similarities to those in Sec-
tion 3.1. pix2pix-based frameworks are used in cases where dif-
ferent image modality data can be co-registered to ensure data
fidelity. CycleGAN-based frameworks are used to handle more
general cases where registration is challenging such as in car-
diac applications. In a study by Wolterink et al. (2017a) for
brain CT image synthesis from MR image, the authors found
that training using unpaired images was even better than us-
ing aligned images. This most likely resulted from the fact
that rigid registration could not very well handle local align-
ment in the throat, mouth, vertebrae, and nasal cavities. Hiasa
et al. (2018) further incorporated gradient consistency loss in
Table 3: A brief summary of quantitative measures used in the reviewed publications listed in Tables 1,4,5and 6.
Abbr. Measures Remarks
Overall image quality without reference
M1 Human observer Gold standard but costly and hard to scale
M2 Breuleux et al. (2011) Kernel density function Estimate the probability density of the generated data and compute the log likelihood of real test data under this distribution
M3 Salimans et al. (2016) Inception score Measure the generated images’ diversity and visual similarity to the real images with the pretrained Inception model
M4 Goodfellow et al. (2014) JS divergence Distance measure between two distributions (used for comparison between normalized color histogram computed from a large
batch of image samples)
M5 Wasserstein distance Distance measure between two distributions (used for comparison between normalized color histogram computed from a large
batch of image samples)
M6 Matkovic et al. (2005) GCF Global contrast factor
M7 K¨
ohler et al. (2013)QvVessel-basedquality metric (noise and blur) for fundus image
M8 Niemeijer et al. (2006) ISC Image structure clustering, a trained classifier based to dierentiate normal from low quality fundus images
M9 Shan et al. (2018) Perceptual loss Dierence of features extracted from a pre-trained VGG net
M10 Shan et al. (2018) Texture loss Gram matrix which is basically the correlation of low-level features, defined as style in style transfer literature
Overall image quality with respect to a groundtruth
M11 NMSE/MAE/MSE (Normalized)mean absolute/square error with respect to a given groundtruth
M12 PSNR/SNR (Peak) signal to noise ratio with respect to a given groundtruth
M13 Wang et al. (2004) SSIM Structural similarity with respect to a given groundtruth
M14 Sheikh and Bovik (2004) VIF Visual information fidelity with regard to a given groundtruth
M15 Wang and Bovik (2002) UQI Universal quality index with regard to a givengroundtruth
M16 Sheikh et al. (2005) IFC Information Fidelity Criterion
M17 Zhang et al. (2011) FSIM A low-level feature based image quality assessment metric with regard to a givengroundtruth
M18 Zhang et al. (2018b) LPIPS Learned perceptual image patch similarity
M19 Pluim et al. (2003) Mutual information Commonly used in cross modality registration in evaluating the alignment of two images
M20 NMI/MI (Normalized) median intensity, used to measure color consistancy of histology images
M21 Lee Rodgers and Nicewander (1988) Cross correlation Globalcorrelation between two images
M22 Low (2010) Clinical measure Dose dierence, gamma analysis for CT
M23 Seitzer et al. (2018) SIS Semantic interpretability score, essentially the dice loss of a pre-trained downstream segmentor
Local image quality
M24 Line profile Measure the loss of spatial resolution
M25 Noise level Standard deviation of intensities in a local smooth region
M26 CBR Contrast to background ratio, measure the local contrast loss
M27 Kinahan and Fletcher (2010) SUV Standard uptake value, a clinical measure in oncology for local interest region, should not vary too much in reconstruction
M28 NPS Noisepower spectrum
Image quality analysis by auxiliary task
M29 Taskspecific statistics Down stream task (e.g. for coronary calcium quantification)
M30 Classification Down stream task
M31 Detection Downstream task (e.g. for lesion/hemorrhage)
M32 Segmentation Down stream task
M33 Cross modality registration Down stream task
M34 Depth estimation Down stream task
the training to improve accuracy at the boundaries. Zhang et al.
(2018d) found that using only cycle loss in the cross modality
synthesis was insucient to mitigate geometric distortions in
the transformation. Therefore, they employed a shape consis-
tency loss that is obtained from two segmentors (segmentation
network). Each segmentor segments the corresponding image
modality into semantic labels and provides implicit shape con-
straints on the anatomy during translation. To make the whole
system end-to-end trainable, semantic labels of training im-
ages from both modalities are required. Zhang et al. (2018c)
and Chen et al. (2018a) proposed using a segmentor also in the
cycle transfer using labels in only one modality. Therefore, the
segmentor is trained oine and fixed during the training of the
image transfer network. As reviewed in Section 2, UNIT and
CycleGAN are two equally valid frameworks for unpaired cross
modality synthesis. It was found that these two frameworks per-
formed almost equally well for the transformation between T1
and T2-weighted MR images (Welander et al.,2018). Papers
related to cross modality medical image synthesis are summa-
rized in Table 5.
3.2.3. Other conditional synthesis
Medical images can be generated by constraints on segmen-
tation maps, text, locations or synthetic images etc. This is use-
ful to synthesize images in uncommon conditions, such as lung
nodules touching the lung border (Jin et al.,2018b). More-
over, the conditioned segmentation maps can also be generated
from GANs (Guibas et al.,2017) or from a pretrained segmen-
tation network (Costa et al.,2017a), by making the generation
a two stage process. Mok and Chung (2018) used cGAN to
augment training images for brain tumour segmentation. The
generator was conditioned on a segmentation map and gener-
ated brain MR images in a coarse to fine manner. To ensure
the tumour was well delineated with a clear boundary in the
generated image, they further forced the generator to output the
tumour boundaries in the generation process. The full list of
synthesis works is summarized in Table 6.
3.3. Segmentation
Generally, researchers have used pixel-wise or voxel-wise
loss such as cross entropy for segmentation. Despite the fact
that U-net (Ronneberger et al.,2015) was used to combine both
low-level and high-level features, there is no guarantee of spa-
tial consistency in the final segmentation map. Traditionally,
conditional random field (CRF) and graph cut methods are usu-
ally adopted for segmentation refinement by incorporating spa-
tial correlation. Their limitation is that they only take into ac-
count pair-wise potentials which might cause serious boundary
leakage in low contrast regions. On the other hand, adversarial
losses as introduced by the discriminator can take into account
high order potentials (Yang et al.,2017a). In this case, the dis-
criminator can be regarded as a shape regulator. This regular-
ization eect is more prominent when the object of interest has
a compact shape, e.g. for lung and heart mask but less use-
ful for deformable objects such as vessels and catheters. This
regulation eect can be also applied to the internal features of
the segmentor to achieve domain (dierent scanners, imaging
protocols, modality) invariance (Kamnitsas et al.,2017;Dou
Table 4: Unconditional medical image synthesis publications. A brief description of the quantitative measures and datasets can be found in Tables 3and 7.
Publications Method Dataset Measures Remarks
Chuquicusma et al. (2017) DCGAN D4 M1 [Lung nodule]
Frid-Adar et al. (2018) DCGAN /ACGAN M30 [Liver lesion] Generating each lesion class separately (with DCGAN) is than generating all classes at once (using ACGAN)
Bowles et al. (2018a) PGGAN M32 [Brain] Joint learning of image and segmentation map
Calimeri et al. (2017) LAPGAN M1, 2, 3 [Brain]
Zhang et al. (2017b) Semi-Coupled-GAN M30 [Heart] Twogenerators coupled with a single discriminator which outputted both a distribution over the image data source and class
Han et al. (2018a) WGAN D20 M1 [Brain]
Beers et al. (2018) PGGAN D21 - [Brain]
Bermudez et al. (2018) DCGAN D23 M1 [Brain]
Mondal et al. (2018) DCGAN* D18, 25 M32 [Brain] Semi-supervised training with labeled, unlabeled, generated data
Bowles et al. (2018a) PGGAN M32 [Brain] Joint learning of image and segmentation map
Salehinejad et al. (2017) DCGAN M30 [Chest] Five dierent GANs to generate five dierent classes of chest disease
Madani et al. (2018b) DCGAN D34 M30 [Chest] Semi-supervised DCGAN can achieve performance comparable with a traditional supervised CNN with an order of mag-
nitude less labeled data
Madani et al. (2018a) DCGAN D34 M30 [Chest] Two GANs to generate normal and abnormal chest X-rays separately
Korkinof et al. (2018) PGGAN
Hu et al. (2017a) WGAN+infoGAN D42 M30, M32 Cell level representation learning
Retinal fundus imaging
Beers et al. (2018) PGGAN – –
Lahiri et al. (2017) DCGAN D43 M30 Semi-supervised DCGAN can achieve performance comparable with a traditional supervised CNN with an order of magnitude less
labeled data
Lahiri et al. (2018) DCGAN D43, 44 M30 Extend the above work by adding an unsupervised loss into the discriminator
Baur et al. (2018b) LAPGAN D28 M4, 11
Baur et al. (2018a) PGGAN D29 M1
Yi et al. (2018) CatGAN+WGAN D27, 30 M30 Semi-supervised skin lesion feature representation learning
et al.,2018). The adversarial loss can also be viewed as a adap-
tively learned similarity measure between the segmented out-
puts and the annotated groundtruth. Therefore, instead of mea-
suring the similarity in the pixel domain, the discriminative net-
work projects the input to a low dimension manifold and mea-
sures the similarity there. The idea is similar to the perceptual
loss. The dierence is that the perceptual loss is computed from
a pre-trained classification network on natural images whereas
the adversarial loss is computed from a network that trained
adaptively during the evolvement of the generator.
Xue et al. (2018) used a multi-scale L1loss in the discrim-
inator where features coming from dierent depths are com-
pared. This was demonstrated to be eective in enforcing the
multi-scale spatial constraints on segmentation maps and the
system achieved state-of-the-art performance in the BRATS 13
and 15 challenges. Zhang et al. (2017c) proposed to use both
annotated and unannotated images in the segmentation pipeline.
The annotated images are used in the same way as in (Xue et al.,
2018;Son et al.,2017) where both element-wise loss and ad-
versarial loss are applied. The unannotated images on the other
hand are only used to compute a segmentation map to confuse
the discriminator. Li and Shen (2018) combined pix2pix with
ACGAN for segmentation of fluorescent microscopy images of
dierent cell types. They found that the introduction of the aux-
iliary classifier branch provides regulation to both the discrimi-
nator and the segmentor.
Unlike these aforementioned segmentation works where ad-
versarial training is used to ensure higher order structure con-
sistency on the final segmentation maps, the adversarial training
scheme in (Zhu et al.,2017b) enforces network invariance to
small perturbations of the training samples in order to reduce
overfitting on small dataset. Papers related to medical image
segmentation are summarized in Table 8.
3.4. Classification
Classification is arguably one of the most successful tasks
where deep learning has been applied. Hierarchical image fea-
tures can be extracted from a deep neural network discrimina-
tively trained with image-wise class labels. GANs have been
used for classification problems as well, either using part of
the generator and discriminator as a feature extractor or di-
rectly using the discriminator as a classifier (by adding an extra
class corresponding to the generated images). Hu et al. (2017a)
used combined WGAN and InfoGAN for unsupervised cell-
level feature representation learning in histopathology images
whereas Yi et al. (2018) combined WGAN and CatGAN for
unsupervised and semi-supervised feature representation learn-
ing for dermoscopy images. Both works extract features from
the discriminator and build a classifier on top. Madani et al.
(2018b), Lahiri et al. (2017) and Lecouat et al. (2018) adopted
the semi-supervised training scheme of GAN for chest abnor-
mality classification, patch-based retinal vessel classification
and cardiac disease diagnosis respectively. They found that
the semi-supervised GAN can achieve performance compara-
Table 5: Cross modality image synthesis publications. In the second column, * following the method denotes some modifications on the basic framework either
on the network architecture or on the employed losses. A brief description of the losses, quantitative evaluation measures and datasets can be found in Tables 2,3
and 7. In the last column, symbol Ëand édenotes whether the corresponding literature used paired training data or not.
Pulications Method Loss Dataset Measures Remarks
Nie et al. (2017,2018) Cascade GAN L1, 2, 4 D16 M11, 12 [Ë]Brain; Pelvis
Emami et al. (2018) cGAN L1, 2 M11, 12, 13 [Ë]Brain
Jin et al. (2018a) CycleGAN L1, 2, 3 M11, 12 [é] Brain
Jiang et al. (2018) CycleGAN* L1, 2, 3, 7, 8 D8 M32 [é] Lung
Chartsias et al. (2017) CycleGAN L1, 3 D9 M32 [é] Heart
Zhang et al. (2018d) CycleGAN* L1, 3, 7 M32 [é][3D] Heart
Huo et al. (2017) CycleGAN* L1, 3, 7 M32 [é] Spleen
Chartsias et al. (2017) CycleGAN L1, 3 M32 [é] Heart
Hiasa et al. (2018) CycleGAN* L1, 3, 4 M19, 32 [é] Musculoskeletal
Wolterink et al. (2017a) CycleGAN L1, 3 M11, 12 [é] Brain
Huo et al. (2018b) CycleGAN L1, 3, 7 M32 [é] Abdomen
Yang et al. (2018b) CycleGAN* L1, 2, 3, 10 M11, 12, 13 [é] Brain
Maspero et al. (2018) pix2pix L1, 2 M11, 22 [Ë] Pelvis
Bi et al. (2017) cGAN L1, 2 M11, 12 [Ë] Chest
Ben-Cohen et al. (2018) FCN+cGAN L1, 2 M11, 12, 31 [Ë] Liver
Armanious et al. (2018c) cGAN* L1, 2, 8, 11 M11, 12, 13, 14, 15, 18 [Ë] Brain
Wei et al. (2018) cascade cGAN L1, 2 M29 [Ë] Brain
Pan et al. (2018) 3D CycleGAN L1, 2, 3 D16 M30 [Ë] Brain
Choi and Lee (2017) pix2pix L1, 2 D16 M13, 29 [Ë] Brain
Synthetic Real
Hou et al. (2017) synthesizer+cGAN L1, 2, 7 D35, 36 M1, 32 [Ë] Histopathology
Real Synthetic
Mahmood et al. (2017) cGAN L1, 12 M34 [é] Endocsocpy
Zhang et al. (2018c) CycleGAN* L1, 3, 7 M32 [é] X-ray
Domain adaption
Chen et al. (2018a) CycleGAN* L1, 3, 7 D32, 33 M32 [é] X-ray
T1 T2 MR
Dar et al. (2018a) CycleGAN L1, 3 D11, 19, 22 M12, 13 [é] Brain
Yang et al. (2018c) cGAN L1, 2 D19 M11, 12, 19, 32, 33 [é] Brain
Welander et al. (2018) CycleGAN, UNIT L1, 2, 3 D24 M11, 12, 19 [é] Brain
Liu (2018) CycleGAN L1, 2, 3 D14 M32 [é] Knee
Yu et al. (2018) cGAN L1, 2 D19 M11, 12, 32 [Ë] [3D] Brain
T1, T2 MRA
Olut et al. (2018) pix2pix* L1, 2, 13 D11 M12, 32 [Ë] Brain
3T 7T MR
Nie et al. (2018) Cascade GAN L1, 2, 4 M11, 12 [Ë] Brain
Histopathology color normalization
Bentaieb and Hamarneh (2018) cGAN+classifier L1, 5, 14 D37, 38, 39 M30 [é]
Zanjani et al. (2018) InfoGAN L1, 2, 12, 16 M20 [é]
Shaban et al. (2018) CycleGAN L1, 2, 3 D37, 40 M12, 13, 17, 30 [é]
Hyperspectral histology H&E
Bayramoglu et al. (2017a) cGAN L1, 2 D41 M12, 13 [Ë] Lung
Table 6: Other conditional image synthesis publications categorized by imaging modality. * following the method denotes some modifications on the basic
framework either on the network architecture or on the employed losses. A brief description of the losses, quantitative evaluation measures and datasets can be
found in Tables 2,3and 7
Publications Conditional information Method Dataset Evaluation
Jin et al. (2018b) (lung nodule) VOI with removed central region [3D] pix2pix* (L1loss considering nodule context) D2 M32
Mok and Chung (2018) Segmentation map Coarse-to-fine boundary-aware D19 M32
Shin et al. (2018) Segmentation map pix2pix D16, 19, 21 M32
Gu et al. (2018) MR CycleGAN D24 M13, 21
Lau et al. (2018) Segmentation map Cascade cGAN - M32
Hu et al. (2018) Gleason score cGAN - -
Hu et al. (2017b) (fetus) Probe location cGAN M1
Tom and Sheet (2018) Segmentation map cascade cGAN D52 M1
Retinal fundus imaging
Zhao et al. (2017) Vessel map cGAN D41, 43, 45 M32
Guibas et al. (2017) Vessel map Dual cGAN D43 M7, 32
Costa et al. (2017a) Vessel map Segmentor+pix2pix D43 M7, 8
Costa et al. (2017b) Vessel map Adversarial VAE+cGAN D43, 46 M8
Appan and Sivaswamy (2018) Vessel map; Lesion map cGAN D46, 47, 48 M7, 31
Iqbal and Ali (2018) Vessel map cGAN D43, 44 M32
Senaras et al. (2018) Segmentation map pix2pix M1
Galbusera et al. (2018) Dierent view; segmentation map pix2pix/CycleGAN – –
Mahapatra et al. (2018b) segmentation map+X-ray pix2pix* (content loss encourage dissimilarity) D33 M30, 32
Oh and Yun (2018) X-ray (for bone supression) pix2pix* (Haar wavelet decomposion) M11, 12, 13, 28
Table 7: Common datasets used in the reviewed literature. In the first column, D stands for Dataset.
Abbre. Dataset Purpose Anatomy Modality
D1 Yi and Babyn (2018)Piglet Denoising Whole body CT
D2 McCollough et al. (2017)LDCT2016 Denoising Abdomen CT
D3 MICCAI2013 Organ segmentation Abdomen, Pelvis CT
D4 Armato III et al. (2015)LIDC-IDRI Lung cancer detection and diagnosis Lung CT
D5 Yan et al. (2018a)DeepLesion Lesion segmentation CT
D6 LiTS2017 Liver tumor segmentation Liver CT
D7 Glocker et al. (2013)Spine Vertebrate localization Spine CT
D8 Aerts et al. (2015)NSCLC-Radiomics Radiomics Lung CT
D9 Zhuang and Shen (2016)MM-WHS Whole heart segmentation Heart CT, MR
D10 Pace et al. (2015)HVSMR 2016 Whole heart and great vessel segmentation Heart, Vessel MR
D11 IXI Analysis of brain development Brain MR
D12 DSB2015 End-systolic/diastolic volumes measurement Heart MR
D13 Mridata MRI reconstruction Knee MR
D14 Ski10 Cartilage and bone segmentation Knee MR
D15 Crimi et al. (2016)BrainLes Lesion segmentation Brain MR
D16 ADNI Alzheimers disease neuroimaging Initiative Brain MR, PET
D17 MAL Brain structure segmentation Brain MR
D18 BRATS2013 Gliomas segmentation Brain MR
D19 BRATS2015 Gliomas segmentation Brain MR
D20 BRATS2016 Gliomas segmentation Brain MR
D21 BRATS2017 Gliomas segmentation, overall survival prediction Brain MR
D22 Bullitt et al. (2005)MIDAS Assessing the eects of healthy aging Brain MR
D23 Resnick et al. (2003)BLSA Baltimore longitudinal study of aging Brain MR
D24 Van Essen et al. (2012)HCP Human connectome project Brain MR
D25 Wang et al. (2019)iSeg2017 Infant brain tissue segmentation Brain MR
D26 UK Biobank Health research Brain, Heart, Body MR
D27 Gutman et al. (2016)ISIC2016 Skin lesion analysis Skin Dermoscopy
D28 Codella et al. (2018)ISIC2017 Skin lesion analysis Skin Dermoscopy
D29 ISIC2018 Skin lesion analysis Skin Dermoscopy
D30 Mendonca et al. (2015)PH2 Skin lesion analysis Skin Dermoscopy
D31 Ballerini et al. (2013)Dermofit Skin lesion analysis Skin Dermoscopy
D32 Jaeger et al. (2014)Montgomery Pulmonary disease detection Chest X-Ray
D33 Shiraishi et al. (2000)JSRT Pulmonary nodule detection Chest X-Ray
D34 NIH PLCO Cancer screening trial for
Prostate, lung, colorectal and ovarian (PLCO) - X-ray; Digital pathology
D35 CBTC2015 Segmentation of nuclei Nuclei Digital pathology
D36 CPM2017 Segmentation of nuclei Nuclei Digital pathology
D37 MITOS-ATYPIA Mitosis detection; Nuclear atypia score evaluation Breast Digital pathology
D38 Sirinukunwattana et al. (2017)GlaS Gland segmentation Colon Digital pathology
D39 K¨
obel et al. (2010)OCHD Carcinoma subtype prediction Ovary Digital pathology
D40 Camelyon16 Lymph node metastases detection Breast Digital pathology
D41 Bayramoglu et al. (2017b)Neslihan Virtual H&E staining Lung Digital pathology
D42 Kainz et al. (2015)CellDetect Cell detection Bone marrow Digital pathology
D43 Staal et al. (2004)DRIVE Blood vessels segmentation Eye Fundus imaging
D44 STARE Structural analysis of the retina Eye Fundus Imaging
D45 Budai et al. (2013)HRF Image quality assessment, segmentation Eye Fundus Imaging
D46 Decenci`
ere et al. (2014)Messidor Segmentation in retinal ophthalmology Eye Fundus Imaging
D47 Prentasic et al. (2013)DRiDB Diabetic retinopathy detection Eye Fundus Imaging
D48 K¨
ainen and Uusitalo (2007)DIARETDB1 Diabetic retinopathy detection Eye Fundus Imaging
D49 Fumero et al. (2011) RIM-ONE Optic nerve head segmentation Eye Fundus Imaging
D50 Hobson et al. (2015)I3A HEp-2 cell classification Skin Fluorescent microscopy
D51 MIVIA HEp-2 cell segmentation Skin Fluorescent microscopy
D52 Balocco et al. (2014)IVUS Vessel inner and outer wall border detection Blood Vessel Ultrasound
D53 Moreira et al. (2012)INbreast Mass segmentation Breast Mammography
D54 Heath et al. (1998)DDSM-BCRP Mass segmentation Breast Mammography
ble with a traditional supervised CNN with an order of magni-
tude less labeled data. Furthermore, Madani et al. (2018b) have
also shown that the adversarial loss can reduce domain overfit-
ting by simply supplying unlabeled test domain images to the
discriminator in identifying cardiac abnormalities in chest X-
ray. A similar work in addressing domain variance in whole
slide images (WSI) has been conducted by Ren et al. (2018).
Most of the other works that used GANs to generate new
training samples have been already mentioned in Section 3.2.1.
These studies applied a two stage process, with the first stage
learned to augment the images and the second stage learned to
perform classification by adopting the traditional classification
network. The two stages are trained disjointedly without any
communication in between. The advantage is that these two
components can be replaced easily if more advanced uncondi-
tional synthesis architectures are proposed whereas the down-
side is that the generation has to be conducted for each class
separately (N models for N classes), which is not memory and
computation ecient. A single model that is capable of per-
forming conditional synthesis of multiple categories is an active
research direction (Brock et al.,2018). Surprisingly, Frid-Adar
et al. (2018) found that using separate GAN (DCGAN) for each
Table 8: Segmentation publications. A brief description of the datasets can be found in Table 7.
Publications Dataset Remarks
Yang et al. (2017a) [3D] [Liver] Generator is essentially a U-net with deep supervisions
Dou et al. (2018) D9 Ensure that the feature distribution of images from both domains (MR and CT) are indistinguishable
Rezaei et al. (2018a) D6 Additional refinement network, patient-wise batchNorm, recurrent cGAN to ensure temporal consistancy
Sekuboyina et al. (2018) D7 Adversarial training based on EBGAN; Butterfly shape network to combine two views
Xue et al. (2018) D18, 19 A multi-scale L1loss in the discriminator where features coming from dierent depth are compared
Rezaei et al. (2017) D21 The generator takes heterogenous MR scans of various contrast as provided by BRATS 17 challenge
Rezaei et al. (2018b) D10 A cascade of cGANs in segmenting myocardium and blood pool
Li et al. (2017) D21 The generator takes heterogenous MR scans of various contrast as provided by BRATS 17 challenge
Moeskops et al. (2017) D17, 18
Kohl et al. (2017) [Prostate] Improved sensitivity
Huo et al. (2018a) [Spleen] Global convolutional network (GCN) with a large receptive field as the generator
Kamnitsas et al. (2017) Regulate the learned representation so that the feature representation is domain invariant
Dou et al. (2018) D9 Ensure that the feature distribution of images from both domains (MR and CT) are indistinguishable
Rezaei et al. (2018a) D21 Additional refinement network, patient-wise batchNorm, recurrent cGAN to ensure temporal consistency
Xu et al. (2018) Joint learning (segmentation and quantification); convLSTM in the generator for spatial-temporal processing; Bi-LSTM in the discriminator to learn relation between tasks
Han et al. (2018b) Local-LSTM in the generator to capture spatial correlations between neighbouring structures
Zhao et al. (2018) D16 Deep supervision; Discriminate segmentation map based on features extracted from a pre-trained network
Retinal fundus imaging
Son et al. (2017) D43, 44 Deep architecture is better for discriminating whole images and has less false positives with fine vessels
Zhang et al. (2017c) D38 Use both annotated and unannotated images in the segmentation pipeline
Shankaranarayana et al. (2017) D49
Dai et al. (2017b) D32, 33 Adversarial loss is able to correct the shape inconsistency
Wang et al. (2017a) Basal membrane segmentation
fluorescent microscopy
Li and Shen (2018) D50, 51 pix2pix +ACGAN; Auxiliary classifier branch provides regulation to both the discriminator and the segmentor
Izadi et al. (2018) D31 Adversarial training helps to refine the boundary precision
Zhu et al. (2017b) D53, 54 Enforce network invariance to small perturbations of the training samples in order to reduce overfitting on small size dataset
Tuysuzoglu et al. (2018) Joint learning (landmark localization +prostate contour segmentation); Contour shape prior imposed by the discriminator
lesion class resulted in better performance in lesion classifica-
tion than using a unified GAN (ACGAN) for all classes. The
underlying reason remains to be explored. Furthermore, Fin-
layson et al. (2018) argue that images generated from GANs
may serve as an eective augmentation in the medium-data
regime, but may not be helpful in a high or low-data regime.
3.5. Detection
The discriminator of GANs can be utilized to detect abnor-
malities such as lesions by learning the probability distribution
of training images depicting normal pathology. Any image that
falls out of this distribution can be deemed as abnormal. Schlegl
et al. (2017) used the exact idea to learn a manifold of normal
anatomical variability and proposed a novel anomaly scoring
scheme based on the fitness of the test image’s latent code to
the learned manifold. The learning process was conducted in
an unsupervised fashion and eectiveness was demonstrated
by state-of-the-art performance of anomaly detection on opti-
cal coherence tomography (OCT) images. Alex et al. (2017)
used GAN for brain lesion detection on MR images. The gen-
erator was used to model the distribution of normal patches
and the trained discriminator was used to compute a posterior
probability of patches centered on every pixel in the test image.
Chen and Konukoglu (2018) used an adversarial auto-encoder
to learn the data distribution of healthy brain MR images. The
lesion image was then mapped to an image without a lesion by
exploring the learned latent space, and the lesion could be high-
lighted by computing the residual of these two images. We can
see that all the detection studies targeted for abnormalities that
are hard to enumerate.
In the image reconstruction section, it has been observed
that if the target distribution is formed from medical images
without pathology, lesions within an image could be removed
in the CycleGAN-based unpaired image transfer due to the dis-
tribution matching eect. However, it can be seen here that if
the target and source domain are of the same imaging modal-
ity diering only in terms of normal and abnormal tissue, this
adverse eect can actually be exploited for abnormality detec-
tion Sun et al. (2018).
3.6. Registration
cGAN can also be used for multi-modal or uni-modal im-
age registration. The generator in this case will either generate
transformation parameters, e.g. 12 numbers for 3D ane trans-
formation, deformation field for non-rigid transformation or di-
rectly generate the transformed image. The discriminator then
discriminates aligned image pairs from unaligned image pairs.
A spatial transformation network (Jaderberg et al.,2015) or a
deformable transformation layer (Fan et al.,2018) is usually
plugged in between these two networks to enable end-to-end
training. Yan et al. (2018b) performed prostate MR to transrec-
tal ultrasound (TRUS) image registration using this framework.
The paired training data was obtained through manual registra-
tion by experts. Yan et al. (2018b) employed a discriminator
to regularize the displacement field computed by the generator
and found this approach to be more eective than the other reg-
ularizers in MR to TRUS registration. Mahapatra et al. (2018a)
used CycleGAN for multi-modal (retinal) and uni-modal (MR)
deformable registration where the generator produces both the
transformed image and the deformation field. Mahapatra et al.
(2018c) took one step further and explored the idea of joint
segmentation and registration with CycleGAN and found their
method performs better than the separate approaches for lung
X-ray images. Tanner et al. (2018) employed CycleGAN for
deformable image registration between MR and CT by first
transforming the source domain image to the target domain and
then employing a mono-modal image similarity measure for the
registration. They found this method can achieve at best sim-
ilar performance with the traditional multi-modal deformable
registration methods.
3.7. Other works
In addition to the tasks described in the aforementioned sec-
tions, GANs have also been applied in other tasks discussed
here. For instance, cGAN has been used for modelling patient
specific motion distribution based on a single preoperative im-
age (Hu et al.,2017c), highlighting regions most accountable
for a disease (Baumgartner et al.,2017) and re-colorization
of endoscopic video data (Ross et al.,2018). In (Mahmood
et al.,2018) pix2pix was used for treatment planning in radio-
therapy by predicting the dose distribution map from CT image.
WGAN has also been used for modelling the progression of
Alzheimer’s disease (AD) in MRI. This is achieved by isolating
the latent encoding of AD and performing arithmetic operation
in the latent space (Bowles et al.,2018b).
4. Discussion
In the years 2017 and 2018, the number of studies applying
GANs has risen significantly. The list of these papers reviewed
for our study can be found on our 1GitHub repository.
About 46% of these papers studied image synthesis, with
cross modality image synthesis being the most important appli-
cation of GANs. MR is ranked as the most common imaging
modality explored in the GAN related literature. We believe
one of the reasons for the significant interest in applying GANs
for MR image analysis is due to the excessive amount of time
spent on the acquisition of multiple sequences. GANs hold the
potential to reduce MR acquisition time by faithfully generating
certain sequences from already acquired ones. A recent study
in image synthesis across dierent MR sequences using Colla-
GAN shows the irreplaceable nature of exogenous contrast se-
quence, but reports the synthesis of endogenous contrast such
as T1, T2, from each other with high fidelity (Lee et al.,2019).
A second reason for the popularity of GANs in MR might be
because of large number of publicly available MR datasets as
shown in Table 7.
Another 37% of these studies fall into the group of recon-
struction and segmentation due to the popularity of image-to-
image translation frameworks. Adversarial training in these
cases imposes a strong shape and texture regulation on the gen-
erator’s output which makes it very promising in these two tasks.
For example, in liver segmentation from 3D CT volumes, the
incorporation of adversarial loss significantly improves the seg-
mentation performance on non-contrast CT (has fuzzy liver bound-
ary) than graph cut and CRF (Yang et al.,2017a).
Further 8% of these studies are related to classification. In
these studies, the most eective use case was to combat domain
shift. For the studies that used GAN for data augmentation in
classification, most focused on generating tiny objects that can
be easily aligned, such as nodules, lesions and cells. We be-
lieve it is partly due to the relatively smaller content variation of
these images compared to the full context image which makes
the training more stable with the current technique. Another
reason might be related to the computation budget of the re-
search since training on high resolution images requires a lot
of GPU time. Although there are studies that applied GAN on
synthesizing whole chest-X-ray (Madani et al.,2018a,b), the
eectiveness has only been shown on fairly easy tasks, e.g.
cardiac abnormality classification and on a medium size data
regime, e.g. a couple of thousand images. With the advent
of large volume labeled datasets, such as the CheXpert (Irvin
et al.,2019), it seems there is diminishing return in the employ-
ment of GANs for image generation, especially for classifica-
tion. We would like to argue that GANs are still useful in the
following two cases. First, nowadays the training of a deep neu-
ral network heavily relies on data augmentation to improve the
network’s generalizability on unseen test data and reduce over-
fitting. However, existing data augmentation operations are all
manually designed operations, e.g. rotation, color jittering, and
can not cover the whole variation of the data. Cubuk et al.
(2018) recently proposed to learn an augmentation policy with
reinforcement learning but the search space still consisted of ba-
sic hand-crafted image processing operations. GANs, however,
can allow us to sample the whole data distribution which oers
much more flexibility in augmenting the training data (Bowles
et al.,2018a). For example, styleGAN, is able to generate high
resolution realistic face images with unprecedented level of de-
tails. This could be readily applied to chest X-ray datasets to
generate images of a pathology class that has sucient number
of cases. Second, it is well known that medical data distribu-
tion is highly skewed with its largest mass centered on common
diseases. It is impossible to accumulate enough training data
for rare diseases, such as rheumatoid arthritis, sickle cell dis-
ease. But radiologists have been trained to detect these diseases
in the long tail. Thus, another potential of GANs will be in
synthesizing uncommon pathology cases, most likely through
conditional generation with the conditioned information being
specified by medical experts either through text description or
hand drawn figures.
The remaining studies pertaining to detection, registration
and other applications are so limited that it is hard to draw any
4.1. Future challenges
Alongside many positive utilities of GANs, there are still
challenges that need to be resolved for their employment to
medical imaging. In image reconstruction and cross modality
image synthesis, most works still adopt traditional shallow ref-
erence metrics such as MAE, PSNR, or SSIM for quantitative
evaluation. These measures, however, do not correspond to the
visual quality of the image. For example, direct optimization of
pixel-wise loss produces a suboptimal (blurry) result but pro-
vides higher numbers than using adversarial loss. It becomes
increasingly dicult to interpret these numbers in horizontal
comparison of GAN-based works especially when extra losses
as shown in Table 2are incorporated. One way to alleviate this
problem is to use down stream tasks such as segmentation or
classification to validate the quality of the generated sample.
Another way is to recruit domain experts but this approach is
expensive, time consuming and hard to scale. Recently, Zhang
et al. (2018b) proposed learned perceptual image path similar-
ity (LPIPS), which outperforms previous metrics in terms of
agreement with human judgements. It has been adopted in
MedGAN (Armanious et al.,2018c) for evaluation of the gener-
ated image quality but it would be interesting to see its eective-
ness for dierent types of medical images as compared to sub-
jective measures from experienced human observers in a more
extensive study. For natural images, the unconditional gener-
ated sample quality and diversity is usually measured by incep-
tion score (Salimans et al.,2016), the mean MS-SSIM metric
among randomly chosen synthetic sample pairs (Odena et al.,
2016), or Fr´
echet Inception distance (FID) (Heusel et al.,2017).
The validity of these metrics for medical images remains to be
Cross domain image-to-image translation can be achieved
with both paired and unpaired training data and it oers many
prospective applications in medical imaging as has already been
seen in section 3.2.2. Unpaired training does not have the data
fidelity loss term therefore there is no guarantee of preserva-
tion of small abnormality regions during the translation process.
Cohen et al. (2018) warn against the use of generated images for
direct interpretation by doctors. They observe that trained Cy-
cleGAN networks (for unpaired data) can be subject to bias due
to matching the generated data to the distribution of the target
domain. This system bias comes into being when target do-
main images in the training set have an over or under represen-
tation of certain classes. As an example of exploitation of this
eect, Mirsky et al. (2019) demonstrate the possibility of ma-
licious tampering of 3D medical imaging using 3D conditional
GANs to remove and inject solitary pulmonary nodule into pa-
tient’s CT scans. This system bias also exists in paired cross
domain image-to-image translation with the data fidelity loss
but only happens when the model was trained on normal im-
ages but tested on abnormal images. Cautions should be taken
in training of the translation model and new methods should be
proposed to faithfully preserve local abnormal regions.
4.2. Interesting future applications
Similar to other deep learning neural network models, var-
ious applications of GANs demonstrated in this paper have di-
rect bearing on improving radiology workflow and patient care.
The strength of GANs however lies in their ability to learn in
an unsupervised and/or weakly-supervised fashion. In partic-
ular, we perceive that image-to-image translation achieved by
cGANs can have various other useful applications in medical
imaging. For example, restoration of MR images acquired with
certain artifacts such as motion, especially in a pediatric setting,
may help reduce the number of repeated exams.
Exploring GANs for image captioning task (Dai et al.,2017a;
Shetty et al.,2017;Melnyk et al.,2018;Fedus et al.,2018)
may lead to semi-automatic generation of medical imaging re-
ports (Jing et al.,2017) potentially reducing image reporting
times. Success of adversarial text classification (Liu et al.,2017b)
also prompts potential utility of GANs in improving perfor-
mance of such systems for automatic MR protocol generation
from free-text clinical indications (Sohn et al.,2017). Auto-
mated systems may improve MRI wait times which have been
on the rise (CIHI,2017) as well as enhance patient care. cGANs,
specifically CycleGAN applications, such as makeup removal (Chang
et al.,2018), can be extended to medical imaging with applica-
tions in improving bone x-ray images by removal of artifacts
such as casts to facilitate enhanced viewing. This may aid
radiologists in assessing fine bony detail, potentially allowing
for enhanced detection of initially occult fractures and help-
ing assess the progress of bone healing more eciently. The
success of GANs in unsupervised anomaly detection (Schlegl
et al.,2017) can help achieve the task of detecting abnormali-
ties in medical images in an unsupervised manner. This has the
potential to be further extended for detection of implanted de-
vices, e.g. staples, wires, tubes, pacemaker and artificial valves
on X-rays. Such an algorithm can also be used for prioritiz-
ing radiologists’ work lists, thus reducing the turnaround time
for reporting critical findings (Gal Yaniv,2018). We also ex-
pect to witness the utility of GANs in medical image synthe-
sis from text descriptions (Bodnar,2018), especially for rare
cases, so as to fill in the gap of training samples required for
training supervised neural networks for medical image classi-
fication tasks. The recent work on styleGAN shows the capa-
bility to control (Karras et al.,2018) the high level attributes
of the synthesized image by manipulating the scale and bias
parameters of the AdaIN layer (Huang and Belongie,2017).
Similarly, the SPADE (Park et al.,2019) controls the semantic
layout of the synthesized image by a spatially adaptive normal-
ization layer. Imagine in the future the desired attribute can be
customized and specified in prior and manipulated in a local-
ized fashion. We may then be able to predict the progression of
disease, measure the impact of drug trial as suggested in Bowles
et al. (2018b) but with more fine-grained controls.
Dierent imaging modalities work by exploiting tissue re-
sponse to a certain physical media, such as x-rays or a magnetic
field, and thus can provide complementary diagnostic informa-
tion to each other. As a common practice in supervised deep
learning, images of one modality type are labelled to train a
network to accomplish a desired task. This process is repeated
when switching modalities even if the underlying anatomical
structure is the same, resulting in a waste of human eort. Ad-
versarial training, or more specifically unpaired cross modal-
ity translation, enables reuse of the labels in all modalities and
opens new ways for unsupervised transfer learning (Dou et al.,
2018;Ying et al.,2019).
Finally, we would like to point out that, although there have
many promising results reported in the literature, the adoption
of GANs in medical imaging is still in its infancy and there is
currently no breakthrough application as yet adopted clinically
for GANs-based methods.
Abramian, D., Eklund, A., 2018. Refacing: reconstructing anonymized facial
features using gans. arXiv preprint arXiv:1810.06455 .
Aerts, H., Rios Velazquez, E., Leijenaar, R.T., Parmar, C., Grossmann, P., Car-
valho, S., Lambin, P., 2015. Data from nsclc-radiomics. The cancer imaging
archive .
Alex, V., KP, M.S., Chennamsetty, S.S., Krishnamurthi, G., 2017. Generative
adversarial networks for brain lesion detection, in: SPIE Medical Imaging,
International Society for Optics and Photonics. pp. 101330G–101330G.
Appan, P., Sivaswamy, J., 2018. Retinal image synthesis for cad development,
in: International Conference Image Analysis and Recognition, Springer. pp.
Arjovsky, M., Chintala, S., Bottou, L., 2017. Wasserstein gan. arXiv preprint
arXiv:1701.07875 .
Armanious, K., K¨
ustner, T., Nikolaou, K., Gatidis, S., Yang, B., 2018a. Ret-
rospective correction of rigid and non-rigid mr motion artifacts using gans.
arXiv preprint arXiv:1809.06276 .
Armanious, K., Mecky, Y., Gatidis, S., Yang, B., 2018b. Adversarial inpainting
of medical image modalities. arXiv preprint arXiv:1810.06621 .
Armanious, K., Yang, C., Fischer, M., K¨
ustner, T., Nikolaou, K., Gatidis, S.,
Yang, B., 2018c. Medgan: Medical image translation using gans. arXiv
preprint arXiv:1806.06397 .
Armato III, S.G., McLennan, G., Bidaut, L., McNitt-Gray, M.F., Meyer, C.R.,
Reeves, A.P., Clarke, L.P., 2015. Data from lidc-idri. the cancer imaging
Ballerini, L., Fisher, R.B., Aldridge, B., Rees, J., 2013. A color and texture
based hierarchical k-nn approach to the classification of non-melanoma skin
lesions, in: Color Medical Image Analysis. Springer, pp. 63–86.
Balocco, S., Gatta, C., Ciompi, F., Wahle, A., Radeva, P., Carlier, S., Unal,
G., Sanidas, E., Mauri, J., Carillo, X., et al., 2014. Standardized evaluation
methodology and reference database for evaluating ivus image segmenta-
tion. Computerized medical imaging and graphics 38, 70–90.
Baumgartner, C.F., Koch, L.M., Tezcan, K.C., Ang, J.X., Konukoglu, E.,
2017. Visual feature attribution using wasserstein gans. arXiv preprint
arXiv:1711.08998 .
Baur, C., Albarqouni, S., Navab, N., 2018a. Generating highly realistic images
of skin lesions with gans, in: OR 2.0 Context-Aware Operating Theaters,
Computer Assisted Robotic Endoscopy, Clinical Image-Based Procedures,
and Skin Image Analysis. Springer, pp. 260–267.
Baur, C., Albarqouni, S., Navab, N., 2018b. Melanogans: High resolution skin
lesion synthesis with gans. arXiv preprint arXiv:1804.04338 .
Bayramoglu, N., Kaakinen, M., Eklund, L., Heikkila, J., 2017a. Towards vir-
tual h&e staining of hyperspectral lung histology images using conditional
generative adversarial networks, in: Proceedings of the IEEE Conference on
Computer Vision and Pattern Recognition, pp. 64–71.
Bayramoglu, N., Kaakinen, M., Eklund, L., Heikkila, J., 2017b. Towards vir-
tual h&e staining of hyperspectral lung histology images using conditional
generative adversarial networks, in: International Conference on Computer
Beers, A., Brown, J., Chang, K., Campbell, J.P., Ostmo, S., Chiang, M.F.,
Kalpathy-Cramer, J., 2018. High-resolution medical image synthesis us-
ing progressively grown generative adversarial networks. arXiv preprint
arXiv:1805.03144 .
Ben-Cohen, A., Klang, E., Raskin, S.P., Soer, S., Ben-Haim, S., Konen, E.,
Amitai, M.M., Greenspan, H., 2018. Cross-modality synthesis from ct to pet
using fcn and gan networks for improved automated lesion detection. arXiv
preprint arXiv:1802.07846 .
Bentaieb, A., Hamarneh, G., 2018. Adversarial stain transfer for histopathology
image analysis. IEEE transactions on medical imaging 37, 792–802.
Bermudez, C., Plassard, A.J., Davis, L.T., Newton, A.T., Resnick, S.M., Land-
man, B.A., 2018. Learning implicit brain mri manifolds with deep learning,
in: Medical Imaging 2018: Image Processing, International Society for Op-
tics and Photonics. p. 105741L.
Berthelot, D., Schumm, T., Metz, L., 2017. Began: boundary equilibrium gen-
erative adversarial networks. arXiv preprint arXiv:1703.10717 .
Bi, L., Kim, J., Kumar, A., Feng, D., Fulham, M., 2017. Synthesis of positron
emission tomography (pet) images via multi-channel generative adversarial
networks (gans), in: Molecular Imaging, Reconstruction and Analysis of
Moving Body Organs, and Stroke Imaging and Treatment. Springer, pp. 43–
Bodnar, C., 2018. Text to image synthesis using generative adversarial net-
works. arXiv preprint arXiv:1805.00676 .
Bowles, C., Chen, L., Guerrero, R., Bentley, P., Gunn, R., Hammers, A., Dickie,
D.A., Hern´
andez, M.V., Wardlaw, J., Rueckert, D., 2018a. Gan augmenta-
tion: Augmenting training data using generative adversarial networks. arXiv
preprint arXiv:1810.10863 .
Bowles, C., Gunn, R., Hammers, A., Rueckert, D., 2018b. Modelling the pro-
gression of alzheimer’s disease in mri using generative adversarial networks,
in: Medical Imaging 2018: Image Processing, International Society for Op-
tics and Photonics. p. 105741K.
Breuleux, O., Bengio, Y., Vincent, P., 2011. Quickly generating representative
samples from an rbm-derived process. Neural computation 23, 2058–2073.
Brock, A., Donahue, J., Simonyan, K., 2018. Large scale gan training for high
fidelity natural image synthesis. arXiv preprint arXiv:1809.11096 .
Budai, A., Bock, R., Maier, A., Hornegger, J., Michelson, G., 2013. Robust
vessel segmentation in fundus images. International journal of biomedical
imaging 2013.
Bullitt, E., Zeng, D., Gerig, G., Aylward, S., Joshi, S., Smith, J.K., Lin, W.,
Ewend, M.G., 2005. Vessel tortuosity and brain tumor malignancy: a
blinded study1. Academic radiology 12, 1232–1240.
Calimeri, F., Marzullo, A., Stamile, C., Terracina, G., 2017. Biomedical data
augmentation using generative adversarial neural networks, in: International
Conference on Artificial Neural Networks, Springer. pp. 626–634.
Chang, H., Lu, J., Yu, F., Finkelstein, A., 2018. Pairedcyclegan: Asymmetric
style transfer for applying and removing makeup, in: 2018 IEEE Conference
on Computer Vision and Pattern Recognition (CVPR).
Chartsias, A., Joyce, T., Dharmakumar, R., Tsaftaris, S.A., 2017. Adversar-
ial image synthesis for unpaired multi-modal cardiac data, in: International
Workshop on Simulation and Synthesis in Medical Imaging, Springer. pp.
Chen, C., Dou, Q., Chen, H., Heng, P.A., 2018a. Semantic-aware generative
adversarial nets for unsupervised domain adaptation in chest x-ray segmen-
tation. arXiv preprint arXiv:1806.00600 .
Chen, X., Konukoglu, E., 2018. Unsupervised detection of lesions in
brain mri using constrained adversarial auto-encoders. arXiv preprint
arXiv:1806.04972 .
Chen, Y., Shi, F., Christodoulou, A.G., Xie, Y., Zhou, Z., Li, D., 2018b. Ef-
ficient and accurate mri super-resolution using a generative adversarial net-
work and 3d multi-level densely connected network, in: International Con-
ference on Medical Image Computing and Computer-Assisted Intervention,
Springer. pp. 91–99.
Choi, H., Lee, D.S., 2017. Generation of structural mr images from amyloid
pet: Application to mr-less quantification. Journal of nuclear medicine: of-
ficial publication, Society of Nuclear Medicine .
Chuquicusma, M.J., Hussein, S., Burt, J., Bagci, U., 2017. How to fool radi-
ologists with generative adversarial networks? a visual turing test for lung
cancer diagnosis. arXiv preprint arXiv:1710.09762 .
CIHI, 2017. Wait times for priority procedures in canada. https://secure. report-2017_en.pdf. Ac-
cessed: 2018-10-17.
Clinical Practice Committee, D.S.o.t., 2000. Informed consent for medical pho-
tographs. Genetics in Medicine 2, 353.
Codella, N.C., Gutman, D., Celebi, M.E., Helba, B., Marchetti, M.A., Dusza,
S.W., Kalloo, A., Liopyris, K., Mishra, N., Kittler, H., et al., 2018. Skin le-
sion analysis toward melanoma detection: A challenge at the 2017 interna-
tional symposium on biomedical imaging (isbi), hosted by the international
skin imaging collaboration (isic), in: Biomedical Imaging (ISBI 2018), 2018
IEEE 15th International Symposium on, IEEE. pp. 168–172.
Cohen, J.P., Luck, M., Honari, S., 2018. Distribution matching losses
can hallucinate features in medical image translation. arXiv preprint
arXiv:1805.08841 .
Costa, P., Galdran, A., Meyer, M.I., Abr`
amo, M.D., Niemeijer, M.,
Mendonc¸a, A.M., Campilho, A., 2017a. Towards adversarial retinal image
synthesis. arXiv preprint arXiv:1701.08974 .
Costa, P., Galdran, A., Meyer, M.I., Niemeijer, M., Abr`
amo, M., Mendonc¸a,
A.M., Campilho, A., 2017b. End-to-end adversarial retinal image synthesis.
IEEE Transactions on Medical Imaging .
Creswell, A., White, T., Dumoulin, V., Arulkumaran, K., Sengupta, B.,
Bharath, A.A., 2018. Generative adversarial networks: An overview. IEEE
Signal Processing Magazine 35, 53–65.
Crimi, A., Menze, B., Maier, O., Reyes, M., Handels, H., 2016. Brainlesion:
Glioma, Multiple Sclerosis, Stroke and Traumatic Brain Injuries: First In-
ternational Workshop, Brainles 2015, Held in Conjunction with MICCAI
2015, Munich, Germany, October 5, 2015, Revised Selected Papers. volume
9556. Springer.
Cubuk, E.D., Zoph, B., Mane, D., Vasudevan, V., Le, Q.V., 2018. Autoaugment:
Learning augmentation policies from data. arXiv preprint arXiv:1805.09501
Dai, B., Lin, D., Urtasun, R., Fidler, S., 2017a. Towards diverse and natural
image descriptions via a conditional gan. arXiv preprint arXiv:1703.06029 .
Dai, W., Doyle, J., Liang, X., Zhang, H., Dong, N., Li, Y., Xing, E.P., 2017b.
Scan: Structure correcting adversarial network for chest x-rays organ seg-
mentation. arXiv preprint arXiv:1703.08770 .
Dar, S.U.H., Yurt, M., Karacan, L., Erdem, A., Erdem, E., C¸ ukur, T., 2018a.
Image synthesis in multi-contrast mri with conditional generative adversarial
networks. arXiv preprint arXiv:1802.01221 .
Dar, S.U.H., Yurt, M., Shahdloo, M., Ildız, M.E., C¸ukur, T., 2018b. Syner-
gistic reconstruction and synthesis via generative adversarial networks for
accelerated multi-contrast mri. arXiv preprint arXiv:1805.10704 .
ere, E., Zhang, X., Cazuguel, G., Lay, B., Cochener, B., Trone, C.,
Gain, P., Ordonez, R., Massin, P., Erginay, A., Charton, B., Klein, J.C.,
2014. Feedback on a publicly distributed database: the messidor database.
Image Analysis & Stereology 33, 231–234. URL: http://www.ias- iss.
org/ojs/IAS/article/view/1155, doi:10.5566/ias.1155.
Denton, E.L., Chintala, S., Fergus, R., et al., 2015. Deep generative image
models using a laplacian pyramid of adversarial networks, in: Advances in
neural information processing systems, pp. 1486–1494.
Donahue, J., Kr¨
uhl, P., Darrell, T., 2016. Adversarial feature learning.
arXiv preprint arXiv:1605.09782 .
Dou, Q., Ouyang, C., Chen, C., Chen, H., Heng, P.A., 2018. Unsupervised
cross-modality domain adaptation of convnets for biomedical image seg-
mentations with adversarial loss. arXiv preprint arXiv:1804.10916 .
Dumoulin, V., Belghazi, I., Poole, B., Lamb, A., Arjovsky, M., Mastropietro,
O., Courville, A., 2016. Adversarially learned inference. arXiv preprint
arXiv:1606.00704 .
Emami, H., Dong, M., Nejad-Davarani, S.P., Glide-Hurst, C., 2018. Generating
synthetic cts from magnetic resonance images using generative adversarial
networks. Medical physics .
Fan, J., Cao, X., Xue, Z., Yap, P.T., Shen, D., 2018. Adversarial similarity
network for evaluating image alignment in deep learning based registration,
in: International Conference on Medical Image Computing and Computer-
Assisted Intervention, Springer. pp. 739–746.
Fedus, W., Goodfellow, I., Dai, A.M., 2018. Maskgan: Better text generation
via filling in the . arXiv preprint arXiv:1801.07736 .
Finlayson, S.G., Lee, H., Kohane, I.S., Oakden-Rayner, L., 2018. Towards
generative adversarial networks as a new paradigm for radiology education.
arXiv preprint arXiv:1812.01547 .
Frid-Adar, M., Diamant, I., Klang, E., Amitai, M., Goldberger, J., Greenspan,
H., 2018. Gan-based synthetic medical image augmentation for in-
creased cnn performance in liver lesion classification. arXiv preprint
arXiv:1803.01229 .
Fukushima, K., Miyake, S., 1982. Neocognitron: A self-organizing neural net-
work model for a mechanism of visual pattern recognition, in: Competition
and cooperation in neural nets. Springer, pp. 267–285.
Fumero, F., Alay´
on, S., Sanchez, J.L., Sigut, J., Gonzalez-Hernandez, M.,
2011. Rim-one: An open retinal image database for optic nerve evaluation,
in: 2011 24th international symposium on computer-based medical systems
(CBMS), IEEE. pp. 1–6.
Gal Yaniv, Anna Kuperberg, E.W., 2018. Deep learning algorithm for optimiz-
ing critical findings report turnaround time, in: SIIM.
Galbusera, F., Niemeyer, F., Seyfried, M., Bassani, T., Casaroli, G., Kienle, A.,
Wilke, H.J., 2018. Exploring the potential of generative adversarial networks
for synthesizing radiological images of the spine to be used in in silico trials.
Frontiers in Bioengineering and Biotechnology 6, 53.
GANs, W., 2018. Sparse-view ct reconstruction using. Machine Learning for
Medical Image Reconstruction 11074, 75.
Glocker, B., Zikic, D., Konukoglu, E., Haynor, D.R., Criminisi, A., 2013.
Vertebrae localization in pathological spine ct via dense classification from
sparse annotations, in: International Conference on Medical Image Comput-
ing and Computer-Assisted Intervention, Springer. pp. 262–270.
Goodfellow, I., Pouget-Abadie, J., Mirza, M., Xu, B., Warde-Farley, D., Ozair,
S., Courville, A., Bengio, Y., 2014. Generative adversarial nets, in: Ad-
vances in neural information processing systems, pp. 2672–2680.
Gu, X., Knutsson, H., Eklund, A., 2018. Generating diusion mri scalar
maps from t1 weighted images using generative adversarial networks. arXiv
preprint arXiv:1810.02683 .
Guibas, J.T., Virdi, T.S., Li, P.S., 2017. Synthetic medical images from dual
generative adversarial networks. arXiv preprint arXiv:1709.01872 .
Gulrajani, I., Ahmed, F., Arjovsky, M., Dumoulin, V., Courville, A., 2017.
Improved training of wasserstein gans. arXiv preprint arXiv:1704.00028 .
Gutman, D., Codella, N.C., Celebi, E., Helba, B., Marchetti, M., Mishra, N.,
Halpern, A., 2016. Skin lesion analysis toward melanoma detection: A
challenge at the international symposium on biomedical imaging (isbi) 2016,
hosted by the international skin imaging collaboration (isic). arXiv preprint
arXiv:1605.01397 .
Han, C., Hayashi, H., Rundo, L., Araki, R., Shimoda, W., Muramatsu, S., Fu-
rukawa, Y., Mauri, G., Nakayama, H., 2018a. Gan-based synthetic brain
mr image generation, in: Biomedical Imaging (ISBI 2018), 2018 IEEE 15th
International Symposium on, IEEE. pp. 734–738.
Han, Z., Wei, B., Mercado, A., Leung, S., Li, S., 2018b. Spine-gan: Semantic
segmentation of multiple spinal structures. Medical image analysis 50, 23–
Heath, M., Bowyer, K., Kopans, D., Kegelmeyer, P., Moore, R., Chang, K., Mu-
nishkumaran, S., 1998. Current status of the digital database for screening
mammography, in: Digital mammography. Springer, pp. 457–460.
Heinrich, M.P., Jenkinson, M., Bhushan, M., Matin, T., Gleeson, F.V., Brady,
M., Schnabel, J.A., 2012. Mind: Modality independent neighbourhood de-
scriptor for multi-modal deformable registration. Medical image analysis
16, 1423–1435.
Heusel, M., Ramsauer, H., Unterthiner, T., Nessler, B., Klambauer, G., Hochre-
iter, S., 2017. Gans trained by a two time-scale update rule converge to a
nash equilibrium. arXiv preprint arXiv:1706.08500 .
Hiasa, Y., Otake, Y., Takao, M., Matsuoka, T., Takashima, K., Prince, J.L.,
Sugano, N., Sato, Y., 2018. Cross-modality image synthesis from unpaired
data using cyclegan: Eects of gradient consistency loss and training data
size. arXiv preprint arXiv:1803.06629 .
Hobson, P., Lovell, B.C., Percannella, G., Vento, M., Wiliem, A., 2015. Bench-
marking human epithelial type 2 interphase cells classification methods on
a very large dataset. Artificial intelligence in medicine 65, 239–250.
Hou, L., Agarwal, A., Samaras, D., Kurc, T.M., Gupta, R.R., Saltz, J.H.,
2017. Unsupervised histopathology image synthesis. arXiv preprint
arXiv:1712.05021 .
Hu, B., Tang, Y., Chang, E.I., Fan, Y., Lai, M., Xu, Y., et al., 2017a. Unsuper-
vised learning for cell-level visual representation in histopathology images
with generative adversarial networks. arXiv preprint arXiv:1711.11317 .
Hu, X., Chung, A.G., Fieguth, P., Khalvati, F., Haider, M.A., Wong, A., 2018.
Prostategan: Mitigating data bias via prostate diusion imaging synthesis
with generative adversarial networks. arXiv preprint arXiv:1811.05817 .
Hu, Y., Gibson, E., Lee, L.L., Xie, W., Barratt, D.C., Vercauteren, T., No-
ble, J.A., 2017b. Freehand ultrasound image simulation with spatially-
conditioned generative adversarial networks, in: Molecular Imaging, Re-
construction and Analysis of Moving Body Organs, and Stroke Imaging and
Treatment. Springer, pp. 105–115.
Hu, Y., Gibson, E., Vercauteren, T., Ahmed, H.U., Emberton, M., Moore, C.M.,
Noble, J.A., Barratt, D.C., 2017c. Intraoperative organ motion models with
an ensemble of conditional generative adversarial networks, in: International
Conference on Medical Image Computing and Computer-Assisted Interven-
tion, Springer. pp. 368–376.
Huang, H., Yu, P.S., Wang, C., 2018. An introduction to image synthesis with
generative adversarial nets. arXiv preprint arXiv:1803.04469 .
Huang, X., Belongie, S., 2017. Arbitrary style transfer in real-time with adap-
tive instance normalization, in: Proceedings of the IEEE International Con-
ference on Computer Vision, pp. 1501–1510.
Huang, X., Li, Y., Poursaeed, O., Hopcroft, J.E., Belongie, S.J., 2017. Stacked
generative adversarial networks., in: CVPR, p. 3.
Huo, Y., Xu, Z., Bao, S., Assad, A., Abramson, R.G., Landman, B.A., 2017.
Adversarial synthesis learning enables segmentation without target modality
ground truth. arXiv preprint arXiv:1712.07695 .
Huo, Y., Xu, Z., Bao, S., Bermudez, C., Plassard, A.J., Liu, J., Yao, Y., As-
sad, A., Abramson, R.G., Landman, B.A., 2018a. Splenomegaly segmen-
tation using global convolutional kernels and conditional generative adver-
sarial networks, in: Medical Imaging 2018: Image Processing, International
Society for Optics and Photonics. p. 1057409.
Huo, Y., Xu, Z., Moon, H., Bao, S., Assad, A., Moyo, T.K., Savona, M.R.,
Abramson, R.G., Landman, B.A., 2018b. Synseg-net: Synthetic segmen-
tation without target modality ground truth. IEEE transactions on medical
imaging .
Ioe, S., Szegedy, C., 2015. Batch normalization: Accelerating deep
network training by reducing internal covariate shift. arXiv preprint
arXiv:1502.03167 .
Iqbal, T., Ali, H., 2018. Generative adversarial network for medical images
(mi-gan). Journal of medical systems 42, 231.
Irvin, J., Rajpurkar, P., Ko, M., Yu, Y., Ciurea-Ilcus, S., Chute, C., Mark-
lund, H., Haghgoo, B., Ball, R., Shpanskaya, K., et al., 2019. Chexpert:
A large chest radiograph dataset with uncertainty labels and expert compar-
ison. arXiv preprint arXiv:1901.07031 .
Isola, P., Zhu, J.Y., Zhou, T., Efros, A.A., 2016. Image-to-image translation
with conditional adversarial networks. arXiv preprint arXiv:1611.07004 .
Izadi, S., Mirikharaji, Z., Kawahara, J., Hamarneh, G., 2018. Generative ad-
versarial networks to segment skin lesions, in: Biomedical Imaging (ISBI
2018), 2018 IEEE 15th International Symposium on, IEEE. pp. 881–884.
Jaderberg, M., Simonyan, K., Zisserman, A., et al., 2015. Spatial transformer
networks, in: Advances in neural information processing systems, pp. 2017–
Jaeger, S., Candemir, S., Antani, S., W´
ang, Y.X.J., Lu, P.X., Thoma, G., 2014.
Two public chest x-ray datasets for computer-aided screening of pulmonary
diseases. Quantitative imaging in medicine and surgery 4, 475.
Jiang, J., Hu, Y.C., Tyagi, N., Zhang, P., Rimner, A., Mageras, G.S., Deasy,
J.O., Veeraraghavan, H., 2018. Tumor-aware, adversarial domain adaptation
from ct to mri for lung cancer segmentation, in: International Conference on
Medical Image Computing and Computer-Assisted Intervention, Springer.
pp. 777–785.
Jin, C.B., Jung, W., Joo, S., Park, E., Saem, A.Y., Han, I.H., Lee, J.I., Cui,
X., 2018a. Deep ct to mr synthesis using paired and unpaired data. arXiv
preprint arXiv:1805.10790 .
Jin, D., Xu, Z., Tang, Y., Harrison, A.P., Mollura, D.J., 2018b. Ct-realistic lung
nodule simulation from 3d conditional generative adversarial networks for
robust lung segmentation. arXiv preprint arXiv:1806.04051 .
Jing, B., Xie, P., Xing, E., 2017. On the automatic generation of medical imag-
ing reports. arXiv preprint arXiv:1711.08195 .
Kainz, P., Urschler, M., Schulter, S., Wohlhart, P., Lepetit, V., 2015. You should
use regression to detect cells, in: International Conference on Medical Image
Computing and Computer-Assisted Intervention, Springer. pp. 276–283.
ainen, R., Uusitalo, H., 2007. Diaretdb1 diabetic retinopathy database
and evaluation protocol, in: Medical Image Understanding and Analysis,
Citeseer. p. 61.
Kamnitsas, K., Baumgartner, C., Ledig, C., Newcombe, V., Simpson, J., Kane,
A., Menon, D., Nori, A., Criminisi, A., Rueckert, D., et al., 2017. Unsu-
pervised domain adaptation in brain lesion segmentation with adversarial
networks, in: International Conference on Information Processing in Medi-
cal Imaging, Springer. pp. 597–609.
Kang, E., Koo, H.J., Yang, D.H., Seo, J.B., Ye, J.C., 2018. Cycle consistent ad-
versarial denoising network for multiphase coronary ct angiography. arXiv
preprint arXiv:1806.09748 .
Karras, T., Aila, T., Laine, S., Lehtinen, J., 2017. Progressive growing
of gans for improved quality, stability, and variation. arXiv preprint
arXiv:1710.10196 .
Karras, T., Laine, S., Aila, T., 2018. A style-based generator architecture for
generative adversarial networks. arXiv preprint arXiv:1812.04948 .
Kim, K.H., Do, W.J., Park, S.H., 2018. Improving resolution of mr images with
an adversarial network incorporating images with dierent contrast. Medical
physics .
Kim, T., Cha, M., Kim, H., Lee, J.K., Kim, J., 2017. Learning to discover
cross-domain relations with generative adversarial networks. arXiv preprint
arXiv:1703.05192 .
Kinahan, P.E., Fletcher, J.W., 2010. Positron emission tomography-computed
tomography standardized uptake values in clinical practice and assessing
response to therapy, in: Seminars in Ultrasound, CT and MRI, Elsevier. pp.
obel, M., Kalloger, S.E., Baker, P.M., Ewanowich, C.A., Arseneau, J.,
Zherebitskiy, V., Abdulkarim, S., Leung, S., Duggan, M.A., Fontaine, D.,
et al., 2010. Diagnosis of ovarian carcinoma cell type is highly reproducible:
a transcanadian study. The American journal of surgical pathology 34, 984–
Kohl, S., Bonekamp, D., Schlemmer, H.P., Yaqubi, K., Hohenfellner, M.,
Hadaschik, B., Radtke, J.P., Maier-Hein, K., 2017. Adversarial networks for
the detection of aggressive prostate cancer. arXiv preprint arXiv:1702.08014
ohler, T., Budai, A., Kraus, M.F., Odstrˇ
cilik, J., Michelson, G., Hornegger, J.,
2013. Automatic no-reference quality assessment for retinal fundus images
using vessel segmentation, in: Computer-Based Medical Systems (CBMS),
2013 IEEE 26th International Symposium on, IEEE. pp. 95–100.
Korkinof, D., Rijken, T., O’Neill, M., Yearsley, J., Harvey, H., Glocker, B.,
2018. High-resolution mammogram synthesis using progressive generative
adversarial networks. arXiv preprint arXiv:1807.03401 .
Krizhevsky, A., Sutskever, I., Hinton, G.E., 2012. Imagenet classification with
deep convolutional neural networks, in: Advances in neural information pro-
cessing systems, pp. 1097–1105.
Kurach, K., Lucic, M., Zhai, X., Michalski, M., Gelly, S., 2018. The gan
landscape: Losses, architectures, regularization, and normalization .
Lahiri, A., Ayush, K., Biswas, P.K., Mitra, P., 2017. Generative adversarial
learning for reducing manual annotation in semantic segmentation on large
scale miscroscopy images: Automated vessel segmentation in retinal fundus
image as test case, in: Conference on Computer Vision and Pattern Recog-
nition Workshops, pp. 42–48.
Lahiri, A., Jain, V., Mondal, A., Biswas, P.K., 2018. Retinal vessel segmen-
tation under extreme low annotation: A generative adversarial network ap-
proach. arXiv preprint arXiv:1809.01348 .
Larsen, A.B.L., Sønderby, S.K., Larochelle, H., Winther, O., 2015. Autoen-
coding beyond pixels using a learned similarity metric. arXiv preprint
arXiv:1512.09300 .
Lau, F., Hendriks, T., Lieman-Sifry, J., Sall, S., Golden, D., 2018. Scargan:
chained generative adversarial networks to simulate pathological tissue on
cardiovascular mr scans, in: Deep Learning in Medical Image Analysis and
Multimodal Learning for Clinical Decision Support. Springer, pp. 343–350.
Lecouat, B., Chang, K., Foo, C.S., Unnikrishnan, B., Brown, J.M., Zenati,
H., Beers, A., Chandrasekhar, V., Kalpathy-Cramer, J., Krishnaswamy, P.,
2018. Semi-supervised deep learning for abnormality classification in retinal
images. arXiv preprint arXiv:1812.07832 .
Ledig, C., Theis, L., Husz´
ar, F., Caballero, J., Cunningham, A., Acosta, A.,
Aitken, A., Tejani, A., Totz, J., Wang, Z., et al., 2017. Photo-realistic sin-
gle image super-resolution using a generative adversarial network. arXiv
preprint .
Lee, D., Moon, W.J., Ye, J.C., 2019. Which contrast does matter? towards a
deep understanding of mr contrast using collaborative gan. arXiv preprint
arXiv:1905.04105 .
Lee Rodgers, J., Nicewander, W.A., 1988. Thirteen ways to look at the corre-
lation coecient. The American Statistician 42, 59–66.
Li, Y., Shen, L., 2018. cc-gan: A robust transfer-learning framework for hep-2
specimen image segmentation. IEEE Access 6, 14048–14058.
Li, Z., Wang, Y., Yu, J., 2017. Brain tumor segmentation using an adversar-
ial network, in: International MICCAI Brainlesion Workshop, Springer. pp.
Liao, H., Huo, Z., Sehnert, W.J., Zhou, S.K., Luo, J., 2018. Adversarial sparse-
view cbct artifact reduction, in: International Conference on Medical Image
Computing and Computer-Assisted Intervention, Springer. pp. 154–162.
Litjens, G., Kooi, T., Bejnordi, B.E., Setio, A.A.A., Ciompi, F., Ghafoorian,
M., van der Laak, J.A., van Ginneken, B., S´
anchez, C.I., 2017. A survey on
deep learning in medical image analysis. arXiv preprint arXiv:1702.05747 .
Liu, F., 2018. Susan: segment unannotated image structure using adversarial
network. Magnetic resonance in medicine .
Liu, M.Y., Breuel, T., Kautz, J., 2017a. Unsupervised image-to-image transla-
tion networks, in: Advances in Neural Information Processing Systems, pp.
Liu, P., Qiu, X., Huang, X., 2017b. Adversarial multi-task learning for text
classification. arXiv preprint arXiv:1704.05742 .
Liu, Z., Bicer, T., Kettimuthu, R., Gursoy, D., De Carlo, F., Foster, I., 2019. To-
mogan: Low-dose x-ray tomography with generative adversarial networks.
arXiv preprint arXiv:1902.07582 .
Low, D.A., 2010. Gamma dose distribution evaluation tool, in: Journal of
Physics-Conference Series, p. 012071.
Maas, A.L., Hannun, A.Y., Ng, A.Y., 2013. Rectifier nonlinearities improve
neural network acoustic models, in: Proc. icml, p. 3.
Madani, A., Moradi, M., Karargyris, A., Syeda-Mahmood, T., 2018a. Chest x-
ray generation and data augmentation for cardiovascular abnormality classi-
fication, in: Medical Imaging 2018: Image Processing, International Society
for Optics and Photonics. p. 105741M.
Madani, A., Moradi, M., Karargyris, A., Syeda-Mahmood, T., 2018b. Semi-
supervised learning with generative adversarial networks for chest x-ray
classification with ability of data domain adaptation, in: Biomedical Imag-
ing (ISBI 2018), 2018 IEEE 15th International Symposium on, IEEE. pp.
Mahapatra, D., 2017. Retinal vasculature segmentation using local saliency
maps and generative adversarial networks for image super resolution. arXiv
preprint arXiv:1710.04783 .
Mahapatra, D., Antony, B., Sedai, S., Garnavi, R., 2018a. Deformable medical
image registration using generative adversarial networks, in: Biomedical
Imaging (ISBI 2018), 2018 IEEE 15th International Symposium on, IEEE.
pp. 1449–1453.
Mahapatra, D., Bozorgtabar, B., Thiran, J.P., Reyes, M., 2018b. Ecient
active learning for image classification and segmentation using a sample
selection and conditional generative adversarial network. arXiv preprint
arXiv:1806.05473 .
Mahapatra, D., Ge, Z., Sedai, S., Chakravorty, R., 2018c. Joint registration
and segmentation of xray images using generative adversarial networks, in:
International Workshop on Machine Learning in Medical Imaging, Springer.
pp. 73–80.
Mahmood, F., Chen, R., Durr, N.J., 2017. Unsupervised reverse domain adap-
tion for synthetic medical images via adversarial training. arXiv preprint
arXiv:1711.06606 .
Mahmood, R., Babier, A., McNiven, A., Diamant, A., Chan, T.C., 2018. Auto-
mated treatment planning in radiation therapy using generative adversarial
networks. arXiv preprint arXiv:1807.06489 .
Mao, X., Li, Q., Xie, H., Lau, R.Y., Wang, Z., 2016. Least squares generative
adversarial networks. arXiv preprint ArXiv:1611.04076 .
Mardani, M., Gong, E., Cheng, J.Y., Vasanawala, S., Zaharchuk, G., Alley, M.,
Thakur, N., Han, S., Dally, W., Pauly, J.M., et al., 2017. Deep generative
adversarial networks for compressed sensing automates mri. arXiv preprint
arXiv:1706.00051 .
Maspero, M., Savenije, M.H., Dinkla, A.M., Seevinck, P.R., Intven, M.P.,
Jurgenliemk-Schulz, I.M., Kerkmeijer, L.G., van den Berg, C.A., 2018.
Dose evaluation of fast synthetic-ct generation using a generative adversar-
ial network for general pelvis mr-only radiotherapy. Physics in Medicine &
Biology 63, 185001.
Matkovic, K., Neumann, L., Neumann, A., Psik, T., Purgathofer, W., 2005.
Global contrast factor-a new approach to image contrast. Computational
Aesthetics 2005, 159–168.
McCollough, C.H., Bartley, A.C., Carter, R.E., Chen, B., Drees, T.A., Edwards,
P., Holmes, D.R., Huang, A.E., Khan, F., Leng, S., et al., 2017. Low-dose ct
for the detection and classification of metastatic liver lesions: Results of the
2016 low dose ct grand challenge. Medical physics 44.
Melnyk, I., Sercu, T., Dognin, P.L., Ross, J., Mroueh, Y., 2018. Improved
image captioning with adversarial semantic alignment. arXiv preprint
arXiv:1805.00063 .
Mendonca, T., Celebi, M., Mendonca, T., Marques, J., 2015. Ph2: A public
database for the analysis of dermoscopic images. Dermoscopy Image Anal-
ysis .
Milletari, F., Navab, N., Ahmadi, S.A., 2016. V-net: Fully convolutional neural
networks for volumetric medical image segmentation, in: 3D Vision (3DV),
2016 Fourth International Conference on, IEEE. pp. 565–571.
Mirsky, Y., Mahler, T., Shelef, I., Elovici, Y., 2019. Ct-gan: Malicious
tampering of 3d medical imagery using deep learning. arXiv preprint
arXiv:1901.03597 .
Mirza, M., Osindero, S., 2014. Conditional generative adversarial nets. arXiv
preprint arXiv:1411.1784 .
Miyato, T., Kataoka, T., Koyama, M., Yoshida, Y., 2018. Spectral normalization
for generative adversarial networks. arXiv preprint arXiv:1802.05957 .
Miyato, T., Koyama, M., 2018. cgans with projection discriminator. arXiv
preprint arXiv:1802.05637 .
Moeskops, P., Veta, M., Lafarge, M.W., Eppenhof, K.A., Pluim, J.P., 2017.
Adversarial training and dilated convolutions for brain mri segmentation,
in: Deep Learning in Medical Image Analysis and Multimodal Learning for
Clinical Decision Support. Springer, pp. 56–64.
Mok, T.C., Chung, A.C., 2018. Learning data augmentation for brain tumor
segmentation with coarse-to-fine generative adversarial networks. arXiv
preprint arXiv:1805.11291 .
Mondal, A.K., Dolz, J., Desrosiers, C., 2018. Few-shot 3d multi-modal medical
image segmentation using generative adversarial learning. arXiv preprint
arXiv:1810.12241 .
Moreira, I.C., Amaral, I., Domingues, I., Cardoso, A., Cardoso, M.J., Cardoso,
J.S., 2012. Inbreast: toward a full-field digital mammographic database.
Academic radiology 19, 236–248.
Nie, D., Trullo, R., Lian, J., Petitjean, C., Ruan, S., Wang, Q., Shen, D.,
2017. Medical image synthesis with context-aware generative adversarial
networks, in: International Conference on Medical Image Computing and
Computer-Assisted Intervention, Springer. pp. 417–425.
Nie, D., Trullo, R., Lian, J., Wang, L., Petitjean, C., Ruan, S., Wang, Q., Shen,
D., 2018. Medical image synthesis with deep convolutional adversarial net-
works. IEEE Transactions on Biomedical Engineering .
Niemeijer, M., Abramo, M.D., van Ginneken, B., 2006. Image structure
clustering for image quality verification of color retina images in diabetic
retinopathy screening. Medical image analysis 10, 888–898.
Nowozin, S., Cseke, B., Tomioka, R., 2016. f-gan: Training generative neural
samplers using variational divergence minimization, in: Advances in Neural
Information Processing Systems, pp. 271–279.
Odena, A., Olah, C., Shlens, J., 2016. Conditional image synthesis with auxil-
iary classifier gans. arXiv preprint arXiv:1610.09585 .
Oh, D.Y., Yun, I.D., 2018. Learning bone suppression from dual energy chest
x-rays using adversarial networks. arXiv preprint arXiv:1811.02628 .
Oksuz, I., Clough, J., Bustin, A., Cruz, G., Prieto, C., Botnar, R., Rueckert, D.,
Schnabel, J.A., King, A.P., 2018. Cardiac mr motion artefact correction from
k-space using deep learning-based reconstruction, in: International Work-
shop on Machine Learning for Medical Image Reconstruction, Springer. pp.
Olut, S., Sahin, Y.H., Demir, U., Unal, G., 2018. Generative adversarial
training for mra image synthesis using multi-contrast mri. arXiv preprint
arXiv:1804.04366 .
Pace, D.F., Dalca, A.V., Geva, T., Powell, A.J., Moghari, M.H., Golland, P.,
2015. Interactive whole-heart segmentation in congenital heart disease,
in: International Conference on Medical Image Computing and Computer-
Assisted Intervention, Springer. pp. 80–88.
Pan, Y., Liu, M., Lian, C., Zhou, T., Xia, Y., Shen, D., 2018. Synthesizing
missing pet from mri with cycle-consistent generative adversarial networks
for alzheimer’s disease diagnosis, in: International Conference on Medical
Image Computing and Computer-Assisted Intervention, Springer. pp. 455–
Park, T., Liu, M.Y., Wang, T.C., Zhu, J.Y., 2019. Semantic image synthesis
with spatially-adaptive normalization. arXiv preprint arXiv:1903.07291 .
Pathak, D., Krahenbuhl, P., Donahue, J., Darrell, T., Efros, A.A., 2016. Con-
text encoders: Feature learning by inpainting, in: Proceedings of the IEEE
Conference on Computer Vision and Pattern Recognition, pp. 2536–2544.
Pluim, J.P., Maintz, J.A., Viergever, M.A., 2003. Mutual-information-based
registration of medical images: a survey. IEEE transactions on medical
imaging 22, 986–1004.
Prentasic, P., Loncaric, S., Vatavuk, Z., Bencic, G., Subasic, M., Petkovic, T.,
Dujmovic, L., Malenica-Ravlic, M., Budimlija, N., Tadic, R., 2013. Diabetic
retinopathy image database (dridb): a new database for diabetic retinopathy
screening programs research, in: Image and Signal Processing and Analysis
(ISPA), 2013 8th International Symposium on, IEEE. pp. 711–716.
Quan, T.M., Nguyen-Duc, T., Jeong, W.K., 2018. Compressed sensing mri re-
construction using a generative adversarial network with a cyclic loss. IEEE
transactions on medical imaging 37, 1488–1497.
Radford, A., Metz, L., Chintala, S., 2015. Unsupervised representation learn-
ing with deep convolutional generative adversarial networks. arXiv preprint
arXiv:1511.06434 .
Ran, M., Hu, J., Chen, Y., Chen, H., Sun, H., Zhou, J., Zhang, Y.,
2018. Denoising of 3-d magnetic resonance images using a residual
encoder-decoder wasserstein generative adversarial network. arXiv preprint
arXiv:1808.03941 .
ı, D., Szczotka, A.B., Shakir, D.I., Pereira, S.P., Vercauteren, T., 2018. Ad-
versarial training with cycle consistency for unsupervised super-resolution
in endomicroscopy .
Reed, S., Akata, Z., Yan, X., Logeswaran, L., Schiele, B., Lee, H., 2016a.
Generative adversarial text to image synthesis, in: Proceedings of The 33rd
International Conference on Machine Learning.
Reed, S.E., Akata, Z., Mohan, S., Tenka, S., Schiele, B., Lee, H., 2016b. Learn-
ing what and where to draw, in: Advances in Neural Information Processing
Systems, pp. 217–225.
Ren, J., Hacihaliloglu, I., Singer, E.A., Foran, D.J., Qi, X., 2018. Adversarial
domain adaptation for classification of prostate histopathology whole-slide
images. arXiv preprint arXiv:1806.01357 .
Resnick, S.M., Pham, D.L., Kraut, M.A., Zonderman, A.B., Davatzikos, C.,
2003. Longitudinal magnetic resonance imaging studies of older adults: a
shrinking brain. Journal of Neuroscience 23, 3295–3301.
Rezaei, M., Harmuth, K., Gierke, W., Kellermeier, T., Fischer, M., Yang, H.,
Meinel, C., 2017. A conditional adversarial network for semantic segmen-
tation of brain tumor, in: International MICCAI Brainlesion Workshop,
Springer. pp. 241–252.
Rezaei, M., Yang, H., Meinel, C., 2018a. Conditional generative refinement
adversarial networks for unbalanced medical image semantic segmentation.
arXiv preprint arXiv:1810.03871 .
Rezaei, M., Yang, H., Meinel, C., 2018b. Whole heart and great vessel segmen-
tation with context-aware of generative adversarial networks, in: Bildverar-
beitung f¨
ur die Medizin 2018. Springer, pp. 353–358.
Ronneberger, O., Fischer, P., Brox, T., 2015. U-net: Convolutional networks
for biomedical image segmentation, in: International Conference on Medi-
cal image computing and computer-assisted intervention, Springer. pp. 234–
Ross, T., Zimmerer, D., Vemuri, A., Isensee, F., Wiesenfarth, M., Bodenstedt,
S., Both, F., Kessler, P., Wagner, M., M¨
uller, B., et al., 2018. Exploiting the
potential of unlabeled endoscopic video data with self-supervised learning.
International journal of computer assisted radiology and surgery , 1–9.
Salehinejad, H., Valaee, S., Dowdell, T., Colak, E., Barfett, J., 2017. General-
ization of deep neural networks for chest pathology classification in x-rays
using generative adversarial networks. arXiv preprint arXiv:1712.01636 .
Salimans, T., Goodfellow, I., Zaremba, W., Cheung, V., Radford, A., Chen,
X., 2016. Improved techniques for training gans, in: Advances in Neural
Information Processing Systems, pp. 2226–2234.
Sanchez, I., Vilaplana, V., 2018. Brain mri super-resolution using 3d generative
adversarial networks .
Sangkloy, P., Lu, J., Fang, C., Yu, F., Hays, J., 2016. Scribbler: Controlling deep
image synthesis with sketch and color. arXiv preprint arXiv:1612.00835 .
Schlegl, T., Seeb¨
ock, P., Waldstein, S.M., Schmidt-Erfurth, U., Langs, G., 2017.
Unsupervised anomaly detection with generative adversarial networks to
guide marker discovery, in: International Conference on Information Pro-
cessing in Medical Imaging, Springer. pp. 146–157.
Seitzer, M., Yang, G., Schlemper, J., Oktay, O., W¨
urfl, T., Christlein, V., Wong,
T., Mohiaddin, R., Firmin, D., Keegan, J., et al., 2018. Adversarial and
perceptual refinement for compressed sensing mri reconstruction, in: Inter-
national Conference on Medical Image Computing and Computer-Assisted
Intervention, Springer. pp. 232–240.
Sekuboyina, A., Rempfler, M., Kukaˇ
cka, J., Tetteh, G., Valentinitsch, A.,
Kirschke, J.S., Menze, B.H., 2018. Btrfly net: Vertebrae labelling with
energy-based adversarial learning of local spine prior. arXiv preprint
arXiv:1804.01307 .
Senaras, C., Niazi, M.K.K., Sahiner, B., Pennell, M.P., Tozbikian, G., Lozanski,
G., Gurcan, M.N., 2018. Optimized generation of high-resolution phantom
images using cgan: Application to quantification of ki67 breast cancer im-
ages. PloS one 13, e0196846.
Shaban, M.T., Baur, C., Navab, N., Albarqouni, S., 2018. Staingan: Stain style
transfer for digital histological images. arXiv preprint arXiv:1804.01601 .
Shan, H., Zhang, Y., Yang, Q., Kruger, U., Kalra, M.K., Sun, L., Cong, W.,
Wang, G., 2018. 3-d convolutional encoder-decoder network for low-dose
ct via transfer learning from a 2-d trained network. IEEE transactions on
medical imaging 37, 1522–1534.
Shankaranarayana, S.M., Ram, K., Mitra, K., Sivaprakasam, M., 2017. Joint
optic disc and cup segmentation using fully convolutional and adversar-
ial networks, in: Fetal, Infant and Ophthalmic Medical Image Analysis.
Springer, pp. 168–176.
Sheikh, H.R., Bovik, A.C., 2004. Image information and visual quality, in:
Acoustics, Speech, and Signal Processing, 2004. Proceedings.(ICASSP’04).
IEEE International Conference on, IEEE. pp. iii–709.
Sheikh, H.R., Bovik, A.C., De Veciana, G., 2005. An information fidelity crite-
rion for image quality assessment using natural scene statistics. IEEE Trans-
actions on image processing 14, 2117–2128.
Shetty, R., Rohrbach, M., Hendricks, L.A., Fritz, M., Schiele, B., 2017. Speak-
ing the same language: Matching machine to human captions by adversarial
training, in: Proceedings of the IEEE International Conference on Computer
Vision (ICCV).
Shin, H.C., Tenenholtz, N.A., Rogers, J.K., Schwarz, C.G., Senjem, M.L.,
Gunter, J.L., Andriole, K.P., Michalski, M., 2018. Medical image synthesis
for data augmentation and anonymization using generative adversarial net-
works, in: International Workshop on Simulation and Synthesis in Medical
Imaging, Springer. pp. 1–11.
Shiraishi, J., Katsuragawa, S., Ikezoe, J., Matsumoto, T., Kobayashi, T., Ko-
matsu, K.i., Matsui, M., Fujita, H., Kodera, Y., Doi, K., 2000. Development
of a digital image database for chest radiographs with and without a lung
nodule: receiver operating characteristic analysis of radiologists’ detection
of pulmonary nodules. American Journal of Roentgenology 174, 71–74.
Shitrit, O., Raviv, T.R., 2017. Accelerated magnetic resonance imaging by
adversarial neural network, in: Deep Learning in Medical Image Analysis
and Multimodal Learning for Clinical Decision Support. Springer, pp. 30–
Simard, P.Y., Steinkraus, D., Platt, J.C., 2003. Best practices for convolutional
neural networks applied to visual document analysis, in: null, IEEE. p. 958.
Simonyan, K., Zisserman, A., 2014. Very deep convolutional networks for
large-scale image recognition. arXiv preprint arXiv:1409.1556 .
Sirinukunwattana, K., Pluim, J.P., Chen, H., Qi, X., Heng, P.A., Guo, Y.B.,
Wang, L.Y., Matuszewski, B.J., Bruni, E., Sanchez, U., et al., 2017. Gland
segmentation in colon histology images: The glas challenge contest. Medi-
cal image analysis 35, 489–502.
Sohn, J.H., Trivedi, H., Mesterhazy, J., Al-adel, F., Vu, T., Rybkin, A., Ohliger,
M., 2017. Development and validation of machine learning based natural
language classifiers to automatically assign mri abdomen/pelvis protocols
from free-text clinical indications, in: SIIM.
Son, J., Park, S.J., Jung, K.H., 2017. Retinal vessel segmentation in fun-
doscopic images with generative adversarial networks. arXiv preprint
arXiv:1706.09318 .
Springenberg, J.T., 2015. Unsupervised and semi-supervised learning with cat-
egorical generative adversarial networks. arXiv preprint arXiv:1511.06390
Staal, J., Abramo, M., Niemeijer, M., Viergever, M., van Ginneken, B., 2004.
Ridge based vessel segmentation in color images of the retina. IEEE Trans-
actions on Medical Imaging 23, 501–509.
Sun, L., Wang, J., Ding, X., Huang, Y., Paisley, J., 2018. An adversarial learn-
ing approach to medical image synthesis for lesion removal. arXiv preprint
arXiv:1810.10850 .
Tang, Y., Cai, J., Lu, L., Harrison, A.P., Yan, K., Xiao, J., Yang, L., Summers,
R.M., 2018. Ct image enhancement using stacked generative adversarial
networks and transfer learning for lesion segmentation improvement, in: In-
ternational Workshop on Machine Learning in Medical Imaging, Springer.
pp. 46–54.
Tanner, C., Ozdemir, F., Profanter, R., Vishnevsky, V., Konukoglu, E., Gok-
sel, O., 2018. Generative adversarial networks for mr-ct deformable image
registration. arXiv preprint arXiv:1807.07349 .
Tom, F., Sheet, D., 2018. Simulating patho-realistic ultrasound images using
deep generative networks with adversarial learning, in: Biomedical Imaging
(ISBI 2018), 2018 IEEE 15th International Symposium on, IEEE. pp. 1174–
Tuysuzoglu, A., Tan, J., Eissa, K., Kiraly, A.P., Diallo, M., Kamen, A., 2018.
Deep adversarial context-aware landmark detection for ultrasound imaging.
arXiv preprint arXiv:1805.10737 .
Van Essen, D.C., Ugurbil, K., Auerbach, E., Barch, D., Behrens, T., Bucholz,
R., Chang, A., Chen, L., Corbetta, M., Curtiss, S.W., et al., 2012. The
human connectome project: a data acquisition perspective. Neuroimage 62,
Wang, D., Gu, C., Wu, K., Guan, X., 2017a. Adversarial neural networks
for basal membrane segmentation of microinvasive cervix carcinoma in
histopathology images, in: Machine Learning and Cybernetics (ICMLC),
2017 International Conference on, IEEE. pp. 385–389.
Wang, J., Zhao, Y., Noble, J.H., Dawant, B.M., 2018a. Conditional genera-
tive adversarial networks for metal artifact reduction in ct images of the ear,
in: International Conference on Medical Image Computing and Computer-
Assisted Intervention, Springer. pp. 3–11.
Wang, L., Nie, D., Li, G., Puybareau, ´
E., Dolz, J., Zhang, Q., Wang, F., Xia, J.,
Wu, Z., Chen, J., et al., 2019. Benchmark on automatic 6-month-old infant
brain segmentation algorithms: the iseg-2017 challenge. IEEE transactions
on medical imaging .
Wang, T.C., Liu, M.Y., Zhu, J.Y., Tao, A., Kautz, J., Catanzaro, B., 2017b.
High-resolution image synthesis and semantic manipulation with condi-
tional gans. arXiv preprint arXiv:1711.11585 .
Wang, Y., Yu, B., Wang, L., Zu, C., Lalush, D.S., Lin, W., Wu, X., Zhou, J.,
Shen, D., Zhou, L., 2018b. 3d conditional generative adversarial networks
for high-quality pet image estimation at low dose. NeuroImage 174, 550–
Wang, Z., Bovik, A.C., 2002. A universal image quality index. IEEE signal
processing letters 9, 81–84.
Wang, Z., Bovik, A.C., Sheikh, H.R., Simoncelli, E.P., 2004. Image quality
assessment: from error visibility to structural similarity. IEEE transactions
on image processing 13, 600–612.
Wei, W., Poirion, E., Bodini, B., Durrleman, S., Ayache, N., Stanko, B., Col-
liot, O., 2018. Learning myelin content in multiple sclerosis from multi-
modal mri through adversarial training. arXiv preprint arXiv:1804.08039
Welander, P., Karlsson, S., Eklund, A., 2018. Generative adversarial networks
for image-to-image translation on multi-contrast mr images-a comparison of
cyclegan and unit. arXiv preprint arXiv:1806.07777 .