GAN Augmentation: Augmenting Training Data

using Generative Adversarial Networks

Christopher Bowles1 , Liang Chen1 , Ricardo Guerrero1,6∗ , Paul Bentley2 , Roger

Gunn2,3 , Alexander Hammers4 , David Alexander Dickie5,7 , Maria Valdés
Hernández5 , Joanna Wardlaw5 , and Daniel Rueckert1
arXiv:1810.10863v1 [cs.CV] 25 Oct 2018

1
Department of Computing, Imperial College London, UK
2
Department of Medicine, Imperial College London, UK
3
Imanova Ltd., London, UK
4
PET Centre, Kings College London, UK
5
Department of Neuroimaging Sciences, University of Edinburgh, UK
6
Samsung AI Research Centre (SAIC), Cambridge, UK
7
Institute of Cardiovascular and Medical Sciences, University of Glasgow, UK

Abstract

One of the biggest issues facing the use of machine learning in medical imaging
is the lack of availability of large, labelled datasets. The annotation of medical
images is not only expensive and time consuming but also highly dependent on
the availability of expert observers. The limited amount of training data can
inhibit the performance of supervised machine learning algorithms which often
need very large quantities of data on which to train to avoid overfitting. So far,
much effort has been directed at extracting as much information as possible from
what data is available. Generative Adversarial Networks (GANs) offer a novel
way to unlock additional information from a dataset by generating synthetic
samples with the appearance of real images. This paper demonstrates the feasi-
bility of introducing GAN derived synthetic data to the training datasets in two
brain segmentation tasks, leading to improvements in Dice Similarity Coefficient
(DSC) of between 1 and 5 percentage points under different conditions, with the
strongest effects seen fewer than ten training image stacks are available.

1 Introduction

Data augmentation is commonly used by many deep learning approaches in the

presence of limited training data. Increasing the number of training examples
through the rotation, reflection, cropping, translation and scaling of existing im-
ages is common practice during the training of learning algorithms, allowing for
the number of samples in a dataset to be increased by factors of thousands [12].
Populating the training data with realistic, if synthetic, data in this way can
significantly reduce overfitting and thus not only improve the accuracy but also
∗
All work done while at Imperial College London
the generalisation ability of deep learning approaches. This is of particular im-
portance in Convolutional Neural Networks (CNNs) which cannot easily learn
rotationally invariant features unless there are sufficient examples at different
rotations in the training data. This paper investigates using a GAN to model
the underlying distribution of training data to allow for additional synthetic data
to be sampled and used to augment the real training data.
First proposed in [6], GANs are a class of neural networks which learn to
generate synthetic samples with the same characteristics as a given training
distribution. In the case of images, this involves learning to produce images (via
a generator) which are visually so similar to a set of real images that an adversary
(the discriminator) cannot detect them. The original formulation has been built
on to address problems such as training stability [16], low resolution [10], and
the absence of a true image quality based loss function [2], and applied to tasks
such as super resolution [13], reconstructing images from a minimal data [22]
and anomaly detection [19].
Various methods for using GANs to expand training datasets have been re-
cently proposed. In [20], the authors use an adversarial network to improve the
quality of simulated images, and use these for further training. In [1], the authors
train a conditional GAN on unlabelled data to generate alternative versions of a
given real image, and in [23], the authors use a similar GAN to impose emotions
on neutral faces to expand underrepresented classes. However, the use of non-
conditional GANs to augment training data directly as a preprocessing step with
no additional data has only very recently been explored [5,15], with promising
results in medical image classification tasks.

1.1 Motivation

Imaging features can be divided into two categories, measuring either pertinent
or non-pertinent variance. Pertinent features are those which are important to
whatever information the user wishes to extract. In medical imaging, these are
features such as the size, shape, intensity and location of key components such as
organs or lesions. Non-pertinent features are those which vary between images
but are unrelated to the information the user wishes to extract. Examples of
these are global intensity differences, position within the image field of view
and appearance of unrelated anatomy. Exactly which features are pertinent or
non-pertinent will depend on the application, and may not be known a-priori.
A lot of non-pertinent variance can easily be removed from a dataset. Com-
mon methods include intensity normalisation, cropping, and registration to a
standard space. These processes substantially simplify the data distribution, and
importantly, can be applied to test instances. Keeping too much non-pertinent
variance can not only occlude the diagnostically important information, but also
lead to overfitting, especially in the small datasets often used in medical imaging.
Data augmentation is an alternative to removing non-pertinent variance. One
of the goals of data augmentation is to populate the data with a large amount
of synthetic data in the directions of these non-pertinent sources of variance.
The aim of this is to reduce this variance to noise, removing any coincidental
correlation with labels and preventing its use as a discriminative feature.
As noted in [11], there is a tendency within medical imaging to remove non-
pertinent variance rather than use augmentation. This due to both the ease
with which much of the non-pertinent variance can be removed, and the lack of
suitable augmentation procedures for many sources of non-pertinent variance.
This is reflected in [9], where the authors choose to only employ reflection and
intensity augmentation for brain lesion segmentation, with even the latter omit-
ted when using larger datasets. On the other hand, in [18] the authors benefit
from extensive augmentation in their application of microscopy images, par-
ticularly through random elastic deformations. This demonstrates how careful
consideration of the application will inform which types of augmentation are
appropriate. While random elastic deformations may be an appropriate model
for microscopy images, in which the objects of interest (cells) are generally fluid
and unconstrained, applying the same procedure to brain images could lead to
certain anatomical constraints such as symmetry, rigidity, and structure being
disregarded. In addition, some sources of non-pertinent variance can be nei-
ther removed nor augmented by traditional means. For example, patient specific
variation in non-relevant anatomy, where it may not be possible to remove this
anatomy through cropping, or to define an accurate enough model to augment
this variance with realistic cases.
GANs offer a potentially valuable addition to the arsenal of augmentation
techniques which are currently available. One of the main potential advantages
of GANs is that they take many decisions away from the user, in much the
same way that deep learning removes the need for “hand crafted” features. An
ideal GAN will transform the discrete distribution of training samples into a
continuous distribution, thereby simultaneously applying augmentation to each
source of variance within the dataset. For example, given a sufficient number of
training examples at different orientations, a GAN will learn to produce examples
at any orientation, replicating the effects of applying rotation augmentation.
While orientation is a source of variance which can easily be augmented or
removed using traditional methods, consider instead a more challenging source
of variance such as ventricle size in brain imaging. Again, given a sufficient
number of training examples of patients with different discrete ventricle sizes, a
trained GAN will be able to produce examples along the continuum of all sizes.
To perform the same kind of augmentation using deformations would involve a
complex model of realistic ventricle size, shape and impact on the surrounding
anatomy. By simultaneously learning the distribution of all sources of variance,
the GAN infers this model directly from the available data.
One potential limitation of using GANs for augmentation is their ability to
generate images with a high enough quality. While improvements have been
made, GANs cannot be relied upon to produce images with perfect fidelity. This
is not a problem for traditional augmentation procedures which do not signifi-
cantly degrade the images. However, both [4] and [17] demonstrate that complete
realism is not necessary to improve results with synthetic data. Whether the ad-
vantage of additional data is outweighed by the disadvantage of lower quality
images is one of the questions we address in this paper.

1.2 Contribution

The results reported in [5,15] suggest that GANs can have a significant benefit
when used for data augmentation in some classification tasks. In this paper we
thoroughly investigate this use of GANs in different domains for the purpose of
medical image segmentation. An in depth investigation into the effects of GAN
augmentation is first carried out on a complex multi-class Computed Tomogra-
phy (CT) Cerebrospinal Fluid (CSF) segmentation task using two segmentation
architectures. By choosing not to co-register the images in this dataset, we are
able to examine how GAN augmentation compares and interacts with rotation
augmentation. The transferability of the method is then evaluated by applying
it to a second dataset of Fluid-Attenuated Inversion Recovery (FLAIR) Mag-
netic Resonance (MR) images for the purpose of single-class White Matter Hy-
perintensity (WMH) segmentation. This is a well studied problem, and poses
challenges typical to medical image segmentation tasks.
Aside from establishing whether GAN augmentation can lead to an improve-
ment in network performance, we answer the following five important questions:
– Does the choice of segmentation network affect this improvement?
– How does GAN augmentation compare to rotation augmentation?
– Does the amount of synthetic data added affect this improvement?
– Does the amount of available real data affect this improvement?
– Does the approach generalise to multiple datasets?
We also explore the distribution of generated images to better understand
what modes of augmentation are provided. This allows us to confirm that the
GANs are producing images which are different to those in the dataset. We
show how images are generated with the same pathology, but different unrelated
anatomy, and vice versa, demonstrating the ability to perform these particularly
challenging forms of augmentation.

2 Methods

We use a Progressive Growing of GANs (PGGAN) network [10] to generate syn-

thetic data. PGGAN was chosen on the basis of its training stability at large
image sizes and apparent robustness to hyperparameter selection.Whether the
choice of GAN architecture will affect the quality of the augmentation is un-
clear, however there is evidence [14] to suggest that different GAN architectures
produce results which are, on average, not significantly different from each other.
We train a PGGAN on 80k patches sampled from the available training data
set as a preprocessing step prior to training a segmentation CNN. The PGGAN
is trained on multi-channel image patches containing both the acquired image
and manual segmentation label, thereby learning the manifold containing this
joint data distribution. Synthetic examples are then sampled randomly from
this manifold using the trained generator and used to augment the same 80k
patches, forming the training data used when training the subsequent segmen-
tation network. The only alteration to the default PGGAN architecture was to
concatenate a 32x32 layer of Gaussian noise at the start of the fourth (32x32)
resolution level when training on CT data. This change was found empirically
to produce CT images with a more realistic noise pattern. The networks were
configured to produce images with a size of 128-by-128px with 6 resolution levels.
Segmentation networks were evaluated using training, validation, and test
sets.Performance (measured by DSC) on the validation set was monitored during
training with the best model at the conclusion of training applied to the test set.
A set of experiments were designed to assess effect of introducing GAN de-
rived synthetic data to a segmentation task. In these experiments, a number of
key variables were modified:
Amount of available real data: To simulate a situation with limited training
data, the amount of training images was artificially reduced by randomly se-
lecting a percentage of the available images, prior to sampling the 80k training
patches. We performed experiments with percentage reductions in available data
ranging from 10% to 90%. Note that this reduction is enforced for both the GAN
and segmentation network training stages, ensuring the GAN is never exposed
to more labelled data than the corresponding segmentation network.
Amount of additional synthetic data: To investigate whether the amount of
synthetic data added to the real data affects the performance of a segmentation
network, experiments were run with different amounts of additional synthetic
data. To ensure equal access to the information available in the real data between
experiments, synthetic data is added to the real data, increasing the size of the
dataset, rather than replacing real data. The amount of additional patches is
expressed as a percentage of the real patches. For example an experiment with
+50% synthetic data would use 120k patches (80k real and 40k synthetic).
Dataset: Two different datasets are explored to assess the ability for GAN
augmentation to generalise across segmentation tasks. The first dataset contains
CT images with manually delineated CSF labels split into into 3 classes: cortical
CSF, brain stem CSF and ventricular CSF. Data is split in the same way as
in [3], using the same preprocessing and sampling procedures. This provides 500
manually labelled training image slices, with an additional 282 validation slices,
from 101 subjects. For these experiments, the average DSC is used as the primary
measure of performance, though results across each class are also analysed. The
second dataset contains FLAIR images with manual WMH segmentations. 147
FLAIR image stacks were acquired as described in [21]. These were manually
segmented, before being bias corrected, brain extracted, rigidly co-registered
and intensity normalised as in [8], and randomly split into equal sized training,
validation and test sets. By selecting two dissimilar tasks (multi- and single-
class segmentation) across two modalities (CT and MR) we cover a wide range
of likely applications for GAN augmentation.
Segmentation network: We investigate three different segmentation networks
across the experiments. In [3], the authors show that both UNet and Residual
UNet (UResNet) [7] architectures perform well on this CT dataset, we therefore
choose to explore both of these. The same hyperparameters were used as in [3].
DeepMedic [9] is a popular general purpose segmentation algorithm which has
been shown to perform well in many applications, and was therefore chosen as a
third network to explore. DeepMedic was modified only so as to accept 128x128
2-dimensional (2D) patches. Between these three, we represent the most popular
CNN architectures currently in use.

Augmentation: As discussed in Section [11] extensive augmentation, beyond

simple reflection, is rarely used in brain imaging due to the variety of prepro-
cessing options available and anatomical constraints of the brain. However, in
order to examine the interaction of GAN and rotation augmentation we elect
not to perform coregistration on the CT dataset. Of the other common forms of
augmentation, reflection augmentation is routinely performed in all experiments,
translation augmentation is encapsulated in the patch based approach, intensity
augmentation is obviated by intensity normalisation, and deformations are not
considered due to anatomical constraints (preserving shape, symmetry etc.).

Table 1: Summary of experiments

% of available real % added syn- Segmentation Dataset Augmentation type
data sampled from thetic data network
100, 50, 10 0, 50, 100 UNet, UResNet CT Rotation+GAN
100, 50, 10 0, 100 UNet CT None, GAN, rota-
tion, rotation+GAN
100, 50, 10 0, 12.5, 25, 37.5, UNet CT Rotation+GAN
50, 100
100, 90...20, 10 0, 50 UNet CT Rotation+GAN
100, 50, 10 0, 50, 100 DeepMedic MR GAN

Table 1 summarises the five sets of experiments which were carried out to
answer the questions posed earlier. In each experiment, the segmentation network
is treated as a black box and unchanged. This provides a fair platform upon
which to observe the effects of GAN augmentation by ensuring that any changes
in performance are as a result of the additional synthetic data, and not of changes
in the network itself. GAN training took 36 hours, each UNet took 4 hours, each
Res-UNet took 24 hours and each DeepMedic network took 24 hours on an
Nvidia GTX 1080 Ti or similar GPU. All segmentation experiments on CT were
repeated 8 times, while those on MR were repeated 14 times to compensate for
a higher observed variance. Examples of real and synthetic patches generated
for each dataset can be seen in Figure 1.
 (a) Real CT (b) Synthetic CT (c) Real MRI (d) Synthetic MRI

Fig. 1: Examples of real and GAN generated synthetic patches. Left CSF. Red:
Cortical CSF. Green: Brain stem CSF. Blue: Ventricular CSF. Right: WMH.

3 Results
3.1 Segmentation results
The following tables and graphs show the results over the two sets of experi-
ments. All tables show the average DSC, with the standard deviation in brack-
ets. Results which are statistically different from the baseline (2-tailed t-test, 5%
significance level) are shown in bold.

Table 2: CSF segmentation on CT: Results with different proportions of the

available training data and varying amounts of additional synthetic data using
UNet and UResNet architectures.
Available data
UNet UResNet
100% 50% 10% 100% 50% 10%
Additional

0% 88.9 (0.51) 86.0 (0.50) 76.9 (0.58) 86.8 (0.82) 82.7 (1.55) 72.5 (1.98)
Data

50% 89.2 (0.30) 87.3 (0.46) 78.6 (1.04) 86.3 (1.44) 84.3 (1.31) 74.3 (1.63)
100% 89.3 (0.39) 86.9 (0.36) 78.4 (0.99) 86.3 (1.24) 84.1 (1.32) 74.7 (1.18)

Table 3: CSF segmentation on CT: UNet results with different proportions

of the available training data and different augmentation techniques.
Available data
100% 50% 10%
No augmentation 88.1 (0.32) 85.0 (0.58) 75.1 (0.60)
GAN augmentation 88.4 (0.41) 85.6 (1.33) 76.3 (1.77)
Rotation augmentation 88.9 (0.51) 86.0 (0.50) 76.9 (0.58)
GAN + Rotation augmentation 89.3 (0.39) 86.9 (0.36) 78.4 (0.99)
Fig. 2: CSF segmentation on CT: Left: Average DSC for each class (coloured)
and mean across classes (black) as availability of real data varies. Solid lines show
performance without GAN augmentation, dashed lines show performance with
+50% synthetic data, and dot/dashed lines show the difference, indicating the
improvement seen with GAN augmentation. Right: Average DSC observed using
a UNet as synthetic data is added, when 100%, 50% and 10% of the total amount
of real data is used. Each coloured dot represents an experiment. Black circles
show the mean with filled circles indicating results significantly different from
the baseline.

Table 4: WMH segmentation on MRI: Results with different proportions of

the available training data and varying amounts of additional synthetic data.
Available data
100% 50% 10%
Additional

0% 66.0 (1.26) 61.4 (2.67) 52.2 (6.65)

Data

50% 65.5 (1.21) 63.7 (0.69) 57.2 (4.09)

100% 64.8 (1.34) 62.8 (1.17) 55.7 (4.26)

3.2 Qualitative evaluation

As well as the quantitative segmentation results, the generated MR images were

also compared to their nearest neighbour in the training set to elucidate what
extra information GAN augmentation provides. These images, a subset of which
are shown in Figure 3, were examined looking for cases where: lesions were
duplicated on different anatomy; lesions were changed whilst anatomy stays the
same; the nearest neighbour is substantially different. The latter indicates the
GAN has learned a smooth manifold leading to potentially novel anatomy.
 (a) 5 training images (b) 25 training images (c) 50 training images

Fig. 3: Synthetic images (top of pair) with their nearest neighbours in the training
set (bottom of pair) from GANs trained on patches from 5, 25 and 50 real MR
images. Some local signs of successful augmentation are indicated using green
(same lesions, different anatomy) and yellow (same anatomy, different lesions)
arrows, and novel images (new anatomy and lesions) are shown with blue dots.

4 Discussion

It can be seen from across all of the results that GAN augmentation can provide a
modest but significant improvement in segmentation performance in many cases.
By far the strongest factor controlling the improvement seen is the amount of
real data available for training. There is a clear trend across all results that
the greatest improvements can be seen in the cases where real data is the most
limited. However, Figure 2 suggests that there is perhaps a drop in improvement
seen at the very lowest levels of available data, likely due to there being too little
data to properly train the GAN. Results on the CT data suggest that there are
no circumstances in which using synthetic data leads to worse results even when
large amounts of real data is available. However, this is not reflected in the MR
results in Table 4, where a loss in DSC is observed when all the data is used. This
suggests a tipping point associated with the amount of available data, beyond
which GAN augmentation harms rather than helps.
Another benefit of GAN augmentation can be seen in the DSC observed on
the individual CSF classes in Figure 2. A ratio of 1.35:4.35:1 between ventricu-
lar, cortical and brain stem CSF classes in the training set indicates a moderate
class imbalance, with examples of the former and latter being relatively limited.
Figure 2 shows that it is these two classes which benefit most from GAN augmen-
tation. Brain stem CSF segmentation appears to benefit the most, though this
can be attributed to ventricular CSF segmentation being an inherently easier
proposition, and therefore being consistently well segmented anyway.
Table 2 shows that there is little difference in the effect of GAN augmentation
when using different segmentation networks. This, coupled with the WMH re-
sults, suggests that GAN augmentation may benefit any segmentation network,
regardless of architecture. Similarly, Figure 2 shows that a similar level improve-
ment is seen across a broad range of additional synthetic data quantities. The
amount of synthetic data is therefore not an additional parameter which needs
to be finely tuned. This, coupled with the earlier observation that synthetic data
rarely impairs performance, makes GAN augmentation a practical proposition.
It is interesting to note that the improvements given by using both traditional
and GAN augmentation, as seen in Table 3, are consistently more than the sum
of the improvements given by using the two methods separately. This provides
strong evidence that the additional information provided by the two augmen-
tation methods are independent. It also suggests that when used together they
are potentially synergistic, an observation which agrees with the results in [5].
This could be due to the two methods acting in different ways, with GANs pro-
viding an effective alternative to traditional augmentation when attempting to
interpolate within the training distribution, but cannot extrapolate beyond its
extremes without the aid of traditional augmentation like rotation.
Figure 3 provides an interesting insight into what additional information is
being provided by GAN augmentation. In the case of 5 training images, it is
clear that each generated image is based heavily on an image from the training
set. This is unsurprising as there are very few images to train on, and little
variation which can be learned. However, there are subtle differences present in
the majority of synthetic images. There are cases where lesions present in the real
image are not reproduced in the synthetic image, as well as cases where the shape
and number of lesions present in the synthetic image differ from those in the
real image. Both of these effects can be extremely valuable to prevent overfitting
when training a model - the former decoupling the presence of lesions from
the surrounding anatomy, and the latter providing more variety of pathology.
When the number of training images increases to 25, we begin to see cases where
there are no close matches in the training set, in addition to the cases of differing
anatomy and pathology seen previously. This trend gets even stronger in Figure 3
where all 50 training images are used. There are often substantial differences
between the synthetic images and their closest real image, suggesting that the
GAN has learned to produce data substantially beyond what was provided to
it. We also observe that these modifications appear reasonable in all cases, with
no obvious unrealistic lesions or anatomy being synthesised.

4.1 Conclusion

This paper has investigated augmenting training data using GAN derived syn-
thetic images, and demonstrated that this can improve results across two seg-
mentation tasks. The approach has been shown to work best in cases of limited
data, either through a lack of real data or as a result of class imbalance. GAN
augmentation requires little overhead, involving only the training of a single
out-of-the-box GAN, does not involve optimising additional parameters, and has
been shown to be low-risk by not hurting performance when training data is lim-
ited. A conservative interpretation of the results from the typical tasks explored
here suggests that in cases where 5 − 50 labelled image volumes are available,
augmenting these with an additional 10 − 100% GAN derived synthetic patches
has the potential to lead to significant improvements in DSC.
One major advantage that traditional augmentation has over GAN augmen-
tation is the ability to extrapolate. GANs can provide an effective way to fill in
gaps in the discrete training data distribution and augment sources of variance
which are difficult to augment in other ways, but will not extend the distribu-
tion beyond the extremes of the training data. In general, appropriate traditional
augmentation procedures should be used to extrapolate and extend the manifold
of semantically viable images. GANs can then be used to interpolate between
the discrete points on this manifold, providing an additional data driven source
of augmentation. Future work will involve investigating GAN augmentation in
other areas, and to evaluate the impact of different GAN architectures.

References
1. Antoniou, A., Storkey, A., Edwards, H.: Data augmentation generative adversarial
networks. arXiv preprint arXiv:1711.04340 (2017)
2. Arjovsky, M., Chintala, S., Bottou, L.: Wasserstein GAN. arXiv preprint
arXiv:1701.07875 (2017)
3. Chen, L., Bentley, P., Mori, K., Misawa, K., Fujiwara, M., Rueckert, D.: DRINet
for medical image segmentation. IEEE Transactions on Medical Imaging (2018)
4. Dosovitskiy, A., Fischer, P., Ilg, E., Hausser, P., Hazirbas, C., Golkov, V., Van
Der Smagt, P., Cremers, D., Brox, T.: Flownet: Learning optical flow with convo-
lutional networks. In: International Conference on Computer Vision. pp. 2758–66
(2015)
5. Frid-Adar, M., Klang, E., Amitai, M., Goldberger, J., Greenspan, H.: Synthetic
data augmentation using GAN for improved liver lesion classification. arXiv
preprint arXiv:1801.02385 (2018)
6. Goodfellow, I., Pouget-Abadie, J., Mirza, M., Xu, B., Warde-Farley, D., Ozair,
S., Courville, A., Bengio, Y.: Generative adversarial nets. In: Advances in neural
information processing systems. pp. 2672–80 (2014)
 7. Guerrero, R., Qin, C., Oktay, O., Bowles, C., Chen, L., Joules, R., Wolz, R., Valdés-
Hernández, M., Dickie, D., Wardlaw, J., et al.: White matter hyperintensity and
stroke lesion segmentation and differentiation using convolutional neural networks.
Neuroimage: Clinical 17, 918–34 (2018)
8. Huppertz, H.J., Wagner, J., Weber, B., House, P., Urbach, H.: Automated quanti-
tative FLAIR analysis in hippocampal sclerosis. Epilepsy research 97(1-2), 146–56
(2011)
9. Kamnitsas, K., Ledig, C., Newcombe, V.F., Simpson, J.P., Kane, A.D., Menon,
D.K., Rueckert, D., Glocker, B.: Efficient multi-scale 3d cnn with fully connected
crf for accurate brain lesion segmentation. Medical image analysis 36, 61–78 (2017)
10. Karras, T., Aila, T., Laine, S., Lehtinen, J.: Progressive growing of GANs for
improved quality, stability, and variation. arXiv preprint arXiv:1710.10196 (2017)
11. Krivov, E., Pisov, M., Belyaev, M.: MRI augmentation via elastic registration for
brain lesions segmentation. In: International MICCAI Brainlesion Workshop. pp.
369–80. Springer (2017)
12. Krizhevsky, A., Sutskever, I., Hinton, G.E.: Imagenet classification with deep con-
volutional neural networks. In: Advances in neural information processing systems.
pp. 1097–105 (2012)
13. Ledig, C., Theis, L., Huszár, F., Caballero, J., Cunningham, A., Acosta, A., Aitken,
A.P., Tejani, A., Totz, J., Wang, Z., et al.: Photo-realistic single image super-
resolution using a generative adversarial network. In: CVPR. p. 4. No. 3 (2017)
14. Lucic, M., Kurach, K., Michalski, M., Gelly, S., Bousquet, O.: Are GANs created
equal? a large-scale study. arXiv preprint arXiv:1711.10337 (2017)
15. Madani, A., Moradi, M., Karargyris, A., Syeda-Mahmood, T.: Chest x-ray gen-
eration and data augmentation for cardiovascular abnormality classification. In:
Medical Imaging 2018: Image Processing. vol. 10574, p. 105741M. International
Society for Optics and Photonics (2018)
16. Radford, A., Metz, L., Chintala, S.: Unsupervised representation learning with deep
convolutional generative adversarial networks. arXiv preprint arXiv:1511.06434
(2015)
17. Richter, S.R., Vineet, V., Roth, S., Koltun, V.: Playing for data: Ground truth
from computer games. In: European Conference on Computer Vision. pp. 102–18.
Springer (2016)
18. Ronneberger, O., Fischer, P., Brox, T.: U-net: Convolutional networks for biomedi-
cal image segmentation. In: International Conference on Medical image computing
and computer-assisted intervention. pp. 234–41. Springer (2015)
19. Schlegl, T., Seeböck, P., Waldstein, S.M., Schmidt-Erfurth, U., Langs, G.: Unsu-
pervised anomaly detection with generative adversarial networks to guide marker
discovery. arXiv preprint arXiv:1703.05921 (2017)
20. Shrivastava, A., Pfister, T., Tuzel, O., Susskind, J., Wang, W., Webb, R.: Learning
from simulated and unsupervised images through adversarial training. In: CVPR.
p. 5. No. 4 (2017)
21. Valdés Hernández, M.d.C., Armitage, P.A., Thrippleton, M.J., Chappell, F.,
Sandeman, E., Muñoz Maniega, S., Shuler, K., Wardlaw, J.M.: Rationale, design
and methodology of the image analysis protocol for studies of patients with cerebral
small vessel disease and mild stroke. Brain and behavior 5(12), e00415 (2015)
22. Yeh, R.A., Chen, C., Lim, T.Y., Schwing, A.G., Hasegawa-Johnson, M., Do, M.N.:
Semantic image inpainting with deep generative models. In: CVPR. p. 4. No. 3
(2017)
23. Zhu, X., Liu, Y., Qin, Z., Li, J.: Data augmentation in emotion classification using
generative adversarial networks. arXiv preprint arXiv:1711.00648 (2017)