GAN-based synthetic brain MR image generation

Abstract

In medical imaging, it remains a challenging and valuable goal how to generate realistic medical images completely different from the original ones; the obtained synthetic images would improve diagnostic reliability, allowing for data augmentation in computer-assisted diagnosis as well as physician training. In this paper, we focus on generating synthetic multi-sequence brain Magnetic Resonance (MR) images using Generative Adversarial Networks (GANs). This involves difficulties mainly due to low contrast MR images, strong consistency in brain anatomy, and intra-sequence variability. Our novel realistic medical image generation approach shows that GANs can generate 128 χ 128 brain MR images avoiding artifacts. In our preliminary validation, even an expert physician was unable to accurately distinguish the synthetic images from the real samples in the Visual Turing Test.

Full text

GAN-BASED SYNTHETIC BRAIN MR IMAGE GENERATION
Changhee Han1,Hideaki Hayashi2,Leonardo Rundo3,Ryosuke Araki4,Wataru Shimoda5
Shinichi Muramatsu6,Yujiro Furukawa7,Giancarlo Mauri3,Hideki Nakayama1
1Grad. School of Information Science and Technology, The University of Tokyo, Tokyo, Japan
2Dept. of Advanced Information Technology, Kyushu University, Fukuoka, Japan
3Dept. of Informatics, Systems and Communication, University of Milano-Bicocca, Milan, Italy
4Grad. School of Engineering, Chubu University, Aichi, Japan
5Dept. of Informatics, The University of Electro-Communications, Tokyo, Japan
6Grad. School of Science and Technology, Shinshu University, Nagano, Japan
7Kanto Rosai Hospital, Kanagawa, Japan
ABSTRACT
In medical imaging, it remains a challenging and valuable
goal how to generate realistic medical images completely dif-
ferent from the original ones; the obtained synthetic images
would improve diagnostic reliability, allowing for data aug-
mentation in computer-assisted diagnosis as well as physi-
cian training. In this paper, we focus on generating synthetic
multi-sequence brain Magnetic Resonance (MR) images us-
ing Generative Adversarial Networks (GANs). This involves
difficulties mainly due to low contrast MR images, strong
consistency in brain anatomy, and intra-sequence variability.
Our novel realistic medical image generation approach shows
that GANs can generate 128×128 brain MR images avoiding
artifacts. In our preliminary validation, even an expert physi-
cian was unable to accurately distinguish the synthetic images
from the real samples in the Visual Turing Test.
Index Terms—Generative Adversarial Networks, Syn-
thetic Medical Image Generation, Brain MRI, Data Augmen-
tation, Physician Training, Visual Turing Test
1. INTRODUCTION
Along with classic methods [1], Convolutional Neural Net-
works (CNNs) have recently revolutionized medical image
analysis [2], including brain Magnetic Resonance Imaging
(MRI) segmentation [3]. However, CNN training demands
extensive medical data that are laborious to obtain [4]. To
overcome this issue, data augmentation techniques via recon-
structing original images are common for better performance,
such as geometry and intensity transformations [5, 6].
However, those reconstructed images intrinsically resem-
ble the original ones, leading to limited performance improve-
ment in terms of generalization abilities; thus, generating re-
This work was partially supported by the Graduate Program for Social
ICT Global Creative Leaders of The University of Tokyo by JSPS.
Original Brain
MR Images
Synthetic Images for
Data Augmentation
T1 T1c
T2 FLAIR
Synthetic Images for
Physician Training
(GAN)
Generate
Generate
(Conditional GAN)
Realistic Tumors
with Desired Size/Location
by Adding Conditioning
Realistic Tumors
in Random Locations
Fig. 1. Potential applications of the proposed GAN-based
synthetic brain MR image generation: (1) data augmentation
for better diagnostic accuracy by generating random realistic
images giving insights in classification; (2) physician training
for better understanding various diseases to prevent misdiag-
nosis by generating desired realistic pathological images.
alistic (similar to the real image distribution) but completely
new images is essential. In this context, Generative Adversar-
ial Network (GAN)-based data augmentation has excellently
performed in general computer vision tasks. It attributes to
GAN’s good generalization ability from matching the gen-
erated distribution from noise variables to the real one with
a sharp value function. Especially, Shrivastava et al. (Sim-
GAN) outperformed the state-of-the-art with a relative 21%
improvement in eye-gaze estimation [7].
So, how can we generate realistic medical images com-
pletely different from the original samples? Our aim is to gen-
978-1-5386-3636-7/18/$31.00 ©2018 IEEE 734
2018 IEEE 15th International Symposium on Biomedical Imaging (ISBI 2018)
April 4-7, 2018, Washington, D.C., USA

erate synthetic multi-sequence brain MR images using GANs,
which is essential in medical imaging to increase diagnos-
tic reliability, such as via data augmentation in computer-
assisted diagnosis as well as physician training and teach-
ing (Fig. 1) [8]. However, this is extremely challenging—
MR images are characterized by low contrast, strong visual
consistency in brain anatomy, and intra-sequence variability.
Our novel GAN-based approach for medical data augmen-
tation adopts Deep Convolutional GAN (DCGAN) [9] and
Wasserstein GAN (WGAN) [10] to generate realistic images,
and an expert physician validates them via the Visual Turing
Test [11].
Research Questions. We mainly address two questions:
•GAN Selection: Which GAN architecture is well-
suited for realistic medical image generation?
•Medical Data Augmentation: How can we handle
MR images with specific intra-sequence variability?
Contributions. Our main contributions are as follows:
•MR Image Generation: This research shows that
WGAN can generate realistic multi-sequence brain
MR images, possibly leading to valuable clinical ap-
plications: data augmentation and physician training.
•Medical Image Generation: This research provides
how to exploit medical images with intrinsic intra-
sequence variability towards GAN-based data augmen-
tation for medical imaging.
2. GENERATIVE ADVERSARIAL NETWORKS
Since the breakthrough paper by Goodfellow et al. in
2014 [12], GANs have shown promising results for image
generation in general computer vision [13]. GANs generate
highly realistic images, without a well-defined objective func-
tion associated with difficult training accompanying oscilla-
tions and mode collapse—i.e., a common failure case where
the generator learns with extremely low variety. Whereas
Variational Autoencoders (VAEs) [14], the other most used
deep generative models, have an objective likelihood function
to optimize, and could so generate blurred samples because
of the injected noise and imperfect reconstruction [15].
Therefore, many medical imaging researchers have begun
to use GANs recently, such as in image super-resolution [16],
anomaly detection [17], and estimating CT images from the
corresponding MR images [18]. As GANs allow adding con-
ditioning on the class labels and images, they often use such
conditional GANs to produce desired images, while it makes
learning robust latent spaces difficult.
Differently from a very recent work of GANs for biolog-
ical image synthesis (fluorescence microscopy) [19], to the
best of our knowledge, this is the first GAN-based realistic
brain tumor MR image generation approach aimed at data
augmentation and physician training. Instead of reconstruct-
ing real brain MR images themselves with respect to geom-
etry/intensity, a completely different approach—generating
T1 (Real, 128 × 128/64 × 64)
T2 (Real, 128 × 128/64 × 64)
T1c (Real, 128 × 128/64 × 64)
FLAIR (Real, 128 × 128/64 × 64)
Fig. 2. Example real MR images used for training the GANs:
the resized sagittal multi-sequence brain MRI scans of pa-
tients with HGG on the BRATS 2016 training dataset [20].
novel realistic images using GANs—may become a clinical
breakthrough.
3. MATERIALS AND METHODS
Towards clinical applications utilizing realistic brain MR im-
ages, we generate synthetic brain MR images from the orig-
inal samples using GANs. Here, we compare the most used
two GANs, DCGAN and WGAN, to find a well-suited GAN
between them for medical image generation—it must avoid
mode collapse and generate realistic MR images with high
resolution.
3.1. The BRATS 2016 Dataset
This paper exploits a dataset of multi-sequence brain MR im-
ages to train GANs with sufficient data and resolution, which
was originally produced for the Multimodal Brain Tumor
Image Segmentation Benchmark (BRATS) challenge [20].
In particular, the BRATS 2016 training dataset contains
220 High-Grade Glioma (HGG) and 54 Low-Grade Glioma
(LGG) cases, with T1-weighted (T1), contrast enhanced T1-
weighted (T1c), T2-weighted (T2), and Fluid Attenuation
Inversion Recovery (FLAIR) sequences—they were skull
stripped and resampled to isotropic 1mm ×1mm ×1mm
resolution with image dimension 240×240 ×155; among the
different sectional planes, we use the sagittal multi-sequence
scans of patients with HGG to show that our GANs can gener-
ate a complete view of the whole brain anatomy (allowing for
visual consistency among the different brain lobes), including
also severe tumors for clinical purpose.
3.2. Proposed GAN-based Image Generation Approach
3.2.1. Pre-processing
We select the slices from #80 to #149 among the whole 240
slices to omit initial/final slices, since they convey a negligi-
ble amount of useful information and could affect the training.
The images are resized to both 64 ×64 and 128 ×128 from
240 ×155 for better GAN training (DCGAN architecture re-
sults in stable training on 64 ×64 [9], and so 128 ×128 is
reasonably a high-resolution). Fig. 2 shows some real MR im-
ages used for training; each sequence contains 15,400 images
with 220 patients ×70 slices (61,600 in total).
735

3.2.2. GAN-based MR Image Generation
DCGAN and WGAN generate six types of images as follows:
•T1 sequence (128 ×128) from the real T1;
•T1c sequence (128 ×128) from the real T1c;
•T2 sequence (128 ×128) from the real T2;
•FLAIR sequence (128 ×128) from the real FLAIR;
•Concat sequence (128 ×128) from concatenating the
real T1, T1c, T2, and FLAIR (i.e., feeding the model
with samples from all the MRI sequences);
•Concat sequence (64 ×64) from concatenating the real
T1, T1c, T2, and FLAIR.
Concat sequence refers to a new ensemble sequence for an
alternative data augmentation, containing the features of all
four sequences. We also generate 64 ×64 Concat images to
compare the generation performance in terms of image size.
DCGAN. DCGAN [9] is a standard GAN [12] with a con-
volutional architecture for unsupervised learning; this gen-
erative model uses up-convolutions interleaved with ReLu
non-linearity and batch-normalization.
Let pdata be a generating distribution over data x. The
generator G(z;θg)is a mapping to data space that takes a
prior on input noise variables pz(z), where Gis a neural
network with parameters θg. Similarly, the discriminator
D(x;θd)is a neural network with parameters θdthat takes
either real data or synthetic data and outputs a single scalar
probability that xcame from the real data. The discrimina-
tor Dmaximizes the probability of classifying both training
examples and samples from Gcorrectly while the generator
Gminimizes the likelihood; it is formulated as a minimax
two-player game with value function V(G, D):
min
Gmax
DV(D, G)=Ex∼pdata(x)[log D(x)]
+Ez∼pz(z)[log(1 −D(G(z)))].
(1)
This can be reformulated as the minimization of the Jensen-
Shannon (JS) divergence between the distribution pdata and
another distribution pgderived from pzand G.
DCGAN Implementation Details. We use the same DC-
GAN architecture [9] with no tanh in the generator, ELU as
the discriminator, all filters of size 4×4, and a half channel
size for DCGAN training. A batch size of 64 and Adam
optimizer with 2.0×10−4learning rate were implemented.
WGAN. WGAN [10] is an alternative to traditional GAN
training, as the JS divergence is limited, such as when it is
discontinuous; this novel GAN achieves stable learning with
less mode collapse by replacing it to the Earth Mover (EM)
distance (a.k.a. the Wasserstein-1 metrics):
W(pg,p
r) = inf
p∈(pg,pr)
E(x,x)∼px−x,(2)
where (pg,p
r)is the set of all joint distributions pwhose
marginals are pgand pr, respectively. In other words, p
T1 (DCGAN, 128 × 128)
T2 (DCGAN, 128 × 128)
T1c (DCGAN, 128 × 128)
FLAIR (DCGAN, 128 × 128)
Concat (DCGAN, 128 × 128) Concat (DCGAN, 64 × 64)
Fig. 3. Example synthetic MR images yielded by DCGAN.
implies how much mass must be transported from one distri-
bution to another. This distance intuitively indicates the cost
of the optimal transport plan.
WGAN Implementation Details. We use the same DCGAN
architecture [9] for WGAN training. A batch size of 64 and
Root Mean Square Propagation (RMSprop) optimizer with
5.0×10−5learning rate were implemented.
3.3. Clinical Validation Using the Visual Turing Test
To quantitatively evaluate how realistic the synthetic im-
ages are, an expert physician was asked to constantly clas-
sify a random selection of 50 real/50 synthetic MR im-
ages as real or synthetic shown in a random order for each
GAN/sequence, without previous training stages revealing
which is real/synthetic; Concat images were classified to-
gether with real T1, T1c, T2, and FLAIR images in equal
proportion. The so-called Visual Turing Test [11] uses binary
questions to probe a human ability to identify attributes and
relationships in images. For these motivations, it is com-
monly used to evaluate GAN-generated images, such as for
SimGAN [7]. This applies also to medical images in clinical
environments [21], wherein physicians’ expertise is critical.
4. RESULTS
This section shows how DCGAN and WGAN generate syn-
thetic brain MR images. The results include instances of syn-
thetic images and their quantitative evaluation of the realism
by an expert physician. The training took about 2(1) hours to
train each 128×128 (64×64) sequence on an Nvidia GeForce
GTX 980 GPU, increasingly learning realistic features.
4.1. MR Images Generated by GANs
DCGAN. Fig. 3 illustrates examples of synthetic images by
DCGAN. The images look similar to the real samples. Concat
images combine appearances and patterns from all the four
sequences used in training. Since DCGAN’s value function
could be unstable, it often generates hyper-intense T1-like im-
ages analogous to mode collapse for 64 ×64 Concat images,
while sharing the same hyper-parameters with 128 ×128.
736

Table 1. Visual Turing Test results by a physician for classifying real vs synthetic images. It should be noted that proximity to
50% of accuracy indicates superior performance (chance = 50%).
Accuracy (%) Real Selected as Real Real as Synt Synt as Real Synt as Synt
T1 (DCGAN, 128 ×128)70 26 24 6 44
T1c (DCGAN, 128 ×128)71 24 26 3 47
T2 (DCGAN, 128 ×128)64 22 28 8 42
FLAIR (DCGAN, 128 ×128)54 12 38 8 42
Concat (DCGAN, 128 ×128)77 34 16 7 43
Concat (DCGAN, 64 ×64)54 13 37 9 41
T1 (WGAN, 128 ×128)64 20 30 6 44
T1c (WGAN, 128 ×128)55 13 37 8 42
T2 (WGAN, 128 ×128)58 19 311139
FLAIR (WGAN, 128 ×128)62 16 34 4 46
Concat (WGAN, 128 ×128)66 31 19 15 35
Concat (WGAN, 64 ×64)53 18 32 15 35
T1 (WGAN, 128 × 128)
T2 (WGAN, 128 × 128)
T1c (WGAN, 128 × 128)
FLAIR (WGAN, 128 × 128)
Concat (WGAN, 128 × 128) Concat (WGAN, 64 × 64)
Fig. 4. Example synthetic MR images yielded by WGAN.
WGAN. Fig. 4 shows the example output of WGAN in each
sequence. Outperforming remarkably DCGAN, WGAN suc-
cessfully captures the sequence-specific texture and the ap-
pearance of the tumors while maintaining the realism of the
original brain MR images. As expected, 128 ×128 Concat
images tend to have more messy and unrealistic artifacts than
64 ×64 Concat ones, especially around the boundaries of the
brain, due to the introduction of unexpected intensity patterns.
4.2. Visual Turing Test Results
Table 1 shows the confusion matrix concerning the Visual
Turing Test. Even the expert physician found classifying real
and synthetic images challenging, especially in lower resolu-
tion due to their less detailed appearances unfamiliar in clin-
ical routine, even for highly hyper-intense 64 ×64 Concat
images by DCGAN; distinguishing Concat images was easier
compared to the case of T1, T1c, T2, and FLAIR images be-
cause the physician often felt odd from the artificial sequence.
WGAN succeeded to deceive the physician significantly bet-
ter than DCGAN for all the MRI sequences except FLAIR
images (62% to 54%).
5. CONCLUSION
Our preliminary results show that GANs, especially WGAN,
can generate 128 ×128 realistic multi-sequence brain MR
images that even an expert physician is unable to accurately
distinguish from the real, leading to valuable clinical applica-
tions, such as data augmentation and physician training. This
attributes to WGAN’s good generalization ability with a sharp
value function. In this context, DCGAN might be unsuitable
due to both the inferior realism and mode collapse in terms
of intensity. We only use the slices of interest in training to
obtain desired MR images and generate both original/Concat
sequence images for data augmentation in medical imaging.
This study confirms the synthetic image quality by the
human expert evaluation, but a more objective computational
evaluation for GANs should also follow, such as Classifier
Two-Sample Tests (C2ST) [22], which assesses whether two
samples are drawn from the same distribution. Currently this
work uses sagittal MR images alone, so we will generate
coronal and transverse images in the near future. As this
research uniformly selects middle slices in pre-processing,
better data generation demands developing a classifier to only
select brain MRI slices with/without tumors.
Towards data augmentation, while realistic images give
more insights on geometry/intensity transformations in clas-
sification, more realistic images do not always assure better
data augmentation, so we have to find suitable image reso-
lutions and sequences; that is why we generate both high-
resolution images and Concat images, yet they looked more
unrealistic for the physician. For physician training, generat-
ing desired realistic tumors by adding conditioning requires
exploring extensively the latent spaces of GANs.
Overall, our novel GAN-based realistic brain MR image
generation approach sheds light on diagnostic and prognostic
medical applications; future studies on these applications are
needed to confirm our encouraging results.
737

6. REFERENCES
[1] L. Rundo, C. Militello, G. Russo, et al., “GTVcut for
neuro-radiosurgery treatment planning: an MRI brain
cancer seeded image segmentation method based on a
cellular automata model,” Nat. Comput., pp. 1–16,
2017.
[2] D. Shen, G. Wu, and H.I. Suk, “Deep learning in med-
ical image analysis,” Annu. Rev. Biomed. Eng., vol. 19,
pp. 221–248, 2017.
[3] M. Havaei, A. Davy, D. Warde-Farley, et al., “Brain
tumor segmentation with deep neural networks,” Med.
Image Anal., vol. 35, pp. 18–31, 2017.
[4] D. Rav`
ı, C. Wong, F. Deligianni, et al., “Deep learning
for health informatics,” IEEE J. Biomed. Health Inform.,
vol. 21, no. 1, pp. 4–21, 2017.
[5] O. Ronneberger, P. Fischer, and T. Brox, “U-Net: Con-
volutional networks for biomedical image segmenta-
tion,” in Proc. International Conference on Medical
Image Computing and Computer-Assisted Intervention
(MICCAI). Springer, 2015, pp. 234–241.
[6] F. Milletari, N. Navab, and S. Ahmadi, “V-Net: Fully
convolutional neural networks for volumetric medical
image segmentation,” in Proc. International Conference
on 3D Vision (3DV), 2016, pp. 565–571.
[7] A. Shrivastava, T. Pfister, O. Tuzel, et al., “Learning
from simulated and unsupervised images through ad-
versarial training,” in Proc. IEEE Conference on Com-
puter Vision and Pattern Recognition (CVPR), 2017, pp.
2107–2116.
[8] M. Prastawa, E. Bullitt, and Guido Gerig, “Simulation
of brain tumors in MR images for evaluation of segmen-
tation efficacy,” Med. Image Anal., vol. 13, no. 2, pp.
297–311, 2009.
[9] A. Radford, L. Metz, and S. Chintala, “Unsupervised
representation learning with deep convolutional gen-
erative adversarial networks,” in Proc. International
Conference on Learning Representations (ICLR), arXiv
preprint arXiv:1511.06434, 2016.
[10] M. Arjovsky, S. Chintala, and L. Bottou, “Wasserstein
generative adversarial networks,” in Proc. International
Conference on Machine Learning (ICML), 2017, pp.
214–223.
[11] T. Salimans, I. Goodfellow, W. Zaremba, V. Cheung,
A. Radford, and X. Chen, “Improved techniques for
training GANs,” in Proc. Advances in Neural Informa-
tion Processing Systems (NIPS), 2016, pp. 2234–2242.
[12] I. Goodfellow, J. Pouget-Abadie, M. Mirza, et al., “Gen-
erative adversarial nets,” in Proc. Advances in Neu-
ral Information Processing Systems (NIPS), 2014, pp.
2672–2680.
[13] J. Zhu, T. Park, P. Isola, et al., “Unpaired image-
to-image translation using cycle-consistent adversar-
ial networks,” in Proc. IEEE International Con-
ference on Computer Vision (ICCV), arXiv preprint
arXiv:1703.10593, 2017.
[14] D.P. Kingma and M. Welling, “Auto-encoding vari-
ational Bayes,” in Proc. International Conference
on Learning Representations (ICLR), arXiv preprint
arXiv:1312.6114, 2014.
[15] L. Mescheder, S. Nowozin, and A. Geiger, “Adversar-
ial variational Bayes: unifying variational autoencoders
and generative adversarial networks,” arXiv preprint
arXiv:1701.04722, 2017.
[16] D. Mahapatra, B. Bozorgtabar, S. Hewavitharanage,
et al., “Image super resolution using generative adver-
sarial networks and local saliency maps for retinal image
analysis,” in Proc. International Conference on Medical
Image Computing and Computer-Assisted Intervention
(MICCAI). Springer, 2017, pp. 382–390.
[17] T. Schlegl, P. Seeb ¨
ock, S.M. Waldstein, et al., “Unsu-
pervised anomaly detection with generative adversarial
networks to guide marker discovery,” in Proc. Interna-
tional Conference on Information Processing in Medical
Imaging. Springer, 2017, pp. 146–157.
[18] D. Nie, R. Trullo, J. Lian, et al., “Medical image synthe-
sis with context-aware generative adversarial networks,”
in Proc. International Conference on Medical Image
Computing and Computer-Assisted Intervention (MIC-
CAI). Springer, 2017, pp. 417–425.
[19] A. Osokin, A. Chessel, R.E. Carazo Salas, et al., “GANs
for biological image synthesis,” in Proc. International
Conference on Computer Vision (ICCV), arXiv preprint
arXiv:1708.04692v2, 2017.
[20] B.H. Menze, A. Jakab, S. Bauer, et al., “The multimodal
brain tumor image segmentation benchmark (BRATS),”
IEEE Trans. Med. Imaging, vol. 34, no. 10, pp. 1993–
2024, 2015.
[21] M. J. M. Chuquicusma, S. Hussein, J. Burt, et al.,
“How to fool radiologists with generative adversarial
networks? A visual Turing test for lung cancer diag-
nosis,” arXiv preprint arXiv:1710.09762v1, 2017.
[22] D. Lopez-Paz and M. Oquab, “Revisiting classifier
two-sample tests,” in Proc. International Conference
on Learning Representations (ICLR), arXiv preprint
arXiv:1610.06545, 2017.
738