Abstract
Spatial resolution is one of the main constraints in diffusion Magnetic Resonance Imaging (dMRI). Increasing resolution leads to a decrease in SNR of the diffusion images. Acquiring high resolution images without reducing SNRs requires larger magnetic fields and long scan times which are typically not applicable in the clinical settings. Currently feasible voxel size is around 1 mm\( ^{3} \) for a diffusion image. In this paper, we present a deep neural network based post-processing method to increase the spatial resolution in diffusion MRI. We utilize Generative Adversarial Networks (GANs) to obtain a higher resolution diffusion MR image in the spatial dimension from lower resolution diffusion images. The obtained real data results demonstrate a first time proof of concept that GANs can be useful in super-resolution problem of diffusion MRI for upscaling in the spatial dimension.
E. Albay and U. Demir—Equal contribution.
You have full access to this open access chapter, Download conference paper PDF
Similar content being viewed by others
Keywords
- Magnetic resonance imaging (MRI)
- Diffusion MRI (dMRI)
- Super resolution
- Generative adversarial networks (GANs)
1 Introduction
Water molecules undergo random movement and diffuse in an environment due to second law of thermodynamics. Diffusion phenomenon enables us to map fibrous substances using principles of magnetic resonance imaging (MRI). Diffusion magnetic resonance imaging (dMRI) takes advantage of signal attenuation that takes place due to diffusion of water molecules in a tissue that is being imaged. Although the signal attenuates isotropically in a free water environment, the signal shows varying attenuations in a restricted environment. This gives an opportunity of in vivo imaging of the internal structure of the human brain white matter, which contains fibrous material that restricts water molecules movements in some directions while water molecules move freely in other directions [10].
Even though dMRI allows microscopic imaging of the white matter at very high magnetic fields, spatial resolution of dMRI is restricted clinically because with the current technology, very high magnetic fields cannot be used ante-mortem. Furthermore, long scan times are not clinically feasible for microscopic resolution. At lower magnetic fields, signal to noise ratio (SNR) becomes problematic for small voxel sizes. Currently, for diffusion image volumes, clinically applicable voxel size is about 1 mm\( ^{3} \) [11], which is relatively coarse with respect to underlying microstructure of brain tissue. Diameter of neuronal axons in brain white matter is at most 30 \(\upmu \)m [7], therefore, a typical voxel contains thousands of fiber populations, possibly lying along different directions with crossings, splaying, or kissing architectures. Hence, increasing both spatial resolution and angular resolution of dMRI using post-processing techniques is desirable and would aid in post-analysis of dMRI data.
In this paper, we present a post-processing method to generate higher spatial resolution dMRI volumes based on an end-to-end generative adversarial network (GAN) framework [5]. GANs learn a mapping from low resolution diffusion MRI data to synthesize a high resolution counterpart. Its main difference from conventional methods is that GANs learn a non-linear model from pairs of low resolution-high resolution data rather than performing a blind interpolation.
2 Related Works
Only a few spatial super resolution methods for diffusion MRI were presented in the literature. Conventional methods for super resolution are typically based on up-sampling with interpolation of low resolution data. An early super resolution approach to diffusion data is based on combination of two shifted images to create an up-sampled image [14], which led to blurry results. Alternatively, a track density approach was presented to obtain super resolution in white matter fiber tracts based on tractography information, however, this method does not up-sample the underlying spatial structure of the diffusion images [1]. A Markov chain Monte Carlo method, the Metropolis-Hastings algorithm is utilized by [19] to create a generative model of local information and sharpens images according to local structure while increasing spatial resolution. This is different from our approach as it does not actually directly learn the data distribution. A recently suggested method proposes using RGB image enhancement method with diffusion images, however, not leading to clear results [18]. It is observed that diffusion weighted images are blurry and ODFs are corrupted with respect to the ground truth data. Recent studies on texture synthesis have shown that convolutional neural networks and adversarial training can be successfully applied to super-resolve images at high upscale factors [3, 12]. This was the motivation of our method, which is presented next.
3 Method
We introduce a deep GAN based single slice super-resolution model that takes a down-sampled low resolution dMRI axial slice \(I_{LR}\) and synthesizes its high resolution counterpart \(\hat{I}_{HR}\). Each down-sampled axial slice from a brain volume is upscaled to the desired resolution with a certain scale factor through bi-cubic interpolation, which is called \(I_{LR}^{bc}\). In this paper, we exemplify the spatial super resolution model with an up-sampling factor of two. The generative model takes \(I_{LR}^{bc}\), and tries to expose the high frequency details by exploring the context of low-resolution image. The trained network resolves the blurriness and generates sharp images filled with estimated missing details. Overall flow of our method is depicted in Fig. 1.
During training of the generative model, an adversarial training [5] approach is used in order to produce more realistic looking outputs. Training procedure intends to minimize the combination of an adversarial loss produced by the discriminator network, and a pixel-wise reconstruction loss (an L2 Loss) to conditionally generate samples from the high resolution image distribution. Details of this procedure are described next.
3.1 Generative Adversarial Networks
GANs have been used to figure out distribution of the input data by learning a mapping from a noise variable to the data space [5]. Recent studies show that once the distribution is learned, the model can be used to generate realistic looking samples [6, 12]. Apart from sample generation, GANs are also used to learn a mapping between contextually paired two images [6]. In our super-resolution problem, low resolution image \(I_{LR}^{bc}\) is given as a condition to the generator and it is expected that our model learns a mapping G that translates \(I_{LR}^{bc}\) to \(I_{HR}\).
There are two different neural networks in the adversarial training phase. The generative network G corresponds to mapping function between the input and the output. The purpose of the discriminator network D is distinguishing the real images from the artificially synthesized ones. While the network G aims to fool D, at the same time, D is trained to improve its accuracy. This optimization problem corresponds to a minimax game, which can be formulated as:
As long as D successfully classifies its input, G benefits from the gradient provided by the D network via its adversarial loss.
Generator Network. The architecture of the generative network is ResNet, which is composed by following the guidelines described in [8]. It consists of down-sampling layers, residual blocks and up-sampling layers. There are two down-sampling layers and each one consists of a convolution layer with stride set to 2, batch normalization layer and Leaky ReLU (LReLU) activation. There are six residual blocks in the architecture. The up-sampling blocks recover spatial resolution of the activation maps in order to reach desired height and width for a slice. An up-sampling layer contains resized convolution [13], batch normalization and LReLU activation. Additionally, a \(7 \times 7\) convolution layer with a Tanh activation is added to end of the network.
Discriminator Network. We utilize a patch based discriminator network PatchGAN [6] design which evaluates local patches of the generated image and gives an average score as a measure instead of considering the whole input. This gives more robust results than the vanilla GAN. Our patch based discriminator has 6 convolution layers followed by batch normalization except the first and the last layers. First 5 layers have LReLU activation and the last convolution layer pass its outputs to Sigmoid activation.
3.2 Training Objective
The main objective function is formed by combining the reconstruction and adversarial losses. The total loss function is optimized with back-propagation by using Adam optimizer [9]. L2 pixel-wise distances between the synthesized image and the ground truth are used as reconstruction loss. Even though it forces the network to produce a blurry output, it guides the network to roughly predict texture colors and low frequency details. Discriminator network computes a score according to quality of the generator outputs and, and is used as an adversarial loss as described in Eq. 1.
Total loss function defines the objective used in the training phase. Each component of the total loss function is governed by a coefficient \(\lambda \):
where \(\mathcal {L}_{rec}\) is reconstruction loss and \(\mathcal {L}_{adv}\) is adversarial loss.
4 Experiments
4.1 Dataset
A diffusion dataset obtained from the Human Connectome Project (HCP) is used [17], where 29 diffusion subjects are randomly selected, and 25 are used to train the network and four of them are used in testing. HCP diffusion images are multi-shell, from which a single shell is extracted for each subject, which resulted in 108 diffusion volumes per subject. It was shown that the best b-value with an SNR of 30 for a non-diffusion weighted volume is between 3000 and 4000 s/mm\(^2\) [16]. As the SNR of the HCP data is greater than 30 for the non-diffusion weighted volume, a single shell that has b-value 2000 s/mm\(^2\) is selected. DIPY [4] library is used in all the analysis. All 108 diffusion-weighted volumes including non-diffusion weighted volumes are used in training the network, and the super resolution model is applied for up-sampling of all diffusion volumes in the test stage. In Fig. 2, sample visual results for diffusion-weighted images with \(I^{bc}_{LR}\), \(\hat{I}_{HR}\) and \(I_{HR}\) from a selected subject are shown. It can be seen that \(I^{bc}_{LR}\) is blurry and our network produces \(\hat{I}_{HR}\) image with a success.
4.2 SNR Comparison and FA (Fractional Anisotropy) Maps
SNR values of \(I^{bc}_{LR}\), \(\hat{I}_{HR}\) and \(I_{HR}\) are compared to measure how image generation introduced noise to diffusion data. The same ROI is used to compare each of the images. SNR values are computed according to most signal attenuation direction approach [2].
Corpus Callosum (CC) is segmented automatically using fractional anisotropy (FA) values. SNR values in the CC region in x, y ad z-directions are compared for four different subjects in Table 1. It can be observed that \(\hat{I}_{HR}\) shows closer SNR values to \(I_{HR}\) than those of the \(I^{bc}_{LR}\).
As a second quantitative evaluation, FA histograms are calculated for each subject. Figure 3 depicts the histograms for two of the subjects. The histograms show that \(I_{HR}\) and \(\hat{I}_{HR}\) exhibit very similar distributions for FA values that are greater than 0.4. Other two subjects displayed similar distributions.
The generated FA maps and color FA maps are shown in Fig. 4 for one of the subjects. It can be observed that \(I_{HR}\) and \(\hat{I}_{HR}\) have similar FA and color FA maps while the baseline bi-cubic interpolation introduces attenuation and blur in the FA maps.
4.3 Tensor and ODF Analysis
For further evaluation of the quality of the reconstructed high resolution diffusion volumes, the diffusion tensor models are constructed for \( \hat{I}_{HR} \), \( I_{LR}^{bc} \) and \(I_{HR}\). Figure 5 shows the results for one of the subjects. Similar tensor orientations and strengths at the crossing points of CC and corticospinal tracts (CST) are observed for the \( \hat{I}_{HR} \) and \( I_{HR} \) of test subjects.
The orientation distribution functions (ODFs) are generated using constrained spherical deconvolution (CSD) [15] over the high resolution diffusion volumes. In Fig. 6, reconstructed ODFs are shown for one subject. It was observed that \(I^{bc}_{LR}\) has bigger artifactual side lobes.
5 Conclusions
In this paper, for the first time, an end-to-end super-resolution method based on GANs is presented for dMRI data. This approach does not assume any model, does not simply interpolate existing data but learns a data-driven generative mapping. The experimental quantitative results such as distribution of FA values and SNR as well as qualitative results such as FA maps, color FA maps, reconstructed tensors and ODFs demonstrate that GANs produce promising results to create higher resolution data using low resolution dMRI input. Although our work shows a preliminary proof of concept with GANs to increase the spatial resolution of dMRI twofold, our future work investigates further tuning of networks with larger training sets, increasing the resolution to triple, quadruple, or higher scale factors, and extending our work to angular up-sampling.
References
Calamante, F., Tournier, J.D., Heidemann, R.M., Anwander, A., Jackson, G.D., Connelly, A.: Track density imaging (TDI): validation of super resolution property. Neuroimage 56(3), 1259–1266 (2011)
Descoteaux, M., Deriche, R., Le Bihan, D., Mangin, J.F., Poupon, C.: Multiple q-shell diffusion propagator imaging. Med. Image Anal. Front. Neuroinformatics 15(4), 603–621 (2011)
Dong, C., Loy, C.C., He, K., Tang, X.: Learning a deep convolutional network for image super-resolution. In: Fleet, D., Pajdla, T., Schiele, B., Tuytelaars, T. (eds.) ECCV 2014. LNCS, vol. 8692, pp. 184–199. Springer, Cham (2014). https://doi.org/10.1007/978-3-319-10593-2_13
Garyfallidis, E.: Dipy, a library for the analysis of diffusion MRI data. Front. Neuroinformatics 8, 8 (2014)
Goodfellow, I., et al.: Generative adversarial nets. In: Ghahramani, Z., Welling, M., Cortes, C., Lawrence, N.D., Weinberger, K.Q. (eds.) Advances in Neural Information Processing Systems, vol. 27, pp. 2672–2680. Curran Associates, Inc. (2014)
Isola, P., Zhu, J.Y., Zhou, T., Efros, A.A.: Image-to-image translation with conditional adversarial networks. arxiv (2016)
Johansen-Berg, H., Behrens, T.E.: Diffusion MRI: From Quantitative Measurement to In Vivo Neuroanatomy. Academic Press (2013)
Johnson, J., Alahi, A., Fei-Fei, L.: Perceptual losses for real-time style transfer and super-resolution. In: Leibe, B., Matas, J., Sebe, N., Welling, M. (eds.) ECCV 2016. LNCS, vol. 9906, pp. 694–711. Springer, Cham (2016). https://doi.org/10.1007/978-3-319-46475-6_43
Kingma, D.P., Ba, J.: Adam: a method for stochastic optimization. CoRR abs/1412.6980 (2014). http://arxiv.org/abs/1412.6980
Le Bihan, D., Basser, P.J.: Molecular diffusion and nuclear magnetic resonance. In: Diffusion and Perfusion Magnetic Resonance Imaging, pp. 5–17 (1995)
Le Bihan, D., Iima, M.: Diffusion magnetic resonance imaging: what water tells us about biological tissues. PLoS Biol. 13(7), e1002203 (2015)
Ledig, C., et al.: Photo-realistic single image super-resolution using a generative adversarial network. CoRR abs/1609.04802 (2016)
Odena, A., Dumoulin, V., Olah, C.: Deconvolution and checkerboard artifacts. Distill (2016). https://doi.org/10.23915/distill.00003, http://distill.pub/2016/deconv-checkerboard
Peled, S., Yeshurun, Y.: Superresolution in MRI: application to human white matter fiber tract visualization by diffusion tensor imaging. Magn. Reson. Med. 45(1), 29–35 (2001)
Tournier, J.D., Calamante, F., Connelly, A.: Robust determination of the fibre orientation distribution in diffusion mri: non-negativity constrained super-resolved spherical deconvolution. Neuroimage 35(4), 1459–1472 (2007)
Tournier, J.D., Calamante, F., Gadian, D.G., Connelly, A.: Direct estimation of the fiber orientation density function from diffusion-weighted MRI data using spherical deconvolution. NeuroImage 23(3), 1176–1185 (2004)
Van Essen, D.C., et al.: The human connectome project: a data acquisition perspective. Neuroimage 62(4), 2222–2231 (2012)
Yang, Z., He, P., Zhou, J., Wu, X.: Non-local diffusion-weighted image super-resolution using collaborative joint information. Exp. Ther. Med. 15(1), 217–225 (2018)
Yap, P.-T., An, H., Chen, Y., Shen, D.: A generative model for resolution enhancement of diffusion MRI data. In: Mori, K., Sakuma, I., Sato, Y., Barillot, C., Navab, N. (eds.) MICCAI 2013. LNCS, vol. 8151, pp. 527–534. Springer, Heidelberg (2013). https://doi.org/10.1007/978-3-642-40760-4_66
Author information
Authors and Affiliations
Corresponding author
Editor information
Editors and Affiliations
Rights and permissions
Copyright information
© 2018 Springer Nature Switzerland AG
About this paper
Cite this paper
Albay, E., Demir, U., Unal, G. (2018). Diffusion MRI Spatial Super-Resolution Using Generative Adversarial Networks. In: Rekik, I., Unal, G., Adeli, E., Park, S. (eds) PRedictive Intelligence in MEdicine. PRIME 2018. Lecture Notes in Computer Science(), vol 11121. Springer, Cham. https://doi.org/10.1007/978-3-030-00320-3_19
Download citation
DOI: https://doi.org/10.1007/978-3-030-00320-3_19
Published:
Publisher Name: Springer, Cham
Print ISBN: 978-3-030-00319-7
Online ISBN: 978-3-030-00320-3
eBook Packages: Computer ScienceComputer Science (R0)