Joseph Y. Cheng1, John M. Pauly2, and Shreyas S. Vasanawala1
1Radiology, Stanford University, Stanford, CA, United States, 2Electrical Engineering, Stanford University, Stanford, CA, United States
Synopsis
Model-based accelerated imaging techniques enable high scan
time reductions while maintaining high image quality. These techniques rely on
the ability to accurately estimate the imaging model. This model can be
extended to include information beyond physical limits, such as high-resolution
phase information to promote conjugate symmetry or information of voxels
without signal for a stronger image prior. Thus, we propose a deep learning
approach to estimate the imaging model with latent coil maps. Furthermore, we
jointly train this latent map estimator with a deep-learning-based reconstruction
using adversarial loss, and we demonstrate the effectiveness of this approach
in volumetric knee datasets.
Introduction
Model-based accelerated imaging techniques1–5 enable high scan time reduction (over 8 fold) while maintaining high image quality. At the core of these
methods is the ability to accurately estimate the imaging model. This model
characterizes the process of transforming the desired image to the measurement
domain in k-space. A key component of the transform is the multi-channel coil
array hardware. For optimal performance, coil profile maps must be
characterized for each scan; however, the characterization can be costly in
scan time and computation. Thus, we introduce a deep convolutional neural
network (CNN) to estimate latent coil profile maps to model the imaging
process. For further gains, this network is jointly trained with a
reconstruction neural network with an adversarial loss.Method
Current approaches to estimate the multi-channel model
assume that each channel profile map of a coil array is slowly varying in space1–4. With accurate estimation (a)
of voxels without signal and (b) of high spatial frequency components induced
by field perturbations, additional information can be modeled by the profile
map for improving the reconstruction. Thus, we propose to use unsupervised
learning to avoid biasing our training based on previous methods: the
CNN6 is trained based on how the
output will be applied instead of on an ideal estimate of profile maps. More
specifically, if the conjugate transpose of the maps estimated is applied to
the original multi-channel, the number of channels is reduced (to 1 in this
work). By applying the maps again, the result is the original multi-channel
data. Because no constraint is imposed on the channel-reduced space and
reconstruction will be performed in this space, we refer to the images in this
space as latent coils and the profile maps as latent coil
maps.
As the main goal is to improve the image reconstruction, the
CNN for estimating these latent coil maps (Latent Map Estimator) are jointly trained
with a reconstruction CNN based on compressed sensing7–9 (Reconstruction Network). Overview of the method is depicted in Figure 1. Furthermore,
to improve the perceptual image quality of the reconstruction, adversarial loss
(Discriminator, $$$D_\omega$$$) is used10.
The training consists of three
components to the loss function.
Latent map estimator $$$G_\theta$$$ (Figure 2b) is
trained with loss:
$$\mathcal{L}_L^\theta=\sum_i\left\|G_\theta(u_i)G^H_\theta(u_i)c_i-c_i
\right\|_1,$$ where $$$c_i$$$ is the $$$i$$$-th fully-sampled multi-channel training
example, and $$$u_i$$$ is k-space subsampled according to the subsampling mask
used for future scans. Both $$$u_i$$$ and $$$c_i$$$ are in the image domain. $$$G_\theta$$$
is jointly trained with the reconstruction network $$$R_\phi$$$ (Figure 2c)
with an additional loss:
$$\mathcal{L}_R^{\theta,\phi}=\sum_i \left\|G_\theta(u_i)R_\phi(u_i,G_\theta(u_i))-c_i
\right\|_1-\lambda\log(D_\omega(G_\theta(u_i)R_\phi(u_i,G_\theta(u_i)))),$$
where $$$R_\phi(u_i,G_\theta(u_i))$$$ outputs a latent image which is then
transformed into the multi-channel image domain with $$$G_\theta(u_i)$$$.
Discriminator $$$D_\omega$$$ (Figure 2d) is trained
with loss $$$\mathcal{L}_D^{\omega}$$$: $$\mathcal{L}_D^{\omega}=\sum_i-\log(D_\omega(c_i))-\log(1-D_\omega(R_\phi(u_i,G_\theta(u_i)))),$$
where $$$D_\omega$$$ classifies each image patch as the original fully-sampled $$$c_i$$$
or not.
Proton-density-weighted volumetric knee scans11,12 using an 8-channel knee coil were
used. Cartesian k-space data were first transformed into the hybrid $$$(x,ky,kz)$$$-space
and separated into $$$x$$$-slices. Dataset consisted of 14, 2, and 3 knee
subjects (4480, 640, and 960 slices) for training, validation, and testing. For
training and for final testing, variable-density poisson-disc sampling masks
were used. The networks were jointly trained in TensorFlow13; comparisons were performed
using BART14.
Results
From a subsampled image set, latent coil maps were generated
with high spatial frequencies (Figure 3). These latent coil maps were similar
to the maps computed from ESPIRiT with and without automatic cropping. These
latent coil maps were used to assist in the reconstruction of the subsampled
dataset by transforming the multi-channel image into a latent space that is
specific for de-noising (Figure 4). Reconstruction results for different
approaches are shown in Figure 5. On a NVIDIA 1080 Ti card, the latent coil map
CNN took 40ms and reconstruction CNN took 60ms, whereas ESPIRiT took 600ms and
9.6s without and with cropping. On the same GPU, parallel imaging &
compressed sensing (PICS) with 30 iterations took 2.6s. The reconstruction
network trained without adversarial loss appears smoother agreeing with
previous literature10,15 whereas the network trained
with adversarial loss has more high-resolution texture.Discussion & Conclusion
A single latent coil is used in this work, but the network
can be easily extended to support more coils. The latent space enabled by the
unsupervised training provides an unexplored domain for reconstruction
solutions. Further, the proposed highly-flexible method allows for arbitrary
regularizations to guide the unsupervised training and for training with
arbitrary reconstruction networks. In conclusion, a CNN framework is proposed
and trained with unsupervised learning to estimate latent coil maps. Latent coil maps are rapidly estimated and used for reconstructing
multi-channel images.Acknowledgements
NIH R01-EB009690, NIH R01-EB019241, NIH R01-EB026136, and GE
Healthcare.References
- Griswold, M. A. et al. Generalized autocalibrating partially parallel acquisitions (GRAPPA). Magn. Reson. Med. 47, 1202–1210 (2002).
- Pruessmann, K. P., Weiger, M., Scheidegger, M. B. & Boesiger, P. SENSE: sensitivity encoding for fast MRI. Magn. Reson. Med. 42, 952–62 (1999).
- Lustig, M. & Pauly, J. M. SPIRiT: Iterative self-consistent parallel imaging reconstruction from arbitrary k-space. Magn. Reson. Med. 64, 457–471 (2010).
- Uecker, M. et al. ESPIRiT-an eigenvalue approach to autocalibrating parallel MRI: Where SENSE meets GRAPPA. Magn. Reson. Med. 71, 990–1001 (2014).
- Lustig, M., Donoho, D. & Pauly, J. M. Sparse MRI: The application of compressed sensing for rapid MR imaging. Magn. Reson. Med. 58, 1182–1195 (2007).
- He, K., Zhang, X., Ren, S. & Sun, J. Identity Mappings in Deep Residual Networks. arXiv:1603.05027 [cs.CV] (2016).
- Hammernik, K. et al. Learning a variational network for reconstruction of accelerated MRI data. Magn. Reson. Med. 79, 3055–3071. PMCID: PMC5902683 (2018).
- Diamond, S., Sitzmann, V., Heide, F. & Wetzstein, G. Unrolled Optimization with Deep Priors. arXiv: 1705.08041 [cs.CV] (2017).
- Cheng, J. Y., Chen, F., Alley, M. T., Pauly, J. M. & Vasanawala, S. S. Highly Scalable Image Reconstruction using Deep Neural Networks with Bandpass Filtering. arXiv:1805.03300 [cs.CV] (2018).
- Mardani, M. et al. Deep Generative Adversarial Neural Networks for Compressive Sensing (GANCS) MRI. IEEE Trans. Med. Imaging (2018). doi:10.1109/TMI.2018.2858752
- Epperson, K. et al. Creation of Fully Sampled MR Data Repository for Compressed Sensing of the Knee. in SMRT 22nd Annual Meeting (2013). doi:10.1.1.402.206
- MRI Data. Available at: http://mridata.org/. (Accessed: 14th July 2017)
- Abadi, M. et al. TensorFlow: Large-Scale Machine Learning on Heterogeneous Distributed Systems. arXiv:1603.04467 [cs.DC] (2016).
- Uecker, M. et al. BART: version 0.3.01. (2016). doi:10.5281/zenodo.50726
- Pathak, D., Krahenbuhl, P., Donahue, J., Darrell, T. & Efros, A. A. Context Encoders: Feature Learning by Inpainting. in 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR) 2536–2544 (IEEE, 2016). doi:10.1109/CVPR.2016.278