Cihat Eldeniz1, Weijie Gan1, Sihao Chen1, Jiaming Liu1, Ulugbek S. Kamilov1, and Hongyu An1
1Washington University in St. Louis, Saint Louis, MO, United States
Synopsis
Radial MRI can be used for
reconstructing multiple respiratory phases with retrospective binning. However,
short acquisitions suffer from significant streaking artifacts. Compressed
sensing (CS)-based methods are commonly used; nevertheless, CS is computational
intensive and the image quality depends on the regularization parameters. We
hereby propose a deep learning method that does not need an artifact-free
target during training. The method can reconstruct high-quality volumes with
ten respiratory phases, even for acquisitions close to 1 minute in length. The method
outperforms CS for the same acquisition duration and can yield slightly better
results than Unet3D trained using a surrogate ground truth.
INTRODUCTION
Respiratory motion compromises image
quality in thoracic and abdominal magnetic resonance imaging (MRI). Recently,
self-navigation techniques have been developed to detect respiratory motion from
the data itself[1–4]. Binning the data into respiratory
phases results in sets of undersampled k-space data, leading to poor
signal-to-noise ratio (SNR) and streaking artifacts. To overcome these
challenges, compressed sensing (CS)[5] reconstruction has been employed[2–4,6–8]. For CS methods, the selection of the
regularization parameters is often empirical and can lead to sub-optimal results.
Moreover, the iterative optimization is computationally intensive.
More recently, deep
learning (DL) methods [9,10] have been explored in MR image
reconstruction [11–14]. For the usual DL, an artifact-free
ground truth reference is required for training. However, such a reference can
be difficult to obtain in practice. Recently, a new learning technique called
Noise2Noise was introduced [15] that can be trained without a ground
truth. This technique instead uses pairs of noisy images.
Respiratory binning leads
to a different radial k-space coverage pattern for each respiratory phase,
leading to different artifact patterns. On the other hand, the underlying
patient anatomy remains similar across adjacent respiratory phases. Based on this observation, and inspired by
the Noise2Noise approach, we developed a novel deep learning method that learns
artifact-free MR volumes directly from noisy MR data with streaking artifacts,
without the need of a ground truth. In this study, the odd respiratory phases were
used as the input and the even phases were used as the training targets, and
vice versa. We refer to our new technique as Phase2Phase (P2P) and applied it for
acquisition times ranging from 1 to 5 minutes. METHODS
The
data was acquired using Consistently Acquired Projections for Tuned and Robust
Estimation (CAPTURE)[4], which is based on a T1-weighted
stack-of-stars 3D spoiled gradient-echo sequence with fat suppression[1,16], and comes with consistently acquired
projections for a more robust detection of respiratory motion.
All
experiments were performed on a 3T simultaneous PET/MRI scanner (Siemens
Biograph mMR; Siemens Healthcare, Erlangen, Germany). This Health Insurance
Portability and Accountability Act (HIPAA)-compliant study was performed after
the approval of our Institutional Review Board. 15 healthy volunteers and 17
cancer patients were recruited.
The default
parameters for the CAPTURE acquisition were as follows: TE/TR = 1.69 ms/3.54
ms, matrix size = 320 x 320, FOV = 360 mm x 360 mm, slab thickness = 288 mm,
number of partitions = 48, partial Fourier factor = 6/8 (giving a temporal
resolution of 153.52 ms for the navigator), reconstructed slices per slab = 96
(yielding a slice thickness of 3 mm). The resulting voxel size was 1.125 x
1.125 x 3 mm3. The number of azimuthal angles was 2000, resulting in
a total acquisition time of 5 minutes and 7 seconds, which was longer for large
subjects with a larger number of slices.
After
the k-space data were binned into ten respiratory phases, four image
reconstruction methods were used: (1) multi-coil non-uniform inverse fast
Fourier transform (MCNUFFT), (2) Compressed Sensing (CS), (3) UNet3DPhase – a
3D U-net with the third dimension being the respiratory phase and the target
being the 5-minute CS reconstruction, and (4) Phase2Phase (P2P), which was
inspired by Noise2Noise and performs odd-to-even- as well as even-to-odd-phase
learning without an artifact-free ground truth. Various numbers of radial
spokes (400, 800, 1200, 1600), corresponding, respectively, to about 1-, 2-, 3-
and 4-minute acquisitions, were used to reconstruct the images. Figure 1
details the training and testing of the deep learning methods. 8 healthy
subjects were used for training, 1 was used for validation and 6 were used for
testing together with 17 patients. RESULTS
Figures 2 and 3 demonstrate
reconstructions for 400 radial spokes (approximately 1 minute) for two
different patients. The 2000-spoke CS reconstruction serves as the
gold-standard reference. The 400-spoke CS could not remove all artifacts, while
both Unet3DPhase and P2P images show much fewer artifacts. Moreover, P2P
provided sharper reconstructions when compared to Unet3DPhase. Figures 4 and 5 show Unet3DPhase and P2P reconstructions
for 400, 800, 1200 and 1600 radial spokes. There is a noticeable improvement in
image quality for 800 spokes when compared to 400 spokes for both methods. However,
the images qualities are similar for 800 spokes and beyond.
It is worth noting that,
although the training of the networks takes dozens of hours, it takes only 10
seconds to reconstruct the entire 4D data of size 320x320x96x10.DISCUSSION
P2P can learn to reconstruct
high-quality images without an artifact-free ground truth. Even the 1-minute
acquisition images are of good quality. Furthermore,
P2P can provide sharper images than UNet3DPhase. One possible limitation of P2P
is that the assumption of little motion across adjacent phases may lead to
image blurring. Finally, both Unet3DPhase and P2P outperformed the 400-spoke
CS.CONCLUSION
In summary, we proposed a
new deep learning method (P2P) that can learn to reconstruct artifact-free
images without an artifact-free ground truth. Once trained, the P2P reconstructions
are much faster than CS, making the method extremely suitable for clinical
routine. Furthermore, no ground truth is needed, which saves additional time
during training. Finally, the network trained on healthy subjects can work on
patients with quite different lesion patterns as demonstrated here.Acknowledgements
No acknowledgement found.References
1. Grimm R, Fürst S, Souvatzoglou M, et
al. Self-gated MRI motion modeling for respiratory motion compensation in
integrated PET/MRI. Medical Image Analysis. 2015;19(1):110–120.
2. Feng
L, Grimm R, Block KT, et al. Golden-angle radial sparse parallel MRI: combination
of compressed sensing, parallel imaging, and golden-angle radial sampling for
fast and flexible dynamic volumetric MRI. Magn Reson Med.
2014;72(3):707–717.
3. Feng
L, Axel L, Chandarana H, et al. XD-GRASP: Golden-angle radial MRI with
reconstruction of extra motion-state dimensions using compressed sensing. Magn.
Reson. Med. 2016;75(2):775–788.
4.
Eldeniz C, Fraum T, Salter A, et al. CAPTURE:
Consistently Acquired Projections for Tuned and Robust EstimationA
Self-Navigated Respiratory Motion Correction Approach. Investigative
Radiology. 2018;53(5):293–305.
5. Donoho
DL. Compressed sensing. IEEE Transactions on Information Theory.
2006;52(4):1289–1306.
6. Lustig
M, Donoho D, Pauly JM. Sparse MRI: The application of compressed sensing for
rapid MR imaging. Magn. Reson. Med. 2007;58(6):1182–1195.
7.
Chandarana H, Feng L, Ream J, et al. Respiratory Motion-Resolved Compressed
Sensing Reconstruction of Free-Breathing Radial Acquisition for Dynamic Liver
Magnetic Resonance Imaging. Invest Radiol. 2015;50(11):749–756.
8. Huang
J, Wang L, Zhu Y. Compressed Sensing MRI Reconstruction with Multiple Sparsity
Constraints on Radial Sampling. Mathematical Problems in Engineering.
2019;2019:14.
9.
Schmidhuber J. Deep Learning in Neural Networks: An Overview. Neural
Networks. 2015;61:85–117.
10.
Ronneberger O, Fischer P, Brox T. U-Net: Convolutional Networks for Biomedical
Image Segmentation. arXiv:1505.04597 [cs]. 2015. Available at:
http://arxiv.org/abs/1505.04597. Accessed July 18, 2019.
11. Wang
S, Su Z, Ying L, et al. Accelerating magnetic resonance imaging via deep
learning. In: 2016 IEEE 13th International Symposium on Biomedical Imaging
(ISBI).; 2016:514–517.
12. Han
Y, Yoo J, Kim HH, et al. Deep learning with domain adaptation for accelerated
projection-reconstruction MR. Magnetic Resonance in Medicine.
2018;80(3):1189–1205.
13. Lee
D, Yoo J, Tak S, et al. Deep Residual Learning for Accelerated MRI Using
Magnitude and Phase Networks. IEEE Transactions on Biomedical Engineering.
2018;65(9):1985–1995.
14. Lundervold
AS, Lundervold A. An overview of deep learning in medical imaging focusing on
MRI. Zeitschrift für Medizinische Physik. 2019;29(2):102–127.
15.
Lehtinen J, Munkberg J, Hasselgren J, et al. Noise2Noise: Learning Image
Restoration without Clean Data. arXiv:1803.04189 [cs, stat]. 2018.
Available at: http://arxiv.org/abs/1803.04189. Accessed November 3, 2019.
16.
Chandarana H, Block TK, Rosenkrantz AB, et al. Free-breathing radial 3D
fat-suppressed T1-weighted gradient echo sequence: a viable alternative for
contrast-enhanced liver imaging in patients unable to suspend respiration. Invest
Radiol. 2011;46(10):648–653.