Yuxin Hu1,2, Qiyuan Tian1,2, Yunyingying Xu2, Philip K. Lee1,2, Bruce L. Daniel1,3, and Brian A. Hargreaves1,2,3
1Department of Radiology, Stanford University, Stanford, CA, United States, 2Department of Electrical Engineering, Stanford University, Stanford, CA, United States, 3Department of Bioengineering, Stanford University, Stanford, CA, United States
Synopsis
Low rankness of image/k-space data has been exploited in many different
MRI applications. In this work, we introduce weighted-nuclear-norm minimization
for MRI low-rank reconstruction, in which smaller thresholds are used
for larger eigenvalues to reduce information loss. Our simulation results
demonstrate that weighted nuclear norm could serve as a better rank
approximation compared to (unweighted) nuclear norm. With this technique, we
achieve 10-shot DWI and high-fidelity half-millimetre DWI
reconstruction with significantly reduced ghosting artifacts.
Introduction
Low-rank
regularization
has
achieved success in a variety of MRI reconstruction tasks [1-7]. Figure 1 shows an example
eigendecomposition of an 8-channel brain image along the channel dimension.
Most signal lies in the space corresponding to the largest/first eigenvector, suggesting
low rank along the channel dimension. While rank minimization is NP-hard, its
convex envelope, the nuclear norm, is usually used to approximate the rank to
reduce computational complexity [8].
During the
nuclear norm minimization process, eigenvalues of the low-rank matrix are
soft-thresholded, usually using identical thresholds determined by the
regularization parameter. Components corresponding to larger eigenvalues are
more likely to represent the signal we want to preserve. A weighted-nuclear-norm minimization (WNNM)
assigns different weightings and thus different thresholds, to different
eigenvalues to reduce regularization of larger eigenvalues [9]. In this work,
we apply WNNM to multi-shot diffusion-weighted imaging (DWI) reconstruction.Theory
Weighted nuclear norm
The
weighted nuclear norm, or weighted summation of the eigenvalues of the matrix, can
approximate the rank more accurately than the nuclear norm [10], and may therefore
maintain the signal fidelity better while approaching low-rank approximation. To
validate this, we use two different thresholding methods on the eigenvalues of
matrices constructed from the 8-channel image as in [4] and calculate the normalized root
mean squared error (NRMSE) between the compressed image and the original image.
In the first method, locally low-rank (LLR), thresholds for all eigenvalues are
the same, while in the second, weighted locally low-rank (wLLR), the threshold for
the first eigenvalue is reduced 10-fold. Figure 2 demonstrates that, with the
same nuclear norm after thresholding, wLLR better preserves the signal and
achieves a lower NRMSE. Figure 3 shows compressed images (with about 73%
nuclear norm left) using two methods, and LLR shows greater errors than wLLR. From
the difference image, we can see that Fig. 3b (LLR) loses some structural
details, while Fig. 3c (wLLR) still maintains high quality.
WNNM for multi-shot DWI reconstruction
Compared to single-shot DWI, multi-shot DWI enables higher
resolution with reduced distortion artifacts, but suffers motion-induced
shot-to-shot phase variations [11]. In shot-LLR [7], spatial-shot matrices are
constructed and matrix completion is used to reconstruct the images,
resolving phase variations. However, the number of shots is still limited. Similar
to shot-LLR [4], we use a relaxed model and low-rank regularization term for
multi-shot DWI reconstruction, but use a weighted nuclear norm term to
approximate the rank regularization term as below,
$$$\min_{x} \sum_{s = 1 }^{NS}\frac{1}{2}\left \| E_{s}FS x_{s} - y_{s} \right \|_{2}^{2} + \lambda\sum_{l\in \Omega}\left \| R_{l}x \right \|_{w,*}$$$
where
NS is the number of shots, E, F, and S are the sampling, Fourier transform, and
sensitivity encoding operators, respectively, xs is the image of the s-th
shot, and ys is the acquired k-space data of the s-th shot, |.|w, *
represents the weighted nuclear norm operator, w represents the weightings for
different eigenvalues, and $$$\lambda$$$ is the regularization parameter of the LLR
term. Adaptive weightings based on eigenvalues and the regularization parameter
are used to avoid the complexity of setting them [12, 13].
Methods
Data acquisition
Under
IRB approval, multi-shot EPI DWI data was acquired on six volunteers. Five
volunteers’ brains were scanned on a 3T GE MR750 scanner using a 32-channel
head receive-only coil with 4,6,8 and 10 shots, FOV (field-of-view)=20 cm, in-plane
resolution=0.8 mm isotropic, slice thickness=4 mm,
b-value=1000 s/mm2, and NEX (number of repetitions)=4.
To
evaluate the effect of different in-plane resolutions, one brain volunteer was
scanned on a 3T GE Signa Premier scanner using a 48-channel head
receive-only coil with 6 shots, FOV (field-of-view)=20 cm, in-plane resolution=0.7,0.6
and 0.5 mm isotropic, slice thickness=4 mm, b-value=1000 s/mm2, and NEX=1.
Data preprocessing and reconstruction
The acquired data were processed using the product
algorithm for Nyquist artifact correction and ramp sampling correction.
Sensitivity map was calculated using BART [14] based on the non-diffusion-weighted
data. All data were reconstructed by shot-LLR [7] and the proposed method, both
using FISTA with 200 iterations, a regularization parameter of 0.008, and 8x8
patches for spatial-shot locally low-rank matrices [4]. The proximal operator
of the weighted nuclear norm was to threshold the eigenvalues based on the
corresponding weightings [9].
Results and discussion
Figure
4 shows reconstructions for 4-10 shots.
At 4 shots, shot-LLR and wLLR achieve similar results. As the number of
shots increases, shot-LLR has signal loss (Fig. 4j), and increasing residual
ghosting artifacts (Fig. 4c,d), compared with wLLR (Fig. 4e-h).
Figure
5 shows that shot-LLR leads to signal loss especially in the center of the
brain when compared with wLLR. The signal loss gets more prominent at higher
in-plane resolutions. This suggests that wLLR retains useful information that would
be lost if the same threshold was applied to different eigenvalues.
Different approaches to selecting and
adapting the weights could be explored as WNNM is applied to a wide range of
MRI low-rank reconstruction applications, such as spatio-temporal
imaging [1-3], calibrationless parallel imaging [4,5] and
multispectral imaging [6]. Conclusion
We
demonstrated weighted-nuclear-norm minimization for MRI low-rank reconstruction.
We showed that by using different thresholds for different eigenvalues, we
could achieve 10-shot DWI and high-fidelity half-millimetre brain DWI
reconstruction with reduced ghosting artifacts and information loss compared to
unweighted nuclear-norm methods.
Acknowledgements
R01-EB009055,
P41-EB015891 and GE Healthcare.References
[1] Haldar JP,
Liang ZP. Spatiotemporal imaging with partially separable functions: a matrix
recovery approach. In: Proceedings of the 7th IEEE International Symposium on
Biomedical Imaging: From Nano to Macro; 2010; Rotterdam, Netherlands; 716–719.
[2] Otazo,
Ricardo, Emmanuel Candes, and Daniel K. Sodickson. "Low‐rank plus sparse matrix decomposition
for accelerated dynamic MRI with separation of background and dynamic
components." Magnetic Resonance in Medicine 73.3 (2015):
1125-1136.
[3] Zhang, Tao,
et al. "Fast pediatric 3D free‐breathing
abdominal dynamic contrast enhanced MRI with high spatiotemporal
resolution." Journal of Magnetic Resonance Imaging 41.2 (2015): 460-473.
[4] Trzasko JD,
Manduca A. Calibrationless parallel MRI using CLEAR. In: Proceedings of the
Asilomar Conference on Signals, Systems, and Computers; 2011; Pacific Grove,
CA; 75–79.
[5] Shin, Peter
J., et al. "Calibrationless parallel imaging reconstruction based on
structured low‐rank
matrix completion." Magnetic resonance in medicine 72.4 (2014):
959-970.
[6] Levine,
Evan, et al. "Accelerated three‐dimensional
multispectral MRI with robust principal component analysis for separation of on‐and off‐resonance signals." Magnetic
resonance in medicine 79.3 (2018): 1495-1505.
[7] Hu Y,
Levine EG, Tian Q, et al. Motion‐robust
reconstruction of multishot diffusion‐weighted
images without phase estimation through locally low‐rank regularization. Magn Reson Med.
2019;81:1181–1190.
[8] Candes EJ,
Recht B. Exact matrix completion via convex optimization. Found Comput Math.
2009;9:717–772.
[9] Gu,
Shuhang, et al. "Weighted nuclear norm minimization with application to
image denoising." Proceedings of the IEEE conference on computer
vision and pattern recognition. 2014.
[10] Lu, Canyi,
et al. "Nonconvex nonsmooth low rank minimization via iteratively
reweighted nuclear norm." IEEE Transactions on Image
Processing 25.2 (2015): 829-839.
[11] Wu W.
Miller KL. Image formation in diffusion MRI: A review of recent technical
developments. J Magn Reson Imaging. 2017;46:1–17.
[12] Candes,
Emmanuel J., Michael B. Wakin, and Stephen P. Boyd. "Enhancing sparsity by
reweighted ℓ 1 minimization." Journal of Fourier analysis and
applications 14.5-6 (2008): 877-905.
[13] Gu,
Shuhang, et al. "Weighted nuclear norm minimization and its applications
to low level vision." International journal of computer
vision 121.2 (2017): 183-208.
[14] Uecker M,
Tamir J, Ong F, Holme C, Lustig M. Bart: Version 0.4.01; 2017.