Download as pdf or txt
Download as pdf or txt
You are on page 1of 6

2019 IEEE 7th International Conference on Bioinformatics and Computational Biology

Dynamic MRI Reconstruction Exploiting Partial Separability and t-SVD

Shuli Ma, Huiqian Du*, Qiongzhi Wu, Wenbo Mei


Beijing Institute of Technology
Beijing, China
*Corresponding author
e-mail: duhuiqian@bit.edu.cn

Abstract—In this paper, we proposed a new method to been developed to better represent the inner structure of
reconstruct dynamic magnetic imaging (dMRI) data from tenor. The traditional tensor decomposition methods include
highly undersampled k-t space measurements. First, we use the CP decomposition (CPD) and Tucker decomposition (TD)
partial separability (PS) model to capture the spatiotemporal [10]. In recent years, tensor train decomposition (TTD) [11],
correlations of dMRI data. Then, we introduce a new tensor tensor ring decomposition (TRD) [12-13] and tensor singular
decomposition method named as tensor singular value value (t-SVD) decomposition method [14-17] are proposed.
decomposition (t-SVD) to the reconstruction problem. PS and The disadvantage of CPD is the rank one components cannot
low tensor multi-rank constrains are jointly enforced to be determined easily. Moreover, given a fixed rank,
reconstruct dynamic MRI data. We develop an efficient
computation of an approximation can be numerically
algorithm based on the alternating direction method of
unstable. TD, TTD and TRD can be seen as generalization of
multipliers (ADMM) to solve the proposed optimization
problem. The experimental results demonstrate the superior the CPD, and the truncated decomposition does not yield the
performance of the proposed method. best fit of the original tensor. T-SVD is a novel tensor
decomposition framework in which a new tensor product is
Keywords-dynamic magnetic imaging; partial separability; defined as well as t-SVD, tensor multi-rank. This
low rank tensor; tensor singular value decompositoin decomposition doesn’t unfold a tensor to a matrix so that the
structure of the tensor is maintained. More importantly, it
I. INTRODUCTION gives an optimal approximation of a tensor measured by the
Frobenious norm. Since dMRI data can be decomposed
Dynamic magnetic resonance imaging (dMRI) is a using t-SVD method and has low tensor multi- rank. We
widely used medical imaging modality but the inherently introduce t-SVD into dMRI reconstruction.
slow acquisition time and the limitation of spatial and In this paper, we proposed a new method to reconstruct
temporal resolution limited its application. In conventional dMRI data. First, we use partial separability (PS) to capture
imaging, fully sampled k-t space data s(k, t) are required to inherent spatiotemporal correlation in dMRI data. The PS of
obtain high spatial and temporal resolution. However, it is s(k, t) can be equivalently expressed as the low-rankness of a
difficult to sample k-t space at the Nyquist rate since the Casorati matrix from a set of s(k, t) samples. This method
number of the measurements grows exponentially with the has been demonstrated effective in measurements reduction.
physical dimension. Several model based schemes exploit Then, we introduce the t-SVD to reduce the required k-t
the prior information to reconstruct the dMRI sequences space measurements further. In order to enforce the low
from highly undersampled k-t data. Partial separability (PS) tensor multi-rank constrain, we minimize tensor nuclear
model assumed dMRI data is spatiotemporal correlated [1-2]. norm (TNN) which is the tightest convex relaxation of the
Earlier compressed sensing (CS) based methods assumed tensor multi-rank minimization. We develop an efficient
that the dMRI has sparse representation in various domains algorithm based on the alternating direction method of
[3-7], e.g. wavelet, gradient, Fourier transform domain and multipliers (ADMM) to solve the optimization problem.
etc. Recently, some studies have reported improved results Experimental results demonstrate that our method
obtained by enforcing the low rank or local structured low outperforms the methods which only enforcing PS or TNN
rank constrains [8-9]. These prior information or their constrain alone.
combinations all have been demonstrated useful for dMRI The remainder of this paper is organized as follows. In
reconstruction from the undersampled measurements. section II, we give the brief introduction to t-SVD. In section
It is known dMRI data are high dimensional tensors, III, we introduce our proposed reconstruction model and
while most low rank based reconstruction methods still optimization algorithm in detail. In section IV, we present
require classical two-way analysis. The tensor needs to be the experimental results to demonstrate the effectiveness of
unfolded to a 2D matrix and then enforcing low rank the proposed method. In section V, we give the conclusion.
constrains on the matrix. This unfolding breaks the data
structure and leads to loss of inherent information which II. BRIEF OVERVIEW OF T-SVD
presents in tensors. To address this problem, tensor In the scenario of tensor decomposition, a fiber is defined
decomposition based methods have been proposed to by fixing every index but one of a tensor, a slice is a matrix
preserve the original format of data. Various tensor defined by fixing all but two indices. For a three way tensor
decomposition models and corresponding tensor ranks have

978-1-7281-0641-0/19/$31.00 ©2019 IEEE 179


, we use the Matlab notation (k,:,:) , (:, k , :) and the ith frontal slice of , i.e. ri = rank (
(i )
) . The tensor
(:, :, k ) to denote the kth horizontal, lateral and frontal tubal rank, denoted as rankt ( ) , is defined as the number of
slices, and (:, i, j) , (i, :, j ) and (i, j,:) to denote the nonzero singular tubes of , where comes from the t-
(i, j)th mode-1, mode-2, mode-3 fiber. Specially, we use SVD of   * H . That is
(k ) to represent (:, :, k ) . rankt ( )  #i : (i, i, :)  0  max ri 
i
Definition 1 t-product. The t-product  * of
As in the matrix case, we use tensor nuclear norm (TNN)
 n1 n2 n3 and  n2 n4 n3 is a tensor of size n1  n4  n3
to measure the convex relaxations of the tensor multi-rank.
where the (i, j)th tube denoted by (i, j, :) for i  1, 2, , n1 Definition 4 The tensor-nuclear-norm (TNN). The
and j  1, 2, , n4 of tensor is given by tensor-nuclear-norm(TNN) denoted by is defined as
TNN
k 1
n2
(i, k , :) (k , j, :) , among which denotes the the sum of the singular values of all frontal slices of , is a
circular convolution between two vectors. norm and is the tightest convex relaxation to l1 norm of the
The t-product is analogous to the matrix multiplication tensor multi-rank.
except that circular convolution replaces the multiplication min  min blkdiag ( ) 
operation between the elements, which are now tubes. TNN *
Definition 2 Block diagonal matrix of the tensor. Let The TNN norm is a relaxation for tensor multi-rank
denote the block-diagonal matrix of the tensor in the minimization, there is no clear convex relaxation for tensor-
Fourier transform domain, i.e. tubal-rank minimization.
 (1) 
  III. THE PROPOSED METHOD
 (2) 
 We represent the dMRI data using PS model first and
 blockdiag ( )   

n1n3 n2 n3
then introduce t-SVD in reconstruction. The reconstruction
  problem is formulated as an optimization problem and an
 ( n3 ) 
  ADMM based algorithm is designed to solve the
optimization problem.
where ^ denotes the 1D Fourier transform along the third
dimension. A. The Proposed Reconstruction Model
Theorem 1 For  n1 n2 n3 , the t-SVD of is given
by
   T 
where and are orthogonal tensors of size n1  n1  n3
and n2  n2  n3 respectively. is a rectangular f-diagonal
tensor of size n1  n2  n3 and * denote t-product.
Fig 1 depicts the t-SVD of a three way tensor. The
process can be fulfilled in the Fourier transform domain.
Figure 2. Data acquisition scheme in PS. Red dots represent the training
data i.e. navigator signals, while black dots represent the imaging data. ky
is the phase encoding locations, t is the time locations of frame.

Let  n1 n2 n3 denote a dMRI data. PS model is used


to capture the strong spatiotemporal correction of Casorati
matrix C(s) shown below, where C( ) is an operator
Figure 1. The t-SVD of a tensor of size n1×n2×n3.
mapping s(k, t) data to a matrix, and k stands for (kx, ky).
Theorem 2 Let the t-SVD of  n1 n2 n3 be given by  s(k1,t1) s(k1,t M ) 
   T and for k  min (n1, n2 ) , we define the C ( s)  
  N M 

 s(k N ,t1 ) s(k N ,t M ) 
 i 1 (:, i, :)  (i, i, :)  (:, i, :)T , then
k

k
PS model assumes that Casorati matrix is rank deficient
k  arg min   and its rank R is much smaller than N or M. It can be
M F
decomposed as
where M =  =  |  n1 k n3
,  k n2 n3
.  C (s)  UV  
Definition 3 Tensor multi-rank and tubal rank. The where U  N  R represents a basis for the spatial subspace of
tensor multi-rank is r  n3 1 with its ith entry as the rank of C ( s) , and V  R M contains a basis for the temporal

180
subspace of C(s) . denotes the 2D Fourier transform.   2
 
n  arg min + n   n 1 F
TNN
2
Since s(k, t) is highly undersampled, many entries in the
matrix C(s) are unknown. In order to recover these unknown
 n  n  n   
entries, we need to determine the matrix V and U. Generally, n1

the matrix V is estimated prior to full image reconstruction. It The details for solving problem (3) are as follows.
can be determined from navigator signals (i.e. V is the right The initial value of at the nth iteration step is n0
R dominant singular vectors of the navigator signals). Fig. 2
shows the sampling pattern used for PS model based which can be determined using the Least Square fitting:
methods [2]. The set of red dots (fully sampled navigator   F 1  )        
1
  F ( n 1  n 1

signals along time t) near the center of k-space are used to n0
 
determine the temporal subspace V . The set of black dots
Then, set U n  C ( )V H , An  FU n V , we can get
(randomly undersampled) in the other regions are used to 0 n0 0 0

obtain U.  C ( )  C ()
After V is obtained, the matrix U can be estimated using    U n  F 1( A nV H )  
C ( n 0 )  C ()
n
the Least Square fitting [3], [18]. This method has been
demonstrated useful for dMRI reconstruction from highly where superscript H stands for the Hermitian Transpose,
undersampled k-t space data. In order to exploit the internal and the capital letter A is the Casorati matrix of the Euler
structure of dMRI data we propose to get the matrix U script letter i.e. A  C( ) .
through another way. We introduce t-SVD to the
reconstruction which won’t break the tensor structure of Lastly, n can be obtained from Un and V .
dMRI data. The matrix U is estimated by enforcing the low The details for solving problem (4) are as follows.
tensor multi-rank constrain. We formulate this problem as an According to the particular format (4), we can break it up
(i )
optimization problem and a tensor nuclear norm (TNN) [14- into n3 independent minimization problems. Let denote
16] is added as a regularization term. The reconstruction (i ) (i )
model we proposed is the ith frontal slice of . Similarly define , , then
1
(4) can be separated as:
2
min  F 
 U 2 F TNN    (i ) (i )  (i ) (i )
2
 
n  arg min n + n   n 1
s.t. UV  C ( ) 2
(i)
n * F

where  n n n denotes the undersampled k-t space


1 2 3
for i  1, 2, , n3 . This means each ith frontal slice of n
measurements,  is the undersampling mask, F is the 2D can be calculated through (8). The solution to (8) can be
Fourier transform operator, and  is a regularization found in [19-20].
parameter. Selection of rank R often needs to balance the The whole algorithm we developed for solving problem
representation capability of the model and the numerical (2) is summarized in Table I.
condition of the fitting problem. Since the matrix V is
already estimated from navigator signals, once the matrix U TABLE I. ALGORITHM FOR SOLVING EQ. (2)
is found, the X can be obtained directly. Input: estimated rank R, parameter  ,  , maximum number of
B. Optimization Algorithm iteration nmax , convergence condition tol .
In order to solve the optimization problem in (2), we Calculate V by SVD: Na  UV , where U  L R
,  R R
,
develop an alternating direction method of multipliers R M
, Na is the navigator data matrix. ( L is the number of
V
(ADMM) algorithm. First, we introduce auxiliary variable lines for sampling navigator data at the center of s(k,t) ).
 , then the problem (2) can be recast into a constrained Initialize 0 , 0 .
problem as follows: For n  1, 2,  , until ( n  nmax ) or ( n  tol )
1 2 a) Solve Eq. (6) for
min F  +  n0
, , 2 F TNN

s.t. UV  C ( ),  b) Solve Eq. (7) for U n


c) Update n
More specifically, the associated Lagrangian function of (i )
ADMM is given by b) Solve Eq. (8) for n , and i  1, 2, , n3
 d) Update 
(k )
1 2 2

by Eq. (5)
( , , ) F  +   
2 F TNN
2 F

e) Compute  n  n n 1

Then, each subproblem is simply obtained as follows n 1

End
1 2  2
Output: the recovered image *
.
 U n  argmin F  +  n 1  n 1 F  
U 2 F
2
s.t. U nV  C ( n )

181
IV. EXPERIMENTAL RESULTS TABLE III. PSNR OF THREE METHODS

To estimate the performance of our method, we apply the PSNR with 30% SR (sampling rate)
proposed method to dMRI reconstruction. Simulations are Methods
conducted on undersampled PINCAT phantom data and PINCAT phantom vivo cardiac perfusion MR
cardiac dMRI data. We compare our method with PS model Zero-Filled 29.6368 db 32.9020 db
based method in [3], [18] and TNN based method. All
simulations are carried out on Windows 10 and Matlab TNN 31.9510 db 37.2004 db
R2016a running on a PC with an Intel Core i5 CPU 3.2GHz PS 33.4435 db 34.8554 db
and 16GB of memory. The reconstruction quality is
quantified using the relative l2-norm error (RLNE) and peak- Proposed 36.5985 db 39.3312 db
signal-to-noise ratio (PSNR). RLNE defined as following:
*

 RLNE  F 
F

RLNE is a standard image quality metric indicating the


*
difference between the reconstruction and the gold
standard image .
The original image and sampling pattern used in the
experiments are shown in Fig.3. The navigation data come
from the 10 lines of low frequency region. Other k-t
measurements are obtained from randomly sampled data.
The whole sampling rate shown in Fig.3 is 35%. Tab. Ⅱ
shows the numerical comparison of the relative l2-norm error
(RLNE). Tab. Ⅲ shows the numerical comparison of the
peak-signal-to-noise ratio (PSNR). Furthermore, the RLNE
curves of every frame of vivo cardiac images are depicted in Figure 4. RLNE curves of each frame for reconstructing vivo cardiac
Fig. 4. We can see that our method provides smaller RLNE perfusion MR under the three methods. The sampling mask with sampling
comparing with PS and TNN method for both the PINCAT rate 35% (30 random lines and 10 lines for navigation data).
phantom data and vivo cardiac perfusion MR data.

(a) PS (b) TNN (c) Proposed


(a) (b) (c)

(d) PS error (e) TNN error (f) Proposed error


(d) (e) (f)
Figure 5. The 25th spatial frame: (a) PS; (b) TNN; (c) Proposed; (d)-(f)
Figure 3. The gold standard MR images and sampling mask used in the Error images corresponding to the reconstructions shown in (a)-(c). ( Image
experiment: (a) A spatial frame of PINCAT phantom; (b) A spatial frame (a)-(c) has been normalized such that the grayscale ranges from 0 to 1).
of vivo cardiac perfusion MR; (c) A sampling mask for one frame; (d) The
time profile of PINCAT phantom. (e) The time profile of vivo cardiac In Fig. 5-8, we demonstrate the reconstruction results of
perfusion MR. (f) The sampling mask shown along time t .
three methods from the undersampled measurements with
35% sampling rate.
TABLE II. RLNE OF THREE METHODS
RLNE with different SR (sampling rate)
PINCAT phantom vivo cardiac perfusion MR
Methods 20% 40% 20% 40%
Zero-Filled 0.1488 0.1228 0.2896 0.2413
TNN 0.1200 0.0891 0.1929 0.1350
PS 0.1276 0.0820 0.2839 0.2500
Proposed 0.0822 0.0429 0.1467 0.1155 (a) PS (b) TNN (c) Proposed

182
dMRI data simultaneously. Since PS can capture the
spatiotemporal correlation efficiently and t-SVD involved no
folding or flattening process, the structure of dMRI was
maintained effectively. We formulated the reconstruction
into optimization problem and developed ADMM algorithm
to fulfill reconstruction. The reconstruction results of
(d) PS error (e) TNN error (f) Proposed error PINCAT phantom and vivo cardiac perfusion data
Figure 6. The 50th time profile: (a) PS; (b) TNN; (c) Proposed; (d)-(f)
demonstrated that our proposed method has better
Error images corresponding to the reconstructions shown in (a)-(c). ( Image performance than PS constrain or TNN constrain only. The
(a)-(c) has been normalized such that the grayscale ranges from 0 to 1). proposed method can also be combined with the sparsity
constrain which will be discussed in the future work.
REFERENCES
[1] Z.-P. Liang, “Spatiotemporal imaging with partially separable
functions,” in Proc. IEEE Int. Symp. Biomed. Imaging, 2007, pp.
988–991.
[2] B Zhao, J. P. Haldar, A G Christodoulou, and Z.-P. Liang, “ Image
reconstruction from highly undersampled (k, t)-space data with joint
(a) PS (b) TNN (c) Proposed partial separability and sparsity constraints,” IEEE Transactions on
Medical Imaging, vol. 31, 2012, pp.1809-20.
[3] M Lustig, J M Santos, D L Donoho, and JM Pauly, “k-t SPARSE:
High frame rate dynamic MRI exploiting spatio-temporal sparsity,” in
Proc. ISMRM, 2006, pp. 2420.
[4] P E Z Larson, S Hu, M Lustig, A BKerr, S J Nelson, J Kurhanewicz,
J M Pauly, and D B Vigneron, “Fast dynamic 3D MR spectroscopic
imaging with compressed sensing and multiband excitation pulses for
(d) PS error (e) TNN error (f) Proposed error hyperpolarized 13C studies,” Magnetic resonance in medicine, vol.
65, 2011, pp. 610-619.
Figure 7. The 25th spatial frame: (a) PS; (b) TNN; (c) Proposed; (d)-(f) [5] Y Yu, J Jin, F Liu, and S Crozier, “Multidimensional compressed
Error images corresponding to the reconstructions shown in (a)-(c). ( Image sensing MRI using tensor decomposition-based sparsifying
(a)-(c) has been normalized such that the grayscale ranges from 0 to 1). transform,” PlosOne, vol, 9, 2014, pp. e98441.
[6] Y Wang, N Cao, Z Liu, and Y Zhang, “Real-time dynamic MRI using
parallel dictionary learning and dynamic total variation,”
Neurocomputing, vol. 238, 2017, pp. 410-419.
[7] Y Liu, S Wu, X Huang, B Chen, and C Zhu, “Hybrid CS-DMRI:
Periodic Time-Variant Subsampling and Omnidirectional Total
Variation Based Reconstruction,” IEEE transactions on medical
imaging, vol. 36, 2017, pp. 2148-2159.
[8] S G Lingala, Y Hu, E Dibella, and M Jacob, “Accelerated dynamic
(a) PS (b) TNN (c) Proposed MRI exploiting sparsity and low-rank structure: k-t SLR,” IEEE
Transactions on Medical Imaging, vol. 30, 2011, pp. 1042-1054.
[9] J He, Q Liu, A G Christodoulou, C Ma, F Lam, and Z.-P. Liang,
“Accelerated High-Dimensional MR Imaging With Sparse Sampling
Using Low-Rank Tensors,” IEEE Transactions on Medical Imaging,
vol. 35, 2016, pp. 2119-2129.
[10] T G Kolda and B WB ader, “Tensor Decompositions and
Applications,” Siam Review, vol. 51, 2009, pp. 455-500.
(d) PS error (e) TNN error (f) Proposed error [11] I Oseledets and E Tyrtyshnikov, “TT-cross approximation for
multidimen- sional arrays,” Linear Algebra & Its Applications, vol.
Figure 8. The 50th time profile: (a) PS; (b) TNN; (c) Proposed; (d)-(f)
432, 2010, pp.70- 88.
Error images corresponding to the reconstructions shown in (a)-(c). ( Image
(a)-(c) has been normalized such that the grayscale ranges from 0 to 1). [12] Q Zhao, G Zhou, S Xie, and LQ Zhang, “Tensor Ring
Decomposition,” 2016, unpublished.
Fig. 5(a-c) - 8(a-c) show the reconstructed spatial frames [13] W Wang, V Aggarwal, and S Aeron, “Efficient low rank tensor ring
and time profiles. Fig. 5(d-f) - 8(d-f) are corresponding error completion,” Rn, vol.1, 2017, pp. 1.
images. We can see that our method provides better [14] Z Zhang and S Aeron, “Exact Tensor Completion Using t-SVD,”
IEEE Transactions on Signal Processing, vol. 65, 2017, pp.1511-1526.
reconstruction results than PS or TNN based method, which
[15] Z Zhang, G Ely, S Aeron, N Hao, and M Kilmer, “Novel methods for
demonstrates that the combination of PS and TNN multilinear data completion and de-noising based on tensor SVD,” in
constraints can yield better recovery results than only Proc. IEEE Conf. Comput. Vis. Pattern Recognit., Jun. 2014, pp.
enforcing PS or TNN constrain alone. 3842–3849.
[16] O Semerci , N Hao, and M E Kilmer, “Tensor-Based Formulation and
V. CONCLUSION Nuclear Norm Regularization for Multienergy Computed
Tomography,” IEEE Transactions on Image Processing, vol. 23, 2014,
In this paper we proposed a dMRI reconstruction method pp. 1678-1693.
which exploited partial separability and low multi-rank of

183
[17] M E. Kilmer, K Braman, and N Hao, “Third-order Tensors As [19] JF Cai, E J Candes, and ZW Shen, “A Singular Value Thresholding
Operators On Matrices: A Theoretical And Computational Algorithm for Matrix Completion,” Siam Journal on Optimization,
Framework With Applications In Imaging,” Siam Journal on Matrix vol. 20, 2008, pp. 1956- 1982.
Analysis & Applications, vol. 34, 2013, pp. 148-172. [20] M Yin, S Xie, Z Wu, Y Zhang, and J Gaoet, "Subspace Clustering via
[18] B Zhao, J. P. Haldar and Z.-P. Liang, “PSF model-based Learning an Adaptive Low-rank Graph," IEEE Transactions on Image
reconstruction with sparsity constraint: algorithm and application to Processing, 2018, pp. 1.
real-time cardiac MRI,” in Proc IEEE Eng Med Biol Soc, 2010.

184

You might also like