Professional Documents
Culture Documents
com
Abstract
This works addresses the problem of reconstructing multiple T1- or T2-weighted images of the same anatomical cross section from
partially sampled K-space data. Previous studies in reconstructing magnetic resonance (MR) images from partial samples of the K-space used
compressed sensing (CS) techniques to exploit the spatial correlation of the images (leading to sparsity in wavelet domain). Such techniques
can be employed to reconstruct the individual T1- or T2-weighted images. However, in the current context, the different images are not really
independent; they are images of the same cross section and, hence, are highly correlated. We exploit the correlation between the images,
along with the spatial correlation within the images to achieve better reconstruction results than exploiting spatial correlation only.
For individual MR images, CS-based techniques lead to a sparsity-promoting optimization problem in the wavelet domain. In this article,
we show that the same framework can be extended to incorporate correlation between images leading to group/row sparsity-promoting
optimization. Algorithms for solving such optimization problems have already been developed in the CS literature. We show that significant
improvement in reconstruction accuracy can be achieved by considering the correlation between different T1- and T2-weighted images. If the
reconstruction accuracy is considered to be constant, our proposed group sparse formulation can yield the same result with 33% less K-space
samples compared with simple sparsity-promoting reconstruction. Moreover, the reconstruction time by our proposed method is about two to
four times less than the previous method.
© 2011 Elsevier Inc. All rights reserved.
Keywords: T1- and T2-weighted MR imaging; quantitative MRI; optimization; fast imaging; compressed sensing; group-sparsity; MMV
acceleration. Software-based acceleration is based on the where y is the K-space samples, x is the underlying image
2×1
theory of compressed sensing (CS) [2,3]. Compressed expressed in vectorized form, F: ℝN →ℂm×1 is the Fourier
sensing-based techniques exploit the spatial correlation in mapping from the real image space to the complex frequency
an MR image. They obtain only partial K-space samples domain and η is white Gaussian noise.
[4,5] of an image and then reconstruct it from these partial The underdetermined inverse problem (1) does not
samples by exploiting the correlation present in the image. have a unique solution. To reconstruct the underlying
In this article, we propose to further reduce the number of image, some prior information is necessary. In CS-
required partial K-space samples (thereby increasing accel- based MR reconstruction techniques, it is assumed that
eration) than required by conventional CS technique for the image to be reconstructed is approximately sparse
reconstructing T1/T2-weighted images by exploiting the in the wavelet domain [4,5]. This turns out to be a
interimage correlation present among the different images as realistic assumption and yields extremely good recon-
well as the spatial correlation within the image. struction results when partially sampled K-space data
Compressed sensing exploits only the spatial correlation are used.
of an MR image. In the current problem, one should The wavelet analysis and synthesis equations are:
remember that the T1/T2-weighted images correspond to the
same cross section. Thus, there is correlation among different Analysis : a = Wx ð2aÞ
images of the T1/T2-weighted images. We propose to
exploit both the spatial correlation in individual images as
well as the correlation among the images to further reduce
the K-space required for reconstruction. Synthesis : x = W T a ð2bÞ
The problem of jointly reconstructing multiple T1- or T2-
weighted MR images has not be identified or attempted The wavelet coefficient vector α is approximately s-
before. In this work, we will show how this problem can be sparse, i.e., only s coefficients are nonzeroes, while the rest
formulated into either a group sparse optimization problem are zeroes or very close to zero. Wavelet transform is
or a row sparse multiple measurement vector (MMV) orthogonal, in which case, the dimensionality of the original
optimization problem. These problems can be solved using image and the transform coefficients is the same, or the
existing algorithms. wavelet transform is a tight-frame, in which case, the
The rest of the article is organized into several sections. dimensionality of the wavelet coefficients is larger. In this
The following section discusses the mathematical formulation work, we will be using orthogonal wavelets.
behind the T1/T2 joint reconstruction problem. In Section 3, Incorporating the wavelet coefficients into Eq. (1),
the experimental results are provided. Finally in Section 4, the
conclusions of this work and future directions are discussed. y = FW T a + g ð3Þ
2. Reconstruction of T1/T2-weighted images Instead of solving for the image x directly, CS solves for
the wavelet coefficient of the image (α) (Eq. (3)). CS aims at
As mentioned earlier, most traditional MR scanners recovering only the s high-valued wavelet coefficients. The
collects dense K-space samples on a uniform Cartesian grid. coefficients that are close to zero are indistinguishable from
Reconstructing the image from such samples is trivial— noise and cannot be recovered.
applying an inverse Fourier transform. However, such dense In the pixel domain, an N×N image under consideration
sampling is time consuming, and one of the main challenges in has N2 unknowns. However, the pixel values are highly
MRI lies in reducing the MR acquisition time by reducing the correlated locally. The wavelet transform effectively
number of K-space samples collected. First, we will discuss “whitens” the image by removing the spatial correla-
the reconstruction of single MR image from partially sampled tions [6]. This leaves s nearly independent high-valued
K-space data. Later, we will describe how further reduction in wavelet coefficients.
the number of samples can be achieved by reconstructing all The inverse problem (Eq. (3)) has effectively only 2s
the T1 (or T2)-weighted images simultaneously. unknowns—s positions and s values. Intuitively, solving for
these 2s unknowns will require far fewer equations than
2.1. Reconstructing single MR images directly solving for the image, which has N2 unknowns.
The mathematical representation expressing the partially Compressed sensing-based recovery algorithms solves for
sampled K-space data acquisition model for the underlying the sparse wavelet transform coefficients by solving the
image can be expressed as following optimization problem [4],
To get a reasonably good estimate of the signal via l1 different images of the same cross section to achieve better
minimization, the number of K-space samples needed is [3], reconstruction accuracy.
N2
m = O slog ð5Þ 2.3. Group/row sparse reconstruction of
s T1/T2-weighted images
Once the wavelet coefficients are recovered, the image T1- or T2-weighted images are obtained by changing the
is reconstructed by applying the wavelet synthesis repetition time or echo times of the magnetization. Such T1/
equation (2b). T2 weightings lead to differential contrast among the tissues.
However, it must be remembered that all these images are
2.2. Sparse reconstruction of T1/T2-weighted images different representations of the same anatomical cross
The T1- or T2-weighted images are obtained by imaging section. For a particular T1- or T2-weighted image, the
the same object at different repetition and echo times. contrast between two tissues may be high, leading to a sharp
Suppose, the K-space data for T such weighted images have edge, while for some other T1 or T2 weighting, the contrast
been acquired. This is represented by, may be reduced and the edge may be less pronounced.
However, since the underlying anatomy is the same, the
yi = Fxi + gi ; i = 1:::T ð6Þ positions of the tissue boundaries (edges) are preserved.
The wavelet transform effectively encodes the edges in an
where yi is the K-space sample for the ith-weighted image xi. image. When the edge is more pronounced, the wavelet
The problem is to reconstruct the T1/T2-weighted images coefficients are high. When the edge is not so sharp, the
(xi's) given their K-space samples (yi's). wavelet coefficients are low. In smooth areas, the wavelet
A straightforward application of CS will apply Eq. (4) to coefficients are zero. The following toy example illustrates
reconstruct each T1- or T2-weighted individually image the fact.
(for each i). Consider a small image matrix with sharp boundary for a
In this work, we aim at the joint reconstruction of all the particular T1 weighting (Fig. 1A). The 0's correspond to
T1- or T2-weighted images. First, we show how a simple tissue B and the 1's correspond to tissue W. The Haar
sparsity-based joint reconstruction problem can be formu- wavelet transform (Fig. 1B) captures the vertical disconti-
lated for these T1/T2-weighted images. nuity between W and B perfectly. The values in the last
Incorporating the wavelet transform, Eq. (6) can be column should be ignored, the wavelet transform assumes
expressed concisely as follows, that the signal is periodic and computes the boundary
2 3 2 32 3 2 3 between B and W.
y1 FW T 0 0 a1 g1 Fig. 1 shows that when the tissue boundary is
4 ::: 5 = 4 0 ::: 0 54 ::: 5 + 4 ::: 5 pronounced, the Haar wavelet coefficients along the
yT 0 0 FW T aT gT ð7Þ boundary are high. Now consider a different value of T1
weighting, such that the contrast between W and B is less
or y = Ua + g pronounced (Fig. 2A). The Haar wavelet transform of this
matrix is shown in Fig. 2B. We can see now that although the
The wavelet coefficients (αi's) for each T1/T2-weighted position of the high-valued Haar coefficients is the same,
image are sparse. Therefore, the combined vector α̂ is also their values have changed. As the contrast is reduced, the
sparse. Hence, we can invoke a CS recovery algorithm, value of the wavelet coefficients is reduced as well.
namely the l1 minimization to jointly solve for the This is a toy example, but this corroborates the fact that as
wavelet coefficients,
long as the anatomy of the brain slice does not change, the according to their positions as shown in Fig. 4. If we consider
positions of the high-valued wavelet transform coefficients the T coefficient vector α1,r, … αT,r as one group, then all the
will not change for different T1 or T2 weightings. coefficients in the vector α̃ form N2 groups, each group
Mathematically, this means that there should be a high having T coefficients.
degree of mutual correlation between any two T1- and T2- We have argued above that the wavelet coefficients
weighted images of the same anatomical cross section. should have similar values at similar positions. It is known
Fig. 3 shows the scatter plot between the wavelet that the wavelet transform coefficient vectors of individual
coefficients of two randomly chosen T2-weighted images MR images are sparse. If each αi is approximately s-sparse,
of rat's spinal cord. The plot shows that the correlation is then it follows from the argument that the high-valued
almost perfectly linear. coefficients of the different αi's will correspond to similar
The linear relationship corroborates our physical under- positions; we can say that the vector α̃is approximately
standing of the fact why wavelet coefficients of two T1- or “s-group sparse,” i.e., there are approximately only s groups
T2-weighted images have similar valued coefficients at that have high-valued wavelet coefficients and the rest
similar positions. of the groups have zero coefficients or coefficients close
to zero.
2.3.1. Group sparse formulation Therefore, instead of solving the l1 minimization (8), one
Since the wavelet transform we are considering is needs to incorporate group sparsity into the optimization
orthogonal, the wavelet coefficient vector for each image problem. Previous studies [7–9] have proposed the follow-
(αi) is N2 (same as the number of pixels in the image). The ing mixed l2,1 minimization for solving the group sparsity-
coefficients of the vector α̃ in Eq. (14) can be grouped promoting optimization problem,
1 = 2
P
T
The l2-norm a2k;r over the groups of correlated
k −1
k,r
coefficients α k=1…T promotes a dense solution within the
2
group (k), whereas the summation over the l -norm enforces
group sparsity, i.e., the selection of only a few groups.
Although group sparse optimization is being introduced for
the first time in MR image reconstruction, it has been used
(with slightly different measures on group sparsity) previ-
ously in color imaging [10–12].
the multiple T1/T2-weighted images as vectors as in Eq. (7), discussed in Section 2.2. It is assumed that each of the T1/
we stack them as columns of a matrix, T2-weighted images is approximately s-sparse. As there are
T such weighted images, the vector of the wavelet
Y = FW T â + ˆ ð10Þ coefficients of all images α̂ is approximately s·T -sparse.
Recovering ans·T -sparse vector via l1 minimization requires
where Y=[y1…yT], α̂=[α1…αT] and η̂=[η1…ηT] and and FWT 2
about M1 = O Tslog Ns K-space samples [15].
operates on each column of α̂.
However, we have argued in Section 3.3. that α̂ is
In this arrangement, coefficients with the same positions
actually s-group sparse (each group having T coefficients).
are along the rows of the matrix α̂ (see Fig. 5). This
Sparsity-promoting optimization via l1 minimization only
arrangement (10) is called the MMV formulation.
utilizes the information that each image is spatially
As argued earlier, the wavelet coefficients αk,r for T1/T2-
correlated. Spatial correlation leads to sparsity in the wavelet
weighted images of the same cross section will have similar
domain. In this work, we have identified that T1/T2-
values at the same positions. Thus, for each image k, if the
weighted images are correlated among themselves. There-
wavelet coefficient vector αk is s-sparse, then for the matrix α̂
fore, we exploit both the spatial correlation within each
only s-rows will have significantly high-valued coefficients
image and interimage correlation among different T1/T2-
and the rest will be zeroes or close to zeroes. Therefore,
weighted images. This leads to our proposed group sparsity-
according to the MMV formulation, the solution to Eq. (10)
promoting optimization problem.
should be approximately s-row sparse, i.e., s-rows of the
Instead of reconstructing a vector that is s·T-sparse (as
matrix α̂ will have high-valued coefficients, while the rest of
done by l1 minimization), we reconstruct a vector that is
the rows are zeroes or close to zero.
known to be s-group sparse (each group having T co-
To solve the inverse problem (10) with row-sparsity
efficients). We employ l2,1 minimization (9) to reconstruct
constraints, the following optimization has been proposed by
such a group sparse vector. The number of samples required
van den Berg and Friedlander [13],
to reconstruct an s-group sparse vector having T coefficients
argmin jjâjj2;1 subject to jjY − FW T âjjFrobenius Vâ in each group is M2;1 = O Ts + slog Ns
2
[15]. This is
!1 = 2
ã X
N2 X T significantly less than the number of samples required by
where jjâjj2;1 = ak;r
2
: ð11Þ l1 minimization (M1). The number of measurements required
r=1 k =1 by the row sparse MMV formulation (11) is the same as
group sparse formulation.
The objective function for both the group sparsity
Thus, our proposed methods (l2,1 minimization and MMV)
problem (9) and the row sparsity problem (11) are the
expect to achieve the same level of reconstruction accuracy but
same, only the constraints are different.
with lesser number of K-space samples than required by l1
The MMV formulation has previously been used for
minimization. In other words, with the same number of K-space
seismic imaging problems [14] but is being introduced in
samples, our proposed group/row sparsity-based methods will
this article for the first time in the context of MR
yield better reconstruction results.
image reconstruction.
2.4. Benefit of group/row sparse reconstruction
3. Experimental evaluation
In the problem of joint reconstruction of T1/T2 maps by
sparse optimization (8), l1 minimization is employed as The experimental evaluation was carried out on ex vivo
and in vivo T2-weighted images of a rat's spinal cord. The
Table 1
Reconstruction results (NMSE) for ex vivo data
Reconstruction method 40 Lines 80 Lines 120 Lines
CSMRI 0.14 0.06 0.04
ExtSparseMRI 0.14 0.06 0.04
GroupSparseMRI 0.11 0.03 0.01
RowSparseMRI 0.11 0.03 0.01
Fig. 5. Grouping of wavelet coefficients according to their position.
6 A. Majumdar, R.K. Ward / Magnetic Resonance Imaging xx (2011) xxx–xxx
Fig. 6. From top to bottom—reconstructed images of the same slice for 4 different T2 weightings; left column—ground truth; middle column—CSMRI
reconstruction; right column—GroupSparseMRI reconstruction.
A. Majumdar, R.K. Ward / Magnetic Resonance Imaging xx (2011) xxx–xxx 7
In this work, we experimented on T2-weighted images [3] Candès EJ, Romberg J. Tao, Robust uncertainty principles: exact
since these were the only data available to us. However, the signal reconstruction from highly incomplete frequency information.
IEEE Trans Info Theory 2006;52(2):489–509.
techniques in this work are applicable to T1-weighted image [4] Lustig M, Donoho DL, Pauly JM. Sparse MRI: the application of
reconstruction as well. compressed sensing for rapid MR imaging. Magn Reson Med 2007;58
(6):1182–95.
[5] Trzasko J, Manduca A. Highly undersampled magnetic resonance
image reconstruction via homotopic 10-minimization. IEEE Trans
4. Conclusion Med Imaging 2009;28(1):106–21.
[6] Strickland RN, Hee HI. The wavelet transform as a multi-
This work looks at the problem of reconstructing a set resolution matched filter and zero-crossing detector for detecting
of T1- or T2-weighted MR images from their partially microcalcifications in mammograms. Eng Med Biol Mag 1995:
1047–8.
sampled K-space data. Such a set of images is collected [7] Eldar YC, Kuppinger P, Bolcskei H. Block-sparse signals: uncertainty
to construct T1 or T2 maps of the anatomical cross relations and efficient recovery. IEEE Trans Signal Proc 2010;58(6):
section under study [21,22]. 3042–54.
Compressed sensing-based techniques have previously [8] Stojnic M, Parvaresh F, Hassibi B. On the reconstruction of block-
been used to successfully reconstruct individual MR images sparse signals with an optimal number of measurements, arXiv:
0804.0041v1.
from their partial Fourier measurements. For the T1/T2 [9] Huang J, Zhang T. The Benefit of Group Sparsity, arXiv:0901.2962v2.
imaging problem, we show that instead of reconstructing the [10] Majumdar A, Ward RK. Compressive color sensing of color images.
images individually by only exploiting their sparsity in Signal Proc 2010;90(12):3122–7.
wavelet domain (spatial correlation) better results can be [11] Majumdar A, Ward RK. Compressive color imaging with group-
achieved when further information regarding the correlation sparsity on analysis prior. Int Conf Image Proc 2010 (in press).
[12] Majumdar A. Ward, Compressive non-convex group sparsity:
amongst the different T2-weighted images is exploited along application to color imaging. Int Conf Acoustics, Speech Signal Proc
with spatial correlation. 2010:469–72.
We have shown that the correlation among the T1- or T2- [13] van den Berg E, Friedlander MP. Theoretical and empirical results for
weighted images of the same anatomical cross section can be recovery from multiple measurements. IEEE Trans Info Theory
used to formulate a group sparsity-promoting optimization 2010;56(5):2516–27.
[14] Herrmann FJ. Compressive imaging by wavefield inversion
problem or a row sparsity-promoting MMV problem. Our with group sparsity, SEG, 2009, Houston. Technical Report
proposed technique yields significant improvement in the TR-2009-01.
reconstruction results. By thorough experimentation on ex [15] Baraniuk RG, Cevher V, Duarte MF, Hegde C. Model-based
vivo and in vivo T2-weighted images of a rat's spinal cord, compressive sensing, arXiv:0808.3572v5.
we have shown that it is possible to reduce the sampling [16] Xie J, Lai P, Huang F, Li Y, Li D. Cardiac magnetic resonance imaging
using radial k-space sampling and self-calibrated partial parallel
requirement by around 33% (leading to faster scan times) reconstruction. Magn Reson Imaging 2010;28(4):495–506.
compared with standard CS-based methods. Besides, our [17] Zhang S, Block KT, Frahm J. Magnetic resonance imaging in real time:
proposed reconstruction is also significantly faster (two to advances using radial FLASH. J Magn Reson Imaging 2010;31(1):
four times). 101–9.
This study looks at the problem of acquiring fast MRI [18] Matej S, Fessler JA, Kazantsev IG. Iterative tomographic image
reconstruction using Fourier-based forward and back-projectors. IEEE
scans from a signal processing perspective. Previously Trans Med Imaging 2004;23(4):401–12.
hardware-based acceleration methods were used for acquir- [19] O'Connor YZ, Fessler JA. Fourier-based forward and back-projectors
ing T1- or T2-weighted images [23]. In the future, we will in iterative fan-beam tomographic image reconstruction. IEEE Trans
extend our proposed group/row sparsity-based reconstruc- Med Imaging 2006;25(5):582–9.
tion algorithms to the hardware-based acceleration methods [20] Crozier S, Field J, Doddrell DM. Signal-to-noise ratio improvements in
in vivo high resolution micro-volume selected spectroscopy. Magn
to achieve further speed-ups in data acquisition. Reson Imaging 1990;8(3):277–83.
[21] Laule C, Kolind SH, Bjarnason TA, Boon Li DK, MacKay AL. In vivo
multiecho T2 relaxation measurements using variable TR to decrease
References scan time. Magn Reson Imaging 2007;25(6):834–9.
[22] Deichmann R, Adolf H, Nöth U, Morrissey S, Schwarzbauer C. A.
[1] Haacke EM, Brown RR, Thomson MR, Venkatesh R. Magnetic Haase, Fast T2-mapping with SNAPSHOT FLASH imaging. Magn
resonance imaging: physical principals and sequence design. New Reson Imaging 1995;13(4):633–9.
York: John Wiley & Sons; 1999. [23] Block KT, Uecker M, Frahm J. Model-based iterative reconstruction
[2] Donoho D. Compressed sensing. IEEE Trans Info Theory 2006;52(4): for radial fast spin-echo MRI. IEEE Trans Med Imaging 2009;28(11):
1289–306. 1759–69.