Truncation Effects in SENSE Reconstruction: Lei Yuan, Leslie Ying, Dan Xu, Zhi-Pei Liang 4
Truncation Effects in SENSE Reconstruction: Lei Yuan, Leslie Ying, Dan Xu, Zhi-Pei Liang 4
Abstract
Finite sampling is an important practical issue in Fourier imaging systems. Although data truncation effects are well understood in
conventional Fourier imaging where a single uniform receiver channel is used for data acquisition, this issue is not yet fully addressed in
parallel imaging where an array of nonuniform receiver channels is used for sensitivity encoding to enable sub-Nyquist sampling of k-space.
This article presents a systematic analysis of the problem by comparing the truncation effects in parallel imaging with those in conventional
Fourier imaging. Specifically, it derives a convolution kernel function to characterize the truncation effects, which is shown to be
approximately equal to that associated with the conventional Fourier imaging scheme. This article also describes a set of conditions under
which significant differences between the truncation effects in parallel imaging and conventional Fourier imaging occur. The results should
provide useful insight into interpreting and reducing data truncation effects in parallel imaging.
D 2006 Elsevier Inc. All rights reserved.
Keywords: Truncation effects; SENSE; Parallel imaging
Fig. 1. Simulation results to illustrate that the truncation effects are almost the same for SENSE and Fourier reconstructions, as indicated by Eq. (14): (A) True
object function. (B) Sensitivity functions used for generating sensitivity-encoded data for SENSE reconstructions. (C–D) Fourier reconstructions (from a single
uniform channel) with 128 and 32 k-space samples, respectively. (E–F) SENSE reconstructions with 64 and 16 k-space samples for each channel with R = 2.
2 3 2 3
support of B, q(xmB̂)s l (xmB̂) for different m =0, 1, . . ., qð xÞ d 1 ð xÞ
R1 will overlap, producing the well-known aliasing effect 6 q x B̂B 7 6 d 2 ð xÞ 7
U¼6
4
7
5 and d ¼ 6
4 v 5:
7
in d l (x). Based on Eq. (3), we can remove the aliasing error v
in d l (x) to obtain the desired image function q(x). q x ð R 1ÞB̂
B dL ð x Þ
Specifically, rewriting Eq. (3) in matrix form yields
The least-squares solution for q is given by1
U ¼ d;
SU ð4Þ 1
U ¼ SH S SH d; ð5Þ
where
1
2 3 The SENSE reconstruction formula was derived based on the
s1 ð xÞ s1 x B̂B N s1 x ð R 1ÞB̂
B assumption that q(x) is a summation of voxel functions [2]. For finite
6 s2 ð xÞ s2 x B̂
B
N
B 7
s2 x ð R 1ÞB̂
sampling, the voxel model has to be truncated to a finite sum, which is
S¼6
4 v
7;
5 equivalent to the data truncation effect analyzed in this article by assuming
v O v q(x) to be a general support-limited function. When d-type voxel functions
sL ð xÞ sL x B̂B N sL x ð R 1ÞB̂
B are used (as is often the case in practice), the two effects are identical.
L. Yuan et al. / Magnetic Resonance Imaging 24 (2006) 1311–1318 1313
which is the basic SENSE reconstruction formula. which is a periodic function with period B̂. To further relate
Equation (5) implicitly assumes that an infinite number of the truncation effect of multichannel parallel imaging to that
samples D l (nDk̂) is available, which is not valid in practice. of conventional single-channel Fourier imaging, we show in
This article analyzes the truncation effect in SENSE the Appendix that
reconstruction in comparison with those seen in conven- R1
X
tional Fourier imaging using a single uniform data hR ð x Þ ¼ h x mB̂
B ; ð12Þ
acquisition channel with Nyquist sampling. m¼0
2. Theoretical analysis
In conventional Fourier imaging using a single, uniform
data acquisition channel, the Fourier reconstruction of q(x)
from N data points, D(nDk), n = N/2, . . ., N/21, taken at
the Nyquist interval (Dk =1/B) is given by [17]:
N =21
X
q ð xÞ ¼ Dk
q̂ DðnDk Þei2pnDkx ; ð6Þ
n¼N=2
where h(x) is the same as in Eq. (8). Comparing Eq. (13) expression relating the final SENSE reconstruction (with the
with Eq. (3) immediately yields the SENSE reconstruction Dirac-d voxel function) to the ideal image function. A
partial expression was obtained in Ref. [19] for a simplified
q R ð xÞcqð xÞ4hð xÞ ¼ q̂
q̂ q ð xÞ: ð14Þ
case, although no explicit expression for the convolution
kernel function was given. In Refs. [20,21], a spatial
Eq. (14) is a key result of this article; it confirms that the
response function [equivalent to h R (x)] was used to describe
truncation effects in SENSE reconstruction, q̂R (x), can be
the truncation effects in q(x)s l (x), which was incorporated
described by a convolution equation. More interestingly, the
convolution kernel function is approximately equal to that in the image voxel model, but the final point spread function
for the SENSE reconstruction was not derived. Therefore,
associated with the conventional Fourier imaging scheme.
Eq. (14) should provide additional insight into interpreting
This result is illustrated in Fig. 1.
the truncation effects in SENSE reconstruction. In practice,
Note that Eq. (13) was derived under the assumption that
various factors can affect the approximation accuracy of
s l (x) is a smooth function in the sense that the spatial
Eq. (14). In this section, we discuss several key factors
variations of s l (x) over any interval of length 1/(NDk) [the
and some methods to mitigate the truncation effects.
effective width of h(x)] are negligible. When this assump-
tion is violated, we will see noticeable difference in 3.1. Boundary discontinuities
truncation effects in SENSE reconstruction. This and other
It is well understood that discontinuities in q(x) will
related issues will be discussed in Section 3.
produce noticeable Gibbs ringing in its Fourier reconstruc-
tions. When q(x) is represented by the Fourier series,
additional discontinuities can occur at both ends of the
3. Discussion
image function because of periodic extension implicitly
The problem of data truncation in parallel imaging had imposed by the Fourier series. This is not an important
been addressed in previous work especially [19-21]. issue in conventional Fourier imaging, because these
However, to our knowledge, Eq. (14) is the first explicit discontinuities, if any, are usually at the noise level.
Fig. 3. Simulation results to illustrate the effect of inaccurate s l (x) on the truncation artifacts: (A) SENSE reconstruction using M = 32 and R = 4 with the true
sensitivity functions in (B) and (C) SENSE reconstruction using M = 32 and R = 4 with the estimated sensitivity functions in (D). The sensitivity functions were
not estimated correctly in the background region, which leads to additional artifacts on top of the truncation effects in (A).
L. Yuan et al. / Magnetic Resonance Imaging 24 (2006) 1311–1318 1315
Fig. 6. Simulation results to illustrate that the truncation effects can be amplified when the encoding matrix S is poorly conditioned. (A) SENSE reconstruction
with R = 2 and M = 64 using the s l (x) shown in (B). (C) SENSE reconstruction with R = 2 and M = 64 using the s l (x) shown in (D). Note that the S matrix for the
sensitivities in (B) is better conditioned than for those in (D); hence, the artifacts in (A) are smaller than those in (C).
smoother with respect to h R (x). Therefore, the larger the M 3.5. Methods to mitigate truncation effects in SENSE
is, the closer the truncation effects in SENSE are to those in
A number of methods have been proposed for reducing
the conventional Fourier reconstruction. This is illustrated in
Gibbs ringing artifacts in Fourier imaging [23]. Conventional
Fig. 5, where we compare Fourier reconstruction with
filtering methods are rather effective, but improvement
SENSE reconstruction for different truncation levels using a
comes at the expense of spatial resolution. Superresolution
real data set acquired from a brain phantom.
methods [23] require the use of a priori constraints in one
3.4. Effect of an ill-conditioned S matrix form or another. All of these methods are directly applicable
to parallel imaging. When R is much smaller than L, as is
The truncation effect in SENSE is also dependent on the
sometimes the case, specialized methods are also available to
numerical condition of the SENSE coefficient matrix S.
reduce the truncation effects in SENSE reconstruction by
When S is ill conditioned, the data truncation effects can be
taking advantage of the bredundantQ information collected.
enhanced. This can be understood from a matrix perturba-
One such example is reported in Refs. [20,24], which seeks
tion analysis. Specifically, let d = d0+Dd and R =R0+DR in
the minimum-norm solution after incorporating the spatial
Eq. (5), where Dd and DR account for the truncation effects
response function into the SENSE reconstruction equation.
in the sensitivity-weighted image and in the final recon-
We have also observed a reduction of Gibbs ringing artifacts
struction, respectively, while d0 and R0 are the truncation
in regularized SENSE reconstruction if a good regularization
artifact-free components. It is well known [22] that when S
image is available. Detailed discussion of regularized
is ill conditioned, the same Dd can lead to a larger DR. This
SENSE is beyond the scope of the article. Interested
effect is also illustrated in Fig. 6.
readers are referred to Refs. [7,25,26].
In practice, several factors can affect the numerical
condition of S. Coil configuration is a big factor. Increasing
R without increasing L can also worsen the numerical
4. Conclusions
condition of S. Nonetheless, in practical SENSE imaging
experiments, R is often chosen conservatively to protect the This article presented a systematic analysis of the
numerical condition of S. truncation effects in SENSE. It was shown that the
L. Yuan et al. / Magnetic Resonance Imaging 24 (2006) 1311–1318 1317
Appendix A
Proof of Eq. (13).
Proof of Eqs. (10) and (11). dˆl ð xÞ ¼ ½qð xÞsl ð xÞ4hR ð xÞ
Let X
R1
¼ ½qð xÞsl ð xÞ4h x mB̂
B
1; n ¼ pR; for integer p
PR ðnÞ ¼ ; m¼0
0; otherwise
X
R 1 Z
and l
¼ qðsÞsl ðsÞh x mB̂
B s ds
l
1; N =2VnVN =2 1 m¼0
W ðnÞ ¼ :
0; otherwise
XZ l
R1
Then, c qðsÞh x mB̂
B s ds sl x mB̂
B
m¼0 l
X
M=21
dˆl ð xÞ ¼ Dk̂k Dl nDk̂k e i2pnDk̂k x
X
R1
n¼M=2
¼ ðq4hÞ x mB̂
B sl x mB̂
B :
m¼0
X
l
i2pnDkx
¼ Dk̂k Dl ðnDk ÞPR ðnÞW ðnÞe
n¼l Note that the b6Q holds when the variation of s l (x) over
" # " # any interval of 1/(NDk) [the effective width of the main lobe of
X
l X
l h(x)] is negligible. This is reasonable since s l (x) is generally a
¼ Dl ðnDk Þei2pnDkx 4 Dk̂k PR ðnÞW ðnÞei2pnDkx smooth function.
n¼l n¼l
" # " #
X
l X
l
¼ Dl ðnDk Þei2pnDkx 4 Dk̂k W ðmRÞei2pmRDkx References
n¼l m¼l
[1] Sodickson D, Manning W. Simultaneous acquisition of spatial
2 3 harmonics (SMASH): fast imaging with radiofrequency coil arrays.
X
M=21
Magn Reson Med 1997;38:591 – 603.
¼ ½qð xÞsl ð xÞ44Dk̂k ei2pmDk̂k x 5 [2] Pruessmann K, Weiger M, Scheidegger M, Boesiger P. SENSE:
m¼M =2 sensitivity encoding for fast MRI. Magn Reson Med 1999;42:952 – 62.
[3] Griswold MA, Jakob PM, Heidemann RM, Nittka M, Jellus V, Wang
JM, et al. Generalized autocalibrating partially parallel acquisitions
¼ ½qð xÞsl ð xÞ4hR ð xÞ: (GRAPPA). Magn Reson Med 2002;47:1202 – 10.
1318 L. Yuan et al. / Magnetic Resonance Imaging 24 (2006) 1311–1318
[4] Kyriakos WE, Panych LP, Kacher DF, Westin CF, Bao SM, Mulkern susceptibility artifacts in BOLD fMRI. Magn Reson Med 2002;
RV, et al. Sensitivity profiles from an array of coils for encoding and 48:860 – 6.
reconstruction in parallel (SPACE RIP). Magn Reson Med 2000; [16] McGee KP, Debbins JP, Boskamp EB, Blawat L, Angelos L, King KF.
44:301 – 8. Cardiac magnetic resonance parallel imaging at 3.0 Tesla: technical
[5] Heberlein K, Hu X. kSENSE: k-space sensitivity encoding recon- feasibility and advantages. J Magn Reson Imaging 2004;19:291 – 7.
struction. Proceedings of 9th ISMRM Scientific Meeting and [17] Liang ZP, Lauterbur PC. Principles of magnetic resonance imaging: a
Exhibition; 2001. p. 770. signal processing perspective. New York7 IEEE Press; 2000.
[6] Pruessmann KP, Weiger M, Bfrnert P, Boesiger P. Advances in [18] Haacke EM, Brown RW, Thompson MR, Venkatesan R. Magnetic
sensitivity encoding with arbitrary k-space trajectories. Magn Reson resonance imaging: physical principles and sequence design. New
Med 2001;46:638 – 51. York7 Wiley-Liss; 1999.
[7] King KF, Angelos L. SENSE image quality improvement using [19] Zhao X, Prost RW, Li Z, Li SJ. Reduction of artifacts by optimization
matrix regularization. Proceedings of 9th Annual Meeting of of the sensitivity map in sensitivity-encoded spectroscopic imaging.
ISMRM; 2001. Magn Reson Med 2005;53:30 – 4.
[8] Katscher U, Kohler T. Under-determined SENSE. Proceedings of [20] Tsao J, Sánchez J, Boesiger P, Pruessmann KP. Minimum-norm
Workshop on Minimum MR Data Acquisition Methods: Making reconstruction for optimal spatial response in high-resolution SENSE
More with Less. Marco Island, Florida, USA; 2001. p. 42 – 5. imaging. Proceedings of 11th ISMRM Scientific Meeting and
[9] Blaimer M, Breuer F, Mueller M, Heidemann RM, Griswold MA, Exhibition; 2003. p. 14.
Jakob PM. SMASH, SENSE, PILS, GRAPPA, how to choose the [21] Dydak U, Weiger M, Pruessmann KP, Meier D, Boesiger P.
optimal method. Top Magn Reson Imaging 2004;15:223 – 36. Sensitivity-encoded spectroscopic imaging. Magn Reson Med
[10] Bammer R, Auer M, Keeling SL, Augustin M, Stables LA, Prokesch 2001;46:713 – 22.
RW, et al. Diffusion tensor imaging using single-shot SENSE-EPI. [22] Golub GH, Van Loan CF. Matrix computations. 2nd ed. Baltimore
Magn Reson Med 2002;48:128 – 36. (MD)7 Johns Hopkins University Press; 1989.
[11] Larkman DJ, deSouza NM, Bydder M, Hajnal JV. An investigation [23] Liang ZP, Boada F, Constable T, Haacke EM, Lauterbur PC, Smith
into the use of sensitivity-encoded techniques to increase temporal MR. Constrained reconstruction methods in MR imaging. Rev Magn
resolution in dynamic contrast-enhanced breast imaging. J Magn Reson Med 1992;4:67 – 185.
Reson Imaging 2001;14:329 – 35. [24] Sanchez-Gonzalez J, Tsao J, Dydak U, Desco M, Boesiger P,
[12] Larkman DJ, Bydder M, deSouza NM, Williams AD, Bearden FH, Pruessmann PK. Minimum-norm reconstruction for sensitivity-
Hajnal JV. SENSE in the abdomen and pelvis. Proceedings of 9th encoded magnetic resonance spectroscopic imaging. Magn Reson
ISMRM Scientific Meeting and Exhibition; 2001. p. 2012. Med 2006;55:287 – 95.
[13] de Zwart JA, Ledden PJ, Kellman P, van Gelderen P, Duyn JH. Design [25] Liang ZP, Ying L, Xu D, Yuan L. Parallel imaging: some signal
of a SENSE-optimized high-sensitivity MRI receive coil for brain processing issues and solutions. Proceedings of 2004 IEEE Interna-
imaging. Magn Reson Med 2002;47:1218 – 27. tional Symposium on Biomedical Imaging. Arlington, VA, USA,
[14] Weiger M, Pruessmann KP, Boesiger P. Cardiac real-time imaging April; 2004. p. 1204 – 7.
using SENSE. Magn Reson Med 2000;43:177 – 84. [26] Lin FH, Kwong KK, Belliveau JW, Wald LL. Parallel imaging
[15] Weiger M, Pruessmann KP, Õsterbauer R, Bfrnert PB, Boesiger P, reconstruction using automatic regularization. Magn Reson Med
Jezzard P. Sensitivity-encoded single-shot spiral imaging for reduced 2004;51:559 – 67.