You are on page 1of 7

JOURNAL OF MAGNETIC RESONANCE IMAGING 24:683– 689 (2006)

Original Research

EPI Image Reconstruction With Correction of


Distortion and Signal Losses
Guoxiang Liu, PhD,1,2* and Seiji Ogawa, PhD2

artifacts (1-4). Susceptibility-related field variation


Purpose: To derive and implement a method for correcting
geometric distortions and recovering magnetic resonance
along the slice-selective direction can result in severe
imaging (MRI) signal losses caused by susceptibility-in- signal losses, while in-plane inhomogeneity causes geo-
duced magnetic field gradients (SFGs) in regions with large metric distortions. Two distortions usually exist along
static field inhomogeneities in echo-planar imaging (EPI). the phase-encoding (PE) direction: pixel shift and de-
Materials and Methods: Factors to account for field inho-
formation (compression or expansion), which can be
mogeneities and SFGs were added in a traditional EPI found when a difference exists between the field at a
equation that was a simple Fourier transform (FT) for ex- given pixel and the static field or when a field gradient
pressing the actual k-space data of an EPI scan. The in- exists across individual pixels.
verse calculation of this “distorted EPI” equation was used Several methods have been suggested for overcoming
as a kernel to correct geometric distortions and reductions the distortions. Some methods correct the distortion
in intensity during reconstruction. A step-by-step EPI re- directly in an image space (5), while others apply a
construction method was developed to prevent complicated phase correction to k-space data or use an inverse ap-
phase unwrapping problems. Some EPI images of phantom
proach during image construction (1,6), such as conju-
and human brains were reconstructed from standard EPI
k-spaces. gate phase reconstruction (CPR) (6), the Weisskoff
method (7), the conjugate gradient method (CGM) (1,6),
Results: All images were reconstructed using the proposed and the truncated singular value decomposition (TSVD)
multistep method. Geometric distortions were corrected
method (1). In addition, single reference scans (8) and a
and SFG-induced MRI signal losses were recovered.
multireference technique (9,10) have been developed to
Conclusion: Results suggest that applying our method for reduce the N/2 ghosting and eddy current effect from
reconstructing EPI images to reduce distortions and MRI fast switching gradients. Also, some authors have re-
signal losses is feasible.
ported methods for recovering susceptibility-induced
Key Words: EPI; fMRI; distortion; signal loss; susceptibility magnetic field gradient (SFG)-induced signal losses.
effect The simplest method uses thinner slices to reduce the
J. Magn. Reson. Imaging 2006;24:683– 689. field change across a slice (3). However, this method
© 2006 Wiley-Liss, Inc. reduces both the signal-to-noise ratio (SNR) and the
spatial coverage per unit of time. Another method (11–
13) effectively compensates for field inhomogeneities
IN RECENT YEARS, echo-planar imaging (EPI) has be-
using multiple refocusing gradients. However, in prac-
come widely used as a rapid imaging method for appli-
tice, many repetitions are needed to sum up to a uni-
cations such as functional magnetic resonance imaging
form image. This disadvantage limits its value in fMRI
(fMRI), diffusion-weighted imaging, and cardiac imag-
experiments. Some methods using high-order field
ing. In EPI, main field inhomogeneity, large static field
compensation (14,15) and a two-shot technique (16,17)
inhomogeneity, and magnetic field gradients or signifi-
were more efficient but the repetition time (TR) was still
cant changes in the susceptibility at tissue/air and
doubled.
tissue/bone boundaries cause image distortions and
Until now, no method has been able to correct distor-
tions and recover SFG-induced MRI signal losses from a
1
Brain Information Group, National Institute of Information and Com-
standard EPI k-space, and most previous methods are
munications Technology, Kobe, Japan. not applicable for online processing because of the com-
2
Ogawa Laboratories for Brain Function Research, Hamano Life Sci- plicated computation necessary. To overcome the limi-
ence Research Foundation, Tokyo, Japan. tations of the previous methods, we have created a new
*Address reprint requests to: G.L., Brain Information Group, National
Institute of Information and Communications Technology, 588-2, EPI image reconstruction method that can perform on-
Iwaoka, Iwaoka-cho, Nishi-ku, Kobe, Hyougo 651-2492, Japan. line processing (can be inserted into an MRI reconstruc-
E-mail: lgx@po.nict.go.jp tion system and can reconstruct images in real time)
Received June 22, 2005; Accepted May 18, 2006.
and recover SFG-induced MRI signal losses.
DOI 10.1002/jmri.20672
Published online 4 August 2006 in Wiley InterScience (www.interscience. The standard EPI image reconstruction method (sim-
wiley.com). ple inverse Fourier transformation [FT]) is based on
© 2006 Wiley-Liss, Inc. 683
684 Liu and Ogawa

assumptions that the field is homogeneous and that f sfg 共x,y,m⌬k x ,n⌬k y 兲 ⫽ sinc 共koff共x,y兲W/2兲,
there is no susceptibility difference in the slice. Thus,
the standard EPI equation is just like a simple FT equa- where koff共x,y兲 ⫽ ␥G共x,y兲t共m⌬kx,n⌬ky兲
tion. Of course, with strong field inhomogeneity and ⫽ ␥G共x,y兲共TE ⫹ nT ⫾ m⌬t兲, (3)
SFGs, an ideal EPI image cannot be reconstructed us-
ing the standard (inverse FT) method. What the MRI where G is the intravoxel field gradient (in the slice-
scanner does now is very different from a simple FT. selective direction), t is the time from excitation to ac-
Our method reconstructs EPI images from a k-space quisition, and W is the width of the voxel in the SFG
using the inverse calculation of a “corrected” EPI equa- direction (slice-selective direction). The absolute value
tion in which field inhomogeneity factors and SFG fac- of fsfg共x,y,m⌬kx, n⌬ky兲 shows how much the strength of
tors are added. Based on this simple idea, we developed an MRI signal is reduced by SFG-induced dephasing.
a direct inverse calculation method. Additionally, we The value of fsfg共x,y,m⌬kx, n⌬ky兲 is bounded on [–1, ⫹1].
developed a more powerful multistep method that pre- When fsfg共x,y,m⌬kx, n⌬ky兲 ⬍ 0, the current voxel ap-
vents the SFG-induced phase unwrapping problem. pears in a wrapped phase, which is ␲ bigger/smaller
than in the non-SFG cases, so the SFG factor
fsfg共x,y,m⌬kx, n⌬ky兲 takes into account the signal losses
THEORY
and the phase wrapping caused by SFG. The SNR will
In conventional single-shot EPI, the acquired k-space be increased by this “unwrapped” factor in some strong
data, S共m⌬kx, n⌬ky兲, with matrix size N ⫻ N can be SFG cases.
represented as During MRI image reconstruction, an inverse discrete
Fourier transform (DFT) is used instead of an FT. Equa-
tions [1] and [2] can thus be rewritten in discrete form
S共m⌬k x ,n⌬k y 兲 ⫽ 兰兰␳共x,y兲exp共i␥⌬B共x,y兲共nT
without any data losses. Equations [2] and [3] can then
⫾ m⌬t兲兲exp共i共m⌬kxx ⫹ n⌬kyy兲兲dxdy, (1) be expressed in matrix form as

where ␳共x,y兲 is the spin density of the scanned object, ␥ S ⫽ AI, (4)
is the gyromagnetic ratio, n is the nthky line, m is the mth
reading out point, ⫺ Nx/2 ⱕ m ⬍ Nx/2, ⫺Ny/2 ⱕ n where S is the k-space, I is the ideal undistorted image,
⬍ Ny/2, Nx is the number of data points in the readout and A is the transform matrix operator, which is a
direction and Ny are the phase-encoding steps. ⌬t is the combination of the FT factor exp共i共m⌬kxx ⫹ n⌬kyy兲兲,
dwell time, T is the time interval between adjacent ky the SFG factor fsfg共x,y,m⌬kx, n⌬ky兲, and the field inho-
lines, ⌬B共x,y兲 is the magnetic field inhomogeneity, and mogeneity factor exp共i␥⌬B共x,y兲共nT ⫾ m⌬t兲. With a small
⌬kx and ⌬ky are gradient area increments in the readout SFG and a weak inhomogeneity, the transform matrix
and phase-encoding directions, respectively. This equa- operator A will be close to the FT factor. This is the
tion is used to explain why geometric distortion occurs. assumed condition when the standard FT image recon-
However, the equation does not take into account two struction method is used. SFGs and field inhomogene-
factors: T1/T2 relaxations and intravoxel dephasing by ity distort EPI images, but fortunately, Eqs. [2]–[4] show
SFGs. Previous methods (1,5,6,9) based on Eq. [1], cor- that matrix A can be obtained by measuring the field
rected geometric distortions well but did nothing to and SFG maps and then the ideal image I can be cal-
recover SFG-induced MRI signal losses even though culated from the k-space using numerical solution
large SFGs occasionally produced black holes in EPI methods, like SVD (1) or CGM (6).
images. Equations [2]–[4] can be simplified by assuming that
SFG-induced MRI signal loss problems have been ⌬t ⫽ 0. Because ⌬t is much smaller than T, ⌬t
studied by many researchers who want to acquire fMRI ⬍ T/Nx. Equation [4] then becomes
data with blood oxygenation level-dependent (BOLD)
contrast using T2*-weighted EPI sequences. Some S x ⫽ A x I x , x ⫽ 0,1, . . . ,N x ⫺ 1, (5)
methods have reduced these effects (3,4,11–16). Differ-
ent from T1/T2 relaxations, intravoxel dephasing where Sx and Ix are the xth column of S⬘ (the inverse FT
should not be ignored since it is position-dependent, process result of k-space S along the readout direction)
like field inhomogeneity. Thus, Eq. [1] should be rewrit- and the ideal image I, and Ax is the Ny ⫻ Ny matrix
ten as defined by

S共m⌬k x ,n⌬k y 兲 关A x 兴共y k ,y兲 ⫽ f sfg 共x,y,0,y k 兲


⫻exp共i␥⌬B共x,y兲ykT兲
⫽ 兰兰f sfg 共x,y,m⌬k x ,n⌬k y 兲␳共x,y兲exp共i␥⌬B共x,y兲 ⫻ exp共i2␲yky/Ny兲 y ⫽ 0,1 . . . Ny ⫺ 1
⫻ 共nT ⫾ m⌬t兲兲exp共i共m⌬kx ⫹ n⌬kyy兲兲dxdy⬘ (2) yk ⫽ 0,1 . . . Ny ⫺ 1⬘
(6)
where fsfg共x,y,m⌬kx, n⌬ky兲 is the SFG-induced MRI sig-
nal loss factor, which depends on the intravoxel field where 共x,y兲 is a pixel of the EPI image, yk is the ykth
gradient at (x,y) and the time of the relaxation. The readout line. Equation [6] effectively decouples the
effect of dephasing in MRI signals has been studied NxNx ⫻ NyNy linear system into Nx linear systems of size
previously. Based on the previous studies (2), we define NyNy.
EPI Correction of Distortion and Signal Loss 685

Equation 5 can be rewritten to an image equation, field inhomogeneity factors, and is produced by insert-
ing a phase variation into a DFT. ⌰ denotes element-
I x ⫽ A x⫺1 S x , x ⫽ 0,1,. . .,N x ⫺ 1, (7) by-element matrix calculation. When all elements of Asfg
are greater than 0, we define
where Ax⫺1 is the inverse matrix of Ax Since the operator
matrix Ax and a column of the one-dimensional inverse A sfg ⫽ 1 ⫺ A 1⫺sfg . (11)
Fourier transformed distorted k-space Sx are known, in
order to create a column of the EPI image Ix, Ax⫺1 should From Eq. [11], we can obtain
be calculated from Ax. We resolved this problem using
the SVD method (1). Ax is an Ny-dimensional matrix A sfg ⌰A pf ⫽ A pf ⫺ A 1⫺sfg ⌰A pf , (12)
operator that can be expressed in an orthogonal repre-
sentation of the form By and then Eq. [10] can be rewritten as

冘 冘 S ⫽ 共A pf ⫺ A 1⫺sfg ⌰A pf 兲I.
Ny
(13)
Ax ⫽ U VT ⫽ ␴ nu n v nT , (8)
n⫽1
By A⫺1
pf ⫻ Eq. 13, we obtain

where ␴n is the nth singular value and un and vn are A ⫺1 ⫺1


pf S ⫽ A pf 共A pf ⫺ A 1⫺sfg ⌰A pf 兲I
members of the orthonormal sets of vectors in the col-
umns of U and the rows of V. When the system is well ⫽ A ⫺1 ⫺1
pf A pf I ⫺ A pf 共A 1⫺sfg ⌰A pf 兲I.
conditioned, the inverse of the matrix operator Ax is
⫽ I ⫺ A ⫺1
pf 共A 1⫺sfg ⌰A pf 兲I (14)
given as
Then we obtain

Ny
⫺1
1
A ⫽ ជ vជ T .
u (9)
x
␴n n n I ⫽ A ⫺1 ⫺1
pf S ⫹ A pf 共A 1⫺sfg ⌰A pf 兲I. (15)
n⫽1

When field inhomogeneity and SFG factors are too In this equation, we know that A⫺1 pf is the inverse trans-

strong, the system will be ill conditioned, and small formation matrix of FT and field inhomogeneity, so A⫺1 pf S

singular values may be of the same order as the usual is the ideal image when the SFG factor is not taken into
numerical noise encountered in this computation. The account. A⫺1 pf 共 Asfg⌰Apf兲I is the MRI signal loss due to

TSVD method can be used to provide stable solutions in SFG. If we define


order to avoid artifacts from the basis vectors corre-
sponding to those very small singular values. In this I 0 ⫽ A ⫺1
pf S, (16)
method, the singular values are given a threshold and
only those above the threshold are included in the sum- and
mation of Eq. [9]. With appropriate selections of the
truncation level, TSVD can be used as a regularization B ⫽ A ⫺1
pf 共A 1⫺sfg ⌰A pf 兲, (17)
method. However, choosing a truncation level can be
shown to be equivalent to imposing a quadratic con- then Eq. [15] can be rewritten as
straint on the solution.
Unfortunately, such an inverse method does not al-
I ⫽ I 0 ⫹ BI. (18)
ways provide correct results, such as when strong SFGs
make a phase severely wrapped. The problem is that
Then,
the phase is defined only in the principal value range of
关⫺␲,␲兴. Any values outside this range will be wrapped.
During the calculation, the “wrapped phase” problem I ⫽ I 0 ⫹ B共I 0 ⫹ BI兲,
causes errors that lead to incorrect EPI images. The and
problem occurs during the inverse calculation of Ax⫺1


when a calculated phase is wrapped (outside of 关⫺␲,␲兴. N

The basic method of preventing a wrapped phase in- I⫽ BiI0 ⫹ BN ⫹ 1I. (19)
volves reducing the phase variation in the inverse cal- i⫽0
culation. This can be done by calculating over multiple
steps. Based on this idea, we rewrote the above equa- The absolute value of all elements of B are smaller than
tions. First we separated the transformation matrix A in 1, so in Eq. [19], when the number of calculation step N
Eq. [4] into two matrices, is big enough, all terms of BN⫹1I will approach 0, so with
an appropriate value for N, an ideal image without dis-
S ⫽ 共A sfg ⌰A pf 兲I, (10) tortion and MRI signal loss can be represented by


where Asfg is the transformation matrix operator for the N

SFG factor, and it can be produced by using Eq. [3]. Apf I⬇ B iI 0. (20)
is the transformation matrix operator for the FT and the i⫽0
686 Liu and Ogawa

It is obvious from the equations that to make a recon-


struction kernel, we use A1⫺sfg instead of Asfg. A1⫺sfg can
be calculated using a1⫺sfg ⫽ 1 ⫺ asfg (see Eq. [11]),
where asfg ⬎ ⫽ 0, and asfg and A1⫺sfg are elements of
Asfg and A1⫺sfg respectively. When asfg ⬍ 0, we should
use a1⫺sfg ⫽ ⫺ 1 ⫺ asfg to account for the phase
wrapping caused by SFG.
Equations [16], [17], and [20] are the basic equations
of our reconstruction method. Also, these equations
can be simplified by ignoring the time delay in a readout
line like in Eq. [7], and implementing the inverse calcu-
lation A⫺1pf using TSVD. Different from the direct inverse
Figure 1. Voxel of EPI measured in high resolution for field
calculation method that reconstructs ideal images from
and SFG maps. Rz is the slice-selective direction. Rx and Ry
a k-space using one transformation with only one ker- are the readout and phase encoding directions, respectively.
nel, the multistep method can reconstruct images us-
ing several steps with different kernels. Also these dif-
ferent kernels can be combined into one kernel:
SFG scan MRI signal for two adjacent small voxels in
I ⬇ KS the slice-selective direction, which are located in an EPI
voxel by:


N

K⫽ B i A ⫺1 (21) G ⫽ 共␪ 2 ⫺ ␪ 1 兲/共␥TE sfg D兲, (22)


pf
i⫽0
where G is the local SFG of the EPI voxel in the slice-
where K is the kernel that transforms the k-space S to selective direction, ␪1 and ␪2 are the phases of two ad-
the image I. jacent small SFG scan voxels in the slice-selective di-
rection, TEsfg is the echo time of the SFG scan, and D is
the thickness of the SFG scan (the distance between the
METHOD centers of two adjacent small SFG scan voxels in the
Like many previous methods, two steps are involved in slice-selective direction). It can be seen that the SFG
making an undistorted image. The first step involves scan is implemented under field inhomogeneity. But
taking a field map scan, calculating the field and SFG the in-plane field inhomogeneity will not influence the
maps, and calculating the kernels based on Eqs. [5] and SFG map calculation, because the two voxels from
[6] by TSVD for the direct inverse calculation method, or which ␪1 and ␪2 are calculated are at the same in-plane
Eqs. [16], [17], and [20] for the multistep method. The location.
second step involves taking EPI scans and reconstruct- For the direct inverse calculation method, the matrix
ing the images using the kernels in a step along the Ax is made using the inhomogeneity field and SFG
phase-encoding direction instead of the standard in- maps. The kernel matrix of reconstruction along the PE
verse fast FT (FFT). direction of Ax⫺1 can be calculated using the TSVD
method. Selecting the threshold value is a problem
when the system is ill conditioned. A value of 200 was
Field Map and SFG Map
used in our experiments. For the multistep method,
Since field inhomogeneity and SFGs produce distor- A1⫺sfg, Apf, A⫺1 ⫺1
pf , and B ⫽ Apf 共 A1⫺sfg⌰Apf兲 were calculated
tions, the distortions can be corrected by measuring the from the field and SFG maps.
field and SFG maps. Before EPI scans, a magnetic field
map must be acquired. The basic principle is to mea-
RESULTS
sure the change in the phase of the MRI signal for two
different echo times. The simplest way to do this is to The reconstruction methods described here were imple-
acquire two different images with identical parameters mented by inserting this algorithm (written in C⫹⫹ on
apart from the echo time (TE). A multiecho gradient a SIEMENS IDEA system) into the reconstruction sys-
echo sequence is used. Previous researchers (18) have tem of a 3-T SIEMENS Allegra scanner and applying the
demonstrated that in-plane SFGs (perpendicular to the methods to experimental data of phantoms and human
slice-selective direction) cause MRI signal losses, too; so volunteers. The images were acquired using a single-
theoretically, to calculate the intra-SFG map, we shot blipped-EPI sequence with centric reordering with
should measure a voxel of the EPI in high resolution, different TE and TR. The field of view was 22 ⫻ 22 cm2,
such as the one shown in Fig. 1a. The SFG should not the matrix size was 64 ⫻ 64, and the slice was 5-mm
be calculated along a specific direction. However, the thick.
SFGs in the slice-selective direction are the main cause The experimental results of some major correction
of MRI signal losses. In this work, we considered how to methods and our method are shown in Fig. 2. The EPI
recover these losses. As shown in Fig. 1b, an EPI voxel images were acquired with a TE of 53 msec, a TR of 113
is separated only in the slice-selective direction for an msec, and a 752-Hz/pixel bandwidth (BW). The PE di-
SFG map gradient echo scan. To calculate one pixel in rection was from top to bottom. There is an air bubble
an SFG map, we used the difference in the phase of the on top of the phantom seen in Fig. 2a. These figures
EPI Correction of Distortion and Signal Loss 687

Figure 2. Comparison of our method and other methods. a: Gradient echo image; TE ⫽ 10 msec, TR ⫽ 113 msec. b: Original
EPI, TR ⫽ 113 msec, TE ⫽ 53 msec, BW ⫽ 752 Hz. c: Conjugate phase reconstruction. d: Weisskoff method. e: TSVD. f: Direct
inverse calculation method.

illustrate that although the geometrical distortion of an between tissues and air cavities under this slice lead to
EPI image can be corrected using the previous meth- field inhomogeneities, as shown in Fig. 3c, and strong
ods, the reduction in intensity by SFGs (mainly around SFGs, the bright parts in Fig. 3f. Compared with the
the air bubble) can be corrected only using our method. gradient echo (GRE) image in Fig. 3a, the one-shot
Some MRI images of the brain of a human volunteer blipped EPI image in Fig. 3b had geometric distortions
are shown in Figs. 3 and 4. In Fig. 3, EPI distortion was that were caused by the field inhomogeneities shown in
mainly caused by SFGs and significant changes in sus- Fig. 3c. The phase-encoding direction of this EPI scan
ceptibilities, since the strong field inhomogeneities in was from top to bottom, so the weaker field, the dark
the field map in Fig. 3c could not be found when a area in Fig. 3c, made the eyes appear closer to the top.
phantom was measured. The susceptibility difference Also, in the area close to the air cavity, MRI signals

Figure 3. Experimental results (PE direction from top to bottom). a: Gradient echo image; TE ⫽ 10 msec. b: Original EPI; TR ⫽
500 msec; TE ⫽ 20 msec; BW ⫽ 3002 Hz/pixel. c: Field map. d: Reconstructed EPI image from k-space of (b) using the multistep
method. e: Reconstructed EPI image from k-space of (b) using direct inverse calculation method. f: Intravoxel SFG map.
688 Liu and Ogawa

Figure 4. Other results of the same experiment as in Fig. 3 when our multistep method was used. a: Reconstructed EPI image
when only field inhomogeneity factors were taken into account (the number of calculations of step N ⫽ 1. b: Reconstructed EPI
image when N ⫽ 1. c: Recoveries of signal losses in a pixel (center of the black hole in (a)) in a different calculation step.
d: Projections of 34th (center) column of EPI images (blue: original EPI; green: field inhomogeneities induced geometric
distortions corrected; red: geometric distortions and SFG-induced signal losses corrected). e: Projections of 34th (center) column
of reconstructed EPI images when N ⫽ 1, 2 . . . 20.

appeared with a position shift to the bottom by the ⫽ 1, 2 . . . 20 are shown in Fig. 4e. Figure 4c and e
stronger field. Simultaneously, significant signal losses indicate that after step 20, almost no signals were re-
can be found at the locations of bright areas in the SFG constructed. In other words, nearly all lost signals were
map in Fig. 3f. The EPI image reconstructed by the recovered before step 20.
direct inverse calculation method is shown in Fig. 3e. In
this image, the main geometric distortions were cor-
rected, but the signal losses could not be recovered DISCUSSION
correctly, and some new severe calculation errors ap-
peared since the inverse calculation system was ill con- The field and SFG map scans can be combined into a
ditioned (too many terms were truncated by TSVD). A multiecho gradient echo scan, and a slice of the EPI
better EPI image is shown in Fig. 3d, which was recon- scan should be separated into several thinner ones to
structed using the multistep method; both geometric be measured in the combined field–SFG map scan. The
distortions and signal losses by SFGs were corrected. phases of all small voxels in the field–SFG scan located
Figure 4 shows the advantages of our multistep in an EPI voxel should be averaged first for calculating
method. The image in Fig. 4a can be explained as the a field map, and the data of all echoes of the combined
multistep reconstructed EPI image at step zero (when field–SFG map scan can be used to construct an SFG
N ⫽ 0 in Eq. [20]). I is I0 in Eqs. [18]–[20], so MRI signal map just like using an SFG map scan.
losses have not yet been recovered, though the geomet- Our results showed that the EPI image reconstruc-
ric distortions have been corrected, and the shape of the tion method using a multistep calculation resolved the
reconstructed EPI image is close to that of the GRE SFG-induced phase wrapping problems that caused
image Fig. 3a. Figure 4b is the EPI image reconstructed bottlenecks in previous approaches. The proposed
by our multistep method when the number of calcula- methods are based on an inverse calculation of the ideal
tion step N ⫽ 1. In this image, the signal losses were EPI equation (considering field inhomogeneity and SFG
recovered incompletely, but when I0 N ⫽ 20 was used, factors). Strong SFG factors cause very small singular
we got a perfect image shown in Fig. 3d. Figure 4c values in the inverse calculation and cause the system
shows the recoveries of signal losses in a pixel in differ- to be ill conditioned. This is why the direct inverse
ent calculation step. The correction of geometric distor- calculation method does not work well with strong
tions and the recovery of signal losses can be illustrated SFGs (Fig. 3e). The multistep method resolves this
by Fig. 4d, which shows the difference in projections of problem by separating the SFG factors from the inverse
a column in the original EPI image in Fig. 3b, the geo- calculation (Eqs. [17], [19], and [20]). The inverse cal-
metric distortions corrected image in Fig. 4a, and the culation without the SFG factor works better because
final reconstructed image in Fig. 3d. The projections of this system can be kept well conditioned Thus, the
a column of reconstructed EPI images when N multistep method is the main contribution of this work.
EPI Correction of Distortion and Signal Loss 689

Because it is somewhat sensitive to noise in the field REFERENCES


map, the noise, the SFG map, and their effects on the 1. Kadah YM, Hu X. Algebraic reconstruction for magnetic resonance
reconstruction must be considered. Thus, some prepro- imaging under B0 inhomogeneity. IEEE Trans Med Imaging 1998;17.
cessing of the field map and the SFG map (e.g., low-pass 2. Glover CH. 3-D z-shim method for reduction of susceptibility effects
in BOLD fMRI. Magn Reson Med 1999;42:290 –299.
filtering) may be required. 3. Young IR, Cox IJ, Bryant DJ, Bydder GM. The benefits of increasing
In conclusion, we developed a technique for correct- spatial resolution as a means of reducing artifacts due to field
ing distortions in EPI images. Different from previous inhomogeneities. Magn Reson Imaging 1988;6:585–590.
4. Song AW. Single-shot EPI with signal recovery from the suscepti-
methods, our method includes the influence of field
bility-induced losses. J Magn Reson Imaging 2001;46:407– 411.
inhomogeneities and SFGs, which produce distortion 5. Hutton C, Bork A, Josephs O, et al. Image distortion correction
and MRI signal losses in EPI images, as factors in the in fMRI: a quantitative evaluation. Neuroimage 2002;16:217–
EPI image equation. An inverse EPI image equation that 240.
6. Munger P, Crelier GR, Peters TM. An inverse approach to the correc-
takes into account these factors is the key part of our tion of distortion in EPI images. IEEE Trans Med Imaging 2000;19.
method. In other words, our method is based on not 7. Weisskoff RM, Davis TL. Correcting gross distortion on echo planar
only the field map, but also the SFG map. images, In: Proceedings of the 11th Annual Meeting of SMRM,
We compared our method with some previously pro- 1992. p. 4515.
8. Reeder SB, Faranesh AZ, Ergin A, McVeigh ER. A novel object-
posed methods. The results showed that our method independent ‘balanced’ reference scan for echo-planar imaging. J
corrects both geometrical distortions and moderate Magn Reson Imaging 1999;9:847– 852.
MRI signal losses when SFG maps are measured cor- 9. Xin W, Gullberg GT, Parker DL, Zeng GL. Reduction of geometric
and intensity distortions in echo-planar imaging using a multi-
rectly.
reference scan. Magn Reson Med 1997;37:932–944.
This EPI image reconstruction approach was com- 10. Chen NK, Wyrwicz AM. Correction for EPI distortion using multi-echo
bined with an MR system. The system provides practi- gradient-echo imaging. Magn Reson Med 1999;41:1206 –1213.
cal, online (real-time) reconstruction. To reconstruct an 11. Frahm J, Merboldt KD, Hanicke W. Direct FLASH MR imaging of
magnetic field inhomogeneities by gradient compensation. Magn
image after FFT processing along the readout direction, Reson Med 1988;6:474 – 480.
we used our algorithm instead of an FFT along the 12. Ordidge RJ, Deniau JC, Knight RA, Helpern JA. Assessment of
phase-encoding direction. Thus, the number of compu- relative brain iron concentrations using T2-weighted and T2*-
weighted MRI at 3 Tesla. Magn Reson Med 1994;32:335–341.
tations rises from n log(n) to n ⫻ n, where n is the size
13. Constable RT. Functional MR imaging using gradient-echo echo-
of the resolution in the PE direction. Our method also planar imaging in the presence of large static field inhomogeneities.
increases the time required to compute our kernel pa- J Magn Reson Imaging 1995;5:746 –752.
rameters, but they need to be computed only once for a 14. Cho ZM, Ro YM. Reduction of susceptibility artifact in gradient-
echo imaging. Magn Reson Med 1992;23:193–196.
given field map and an SFG map. The extra time spent 15. Glover G, Lai S. Reduction of susceptibility effects in BOLD fMRI
is worthwhile for correcting a large number of images of using tailored RF pulses. In: Proceedings of the 6th Annual Meeting
the same slice (in fMRI studies, for instance). In exper- of ISMRM, Sydney, Australia, 1998. p. 298.
iments, for a 64 ⫻ 64 –pixel slice, the kernel computa- 16. Chen NK, Wyrwicz AM. Removal of intravoxel dephasing artifact in
gradient-echo images using a field-map based RF refocusing tech-
tion time was 10 seconds (Windows 2000 operating nique. Magn Reson Med 1999;42:807– 812.
system on a 2.8 GHz computer). 17. Mao J, Song AW. Intravoxel rephasing of spins dephased by sus-
We corrected SFG-induced MRI signal loss, which is a ceptibility effect for EPI sequences. In: Proceedings of the 7th An-
nual Meeting of ISMRM, Philadelphia, PA, USA, 1999. p. 1982.
kind of distortion. In future work, we will investigate the
18. Deichmann R, et al. Compensation of susceptibility-induced BOLD
possibility of restoring BOLD signal changes (when sensitivity losses in echo-planar fMRI imaging. Neuroimage 2002;
SFG-induced MRI signal losses occur) in fMRI studies. 15:120 –135.

You might also like