[Jeffrey A.

Fessler]

[A review of the use of iterative algorithms]

M
agnetic effects like field inhomogene-
resonance ity are important, then grid-
imaging ding/FFT methods for image
(MRI) is a reconstruction are subopti-
sophisticat- mal, and iterative algorithms
ed and versatile medical based on appropriate models
imaging modality. Tra - can improve image quality,
ditionally, MR images are at the price of increased
© BRAND X PICTURES
reconstructed from the raw computation. This article
measurements by a simple reviews the use of iterative
inverse two-dimensional (2-D) or three-dimensional (3-D) fast algorithms for model-based MR image reconstruction. The ref-
Fourier transform (FFT). However, there are a growing number erences give pointers to some recent work but are by no means
of MRI applications where a simple inverse FFT is inadequate, a comprehensive survey. To see more citations, visit http://www.
e.g., due to non-Cartesian sampling patterns, non-Fourier physi- eecs.umich.edu/~fessler/.
cal effects, nonlinear magnetic fields, or deliberate under-sam-
pling to reduce scan times. Such considerations have led to MRI BACKGROUND
increasing interest in methods for model-based image recon- Any signal processing method aimed at forming images from
struction in MRI. measurement devices such as MRI scanners must consider the
relevant physics. A survey in IEEE Signal Processing Magazine
INTRODUCTION [2] and a book written from a signal processing perspective [3]
The inverse FFT has served the MR community very well as the have described MRI physics well. Here we review the physics in
conventional image reconstruction method for k-space data a somewhat unconventional way that facilitates describing some
with full Cartesian sampling. And for well sampled non-Carte- of the “non-Fourier” aspects of MRI.
sian data, the gridding method with appropriate density com-
pensation factors [1] is fast and effective. But when only MRI PHYSICS
under-sampled data is available, or when non-Fourier physical Standard MRI scanners use a large static magnetic field
S S S S
Digital Object Identifier 10.1109/MSP.2010.936726 B 0 1 r 2 5 B0 1 r 2 k (1)

1053-5888/10/$26.00©2010IEEE IEEE SIGNAL PROCESSING MAGAZINE [81] JULY 2010

 S S S S
to induce a net magnetization M 5 Mx i 1 My j 1 Mz k at each transverse magnetization. The precise relationship is gov-
S S S
point in space in the body being imaged, where i , j , and k erned by the Bloch equation [2]. For most image recon-
denote the unit vectors along the x, y, and z axes, respective- struction purposes, it suffices to consider just two aspects
S
ly, and r 5 1 x, y, z 2 denotes 3-D spatial coordinates. Ideally, of the full relationship: precession and transverse relax-
S
the static field strength B0 1 r 2 would be spatially uniform, i.e., ation. The most important equation in MRI is the Larmor
S
a single constant B0 . In practice, it is never perfectly uniform, relation: v 5 g| B |, which states that the magnetization
due to the unavoidable nonuniformities of all practical coil precesses (around the axis of the applied field) at a frequen-
designs and due to the field strength variations that are cy v that is proportional to the magnitude of the applied
induced by the nonuniform magnetic susceptibilities of differ- field. The constant of proportionality g is called the gyro-
ent tissue types. The electron distributions in different mole- magnetic ratio and is about 42.6 MHz/T for hydrogen pro-
cules also influence the local magnetic environment tons. During a readout, only the longitudinal component of
S
experienced by an atom’s nucleus, called chemical shift. Some B is varied usually, i.e.,
types of MRI scans are robust to such spatial variations of B0 ;
S S S S
others are sensitive to nonuniformities, necessitating correc- B 1 r , t 2 5 Bz 1 r , t 2 k , (4)
tion methods.
S
At equilibrium (which is established within a few seconds so the magnetization precesses around k , i.e., within the trans-
S
for a stationary object), the magnetization M is aligned with verse plane. This property is why the complex representation (3)
the applied static field and its magnitude is proportional to the is convenient, because precession can be expressed using a com-
S
product of B0 1 r 2 and the object-dependent local density of plex phase in this form. In general, the applied longitudinal field
S
(predominately) hydrogen protons or “spins.” This proton den- strength Bz 1 r , t 2 varies both spatially and temporally, so the
sity alone is of only modest interest in MRI; in practice one Larmor relationship describes the instantaneous frequency at a
S S
applies time-varying magnetic fields B 1 r , t 2 that induce time- given spatial location
varying changes in the magnetization
S S
v 1 r , t 2 5 g Bz 1 r , t 2 . (5)
S S S S S S S S
M 1 r , t 2 5 Mx 1 r , t 2 i 1 My 1 r , t 2 j 1 Mz 1 r , t 2 k . (2)
Without loss of generality, let t 5 0 be the time when the excita-
These changes depend on time constants (tissue-dependent tion pulse is completed, and consider some time point t . 0
relaxation parameters) and other factors, and the goal in MRI during the readout. The precession of the transverse magnetiza-
is to form images of aspects of this magnetization. By manipu- tion between time zero and time t corresponds to a net phase
S S
lating the applied field B0 1 r , t 2 appropriately, sometimes in that is the integral of the instantaneous frequency (5), i.e., ide-
conjunction with injected or inhaled contrast agents, one can ally we would have
examine a multitude of different tissue properties. t
S S S
An MRI scan consists of one or more alternations between M 1 r , t 2 5 M 1 r , 0 2 expa2i3 v 1 r , tr 2 dtrb.
0
two stages: excitation and readout. During the excitation stage,
S S
the applied magnetic field B 1 r , t 2 is designed to tip the mag- In practice, microscopic variations in the magnetic field
S
netization vectors M within some slice or slab away from equi- cause the spins within a given voxel to become out of phase
librium, so that they have a component in the transverse plane, over time. So the transverse magnetization vector’s magni-
i.e., the 1 x, y 2 plane. It is convenient to represent this trans- tude decreases approximately exponentially with a time con-
verse component mathematically using a complex function stant T *2 . Accounting for this decay, an accurate model for the
defined as follows: temporal evolution of the transverse magnetization during a
readout is
S S S
M 1 r , t 2 ! Mx 1 r , t 2 1 i My 1 r , t 2 , (3) t
S S * S S
M 1 r , t 2 5 f 1 r 2 e 2t/T2 1 r 2expa2 ig3 Bz 1 r , tr 2 dtrb, (6)
0
where i ! "21. Note that the field components Mx and
S S
My are real physical quantities; the “transverse magnetiza- where f 1 r 2 ! M 1 r , 02 denotes the object’s transverse magne-
S
tion” M 1 r , t 2 is complex solely by definition. The excitation tization immediately after excitation. A typical goal in MRI is
S S
process can be quite complicated to model and is beyond to form an image of f 1 r 2 . The properties of f 1 r 2 depend not
the scope of this article. See [2] for an introduction to the only on spin density, but also on the type of excitation used.
role that signal processing plays in the design of excitation Note that for simplicity of exposition, we focus here on the
S
pulses and [4] for some recent model-based RF pulse case where the object is static so that f 1 r 2 is not a function of
design methods. time t. Generalizations to dynamic imaging are very active
S S
During the readout stage, the applied field B 1 r , t 2 is research areas in MR image reconstruction.
manipulated in ways that help elucidate the transverse The relaxation factor T 2* varies spatially, and often is on the
S
magnetization M 1 r , t 2 . For image reconstruction, it is order of 10 ms. This relatively rapid decay is a significant limi-
essential to model the effects of the applied field on the tation in MRI. If T *2 were longer, then a signal excitation stage

IEEE SIGNAL PROCESSING MAGAZINE [82] JULY 2010

followed by a (lengthy) readout NOTIONS OF SPARSITY imbalance.) Digitally, these two
stage could be sufficient to signals can be combined and
HAVE DEEP ROOTS IN STATISTICAL
form a high-resolution image stored as complex values, i.e.,
S SIGNAL PROCESSING.
of f 1 r 2 . In practice, the rapid we record samples
decay limits how much spatial
information can be recorded in a single readout stage, so such I 1 mDT 2 1 iQ 1 mDT 2 , m 5 1, c, nd,
“single shot” imaging, such as echo-planar imaging (EPI) [5],
provides only modest spatial resolution. Therefore, high-reso- where DT denotes the sampling rate (typically around 1 ms)
lution imaging uses multiple alternations between excitation and nd denotes the number of recorded samples, typically
stages and readout stages, each with different variations of the 64–512 for a given readout stage. Again, the physical quanti-
S
applied field Bz 1 r , t 2 . ties are real, but complex quantities are defined in terms of
those physical quantities for convenience. (In some sys-
DATA ACQUISITION: THE MR SIGNAL tems, digital demodulation is used, but the modeling
S
By Faraday’s law, the time-varying magnetization M r , t will 1 2 remains identical.)
induce an electromotive force (emf) in a nearby coil. The emf
will be proportional to the volume integral of the time deriva- SIGNAL MODEL
S
tive of the magnetization M 1 r , t 2 multiplied by the coil To improve signal-to-noise ratio and reduce acquisition
S
response pattern c 1 r 2 . The resulting electrical potential v 1 t 2 times, the use of multiple receive coils has become increas-
across the receive coil is ingly popular in MRI. Although originally called phased array
imaging [6], a term that resonates with other signal process-
S d S S
1 2 1 2
v t 5 reala 3 c r M r , t d r b, 1 2 (7) ing applications involving multiple receivers, today the use
dt
of multiple receive coils in MRI is usually called parallel
where real(·) denotes the real part of a complex number. The imaging [7].
S S
1 2
coil response c r generally decreases with distance from the Let cl 1 r 2 denote the sensitivity (response pattern) of the lth
coil. If uncorrected, this nonuniformity causes spatial variations coil, for l 5 1, c, L, where L denotes the number of coils. Let
in signal strength that can be a challenge for image processing sl 1 t 2 denote the demodulated “MR signal” associated with the
methods like segmentation algorithms. Numerous correction lth coil, defined as in (9). Substituting (6) into (9) and simplify-
methods have been developed. ing yields the following general forward model for the MR signal
Because the time constant T 2* is on the order of millisec- associated with the lth coil
onds whereas the phase variations in (6) are many MHz, it is
S S * S S S
very reasonable to use a narrow-band approximation when sl 1 t 2 5 3 cl 1 r 2 f 1 r 2 e 2t/T 2 1 r 2e 2if1 r ,t2 d r , (10)
S
evaluating the time derivative of M 1 r , t 2 as needed in (7). The
time derivative of a narrow-band signal is well approximated where the space- and time-varying phase is
S S
by a constant scaling factor d/dtM 1 r , t 2 < c0 M 1 r , t 2 . We t
S S
absorb this constant into the coil response pattern and rewrite f 1 r , t 2 ! 3 1 g Bz 1 r , tr 2 2 v 0 2 dtr. (11)
0
(7) as
In practice, multiple such signals are recorded, one for each exci-
S S S
v 1 t 2 5 reala 3 c 1 r 2 M 1 r , t 2 d r b. (8) tation/readout pair (“shots”). For simplicity of notation, we con-
sider “single shot” imaging; the extension to multiple shots is
The receive coil’s signal is amplified and demodulated using conceptually straightforward but notationally cumbersome. Note
some center frequency v 0 . Ideally, one would use v 0 5 g B0 if that the phase variations (11) are common to all receive coils;
S
the static magnetic field had uniform strength B0 . Usually only the coil response patterns 5 cl 1 r 2 6 differ between coils.
quadrature demodulation is used, yielding separate in-phase
I 1 t 2 and quadrature Q 1 t 2 baseband signals. In the literature, MEASUREMENT MODEL
the demodulated “MR signal” s 1 t 2 is defined (implicitly) as The recorded measurements in a MR scan consist of noisy samples
of the MR signal (10)
S S S
s 1 t 2 ! I 1 t 2 1 iQ 1 t 2 5 lowpass 1 eiv0 tv 1 t 22 5 eiv0 t3c 1 r 2 M 1 r , t 2 d r ,
(9) yli 5 sl 1 ti 2 1 eli, i 5 1, c, nd, l 5 1, c, L, (12)

where the low-pass operation selects the baseband component where yli denotes the ith sample of the lth coil’s signal at time
of the demodulated signal. This complex analog signal is just a ti and nd denotes the number of time samples. Usually the ti
mathematical definition; in practice, the I 1 t 2 and Q 1 t 2 signals values are equally spaced, and often there are one or more time
are each sampled and digitized yielding two digital signals. (One values where the signal is particularly strong due to alignment
can use two separate analog-to-digital (A/D) converters, or a sin- of the magnetization’s phases; these values are called echo
gle A/D converter running at twice the normal rate to avoid I/Q times. The measurement errors eli are very well modeled by

IEEE SIGNAL PROCESSING MAGAZINE [83] JULY 2010

additive, complex, zero-mean, MODEL-BASED METHODS THAT In practice the basis func-
temporally white Gaussian noise ACCOUNT FOR THOSE [PHYSICAL] tions are usually highly local-
[8]. However, there can be cou- EFFECTS ARE PROVING TO BE ized (e.g., voxels), so “center
pling of the noise values between of voxel” approximations like
BENEFICIAL FOR IMPROVING
different coils for the same time the following are nearly always
IMAGE QUALITY.
points, i.e., used, often implicitly
S * S S
Cov 5 eli, ekj 6 5 Slkd[i 2 j 4 , (13) alij < cl 1 r j 2 e 2ti /T 2 1 rj 2e 2if1 rj , ti2 . (17)

where d denotes the Kronecker impulse, and the L 3 L matrix For exceptions, see [10].
S characterizes the noise covariance between coils [7]. Typically the decay due to T *2 is ignored, or it is assumed
implicitly that the total readout time tnd 2 t1 is small relative
LINEAR RECONSTRUCTION PROBLEM to T *2 in which case one can make the approximation
* S * S
Using the measurement model (12) and the signal model (10), e 2ti /T 21 r 2 < e 2t1/T2 1 r 2 . Under this approximation, we can absorb
* S S
the “typical” image reconstruction problem in MRI is to esti- the T*2-weighting effect of e 2t1/T 2 1 r 2 into the unknown image f 1 r 2 .
S
mate the object f 1 r 2 from the measurement vector Combining (12) and (15) in matrix-vector form yields
y 5 1 y1, c, yL 2 , where yl 5 1 yl1, c, yl,nd 2 . (All vectors are
column vectors here.) We first consider model-based image yl 5 Al f 1 el,
reconstruction for this “basic” linear formulation. Because par-
allel imaging is of considerable interest, we continue to consider where f 5 1 f1, c, fN 2 is the vector of parameters (pixel
the general case of L receive coils. A standard single receive coil values) that we wish to estimate from the data y. Stacking up
is a simple special case. all L measurement vectors as y 5 1 y1, c, yL 2 and defining
This is an ill-posed problem because the given measure- the 1 nd L 2 3 N system matrix A 5 1 A1, c, AL 2 yields the
S
ments y are discrete whereas the object f 1 r 2 is an unknown linear model
continuous-space function. To facilitate parametric estimation,
S
we approximate the object f 1 r 2 using a “finite series expansion” y 5 Af 1 e. (18)
as follows:
N At first glance this linear model appears amenable to a vari-
S S S
f 1 r 2 5 a fj b 1 r 2 r j 2 , (14) ety of iterative solution methods. However, a significant chal-
j51
lenge that arises is that in general the elements of A can be
where b 1 # 2 denotes the object basis function, r j denotes the
S
quite complicated in the form above, yet A is too large to store
center of the jth translated basis function, and N is the for typical problem sizes. Most iterative algorithms require
number of parameters. Such approximations are classic in matrix-vector multiplication by A and its transpose; there
the tomographic image reconstruction literature [9] and are are fast algorithms for these operations (without storing A
slowly taking root in the MR community. Minimum L2 norm explicitly) in many special cases of interest [10], [11].
S
methods can postpone the discretization (14) until the final Thus far we have allowed the phase function f 1 r j, ti 2 to be
step of displaying the image, but it is unclear if this approach quite general, without the traditional focus on “Fourier
provides image quality benefits that outweigh its computa- encoding.” Recently there has been interest in investigating
S
tional requirements. For simplicity, hereafter we use rect nonlinear magnetic field variations Bz 1 r , t 2 in (4), and recon-
S S
basis functions b 1 r 2 5 rect 1 r /D 2 , i.e., square pixels of di- struction algorithms have been proposed that use much of the
mension D, so N is the number of pixels, or voxels in 3-D generality in (16) [12], [13]. These are currently specialized
scans. Many other possible basis function choices can be research topics, so we now focus on the more common case of
considered, all of which are imperfect because the true ob- linear field gradients.
ject never satisfies the parametric model (14) exactly.
Nevertheless simple basis functions can provide useful ap- FOURIER ENCODING
proximations. In typical MR scanners, the longitudinal component of the
S
Substituting the basis expansion (14) into the signal model applied field Bz 1 r , t 2 in (4) consists of three components
(10) and simplifying leads to the discrete forward model
S
Bz 1 r , t 2 5 B0 1 DB0 1 r 2 1 G 1 t 2 # r .
S S S
N (19)
sl 1 ti 2 5 a alij fj , (15)
j51
The constant B0 denotes the advertised field strength of the
S
where the elements 5 alij 6 of the system matrix Al associated main static field. The function DB0 1 r 2 denotes the spatial devi-
with the lth coil are given by ations of the field strength from this nominal value. This
function is often called a field map, and in general, it is
S S S * S S S
alij 5 3 b 1 r 2 r j 2 cl 1 r 2 e 2ti /T 2 1 r 2e 2if1 r, ti2 d r . (16) unknown, but it can be estimated by suitable types of

IEEE SIGNAL PROCESSING MAGAZINE [84] JULY 2010

acquisitions and data processing methods [14]. The field gradi- for various choices for the basis functions bik and coefficients ckj
S S S S S S S
ents G 1 t 2 5 Gx 1 t 2 j 1 Gy 1 t 2 j 1 Gz 1 t 2 j consist of three user- [11]. With this type of approximation, we can rewrite matrix-
controlled functions that are the historical key to providing vector multiplication as follows:
spatial information in standard MR imaging. Many different N
S S S
#S
types of MR scans are possible by changing G 1 t 2 . 3Al f 4 i < a bik a 1 ckj cl 1 r j 2 fj 2 e 2i2pk 1 ti2 r
j
.
k j51
Substituting (19) into (11) using v 0 ! gB0 and simplify-
ing yields The inner sum is simply a FFT or NUFFT so this approach is
t relatively fast. Free software for this is available [18].
S
f 1 r , t 2 5 3 g D B0 1 r 2 1 gG 1 t 2 # r dt
S S S

0
RECONSTRUCTION COST FUNCTION
or equivalently Having specified the linear model (18), we now turn to solution
S
methods. Because the noise in MRI measurements is Gaussian,
S
S
#S
e 2if1 r ,t 2 5 e 2iDv01 r 2t e 2i2pk 1t2 r
, (20) a natural approach is to estimate f by minimizing a regularized
least-squares cost function
S S
where Dv 0 1 r 2 ! g DB0 1 g 2 denotes the off-resonance frequen-
cy and the k-space trajectory is defined by f^ 5 arg min C 1 f 2 , C 1 f 2 ! || y 2 Af ||2 1 bR 1 f 2 . (24)
f
t
S 1 S
k 1t2 ! gG 1 t 2 dt. (21) For a single coil, the noise variance in the k-space data is
2p 30
white (uncorrelated with uniform variance), so the usual
Euclidian norm || # || is appropriate. For parallel MRI, noise is
S
Usually the phase accrual e 2iDv01 r 2 t due to off resonance is
undesirable and can distort reconstructed images if ignored. stationary across time samples 1 i 2 , but the norm should
Therefore some image reconstruction methods, particularly in include the inverse of the L 3 L covariance matrix S in
fMRI, account for its effects [10]. In some cases, the map (13) that describes the noise correlation between receive
S
Dv 0 1 r 2 is found from a separate “prescan,” in other cases it is coils [7].
estimated jointly with f [15]. In chemical shift imaging, e.g., If the k-space samples lie on an equally spaced grid
S
to separate fat and water components, the term Dv 0 1 r 2 (Cartesian sampling) with appropriate sample spacings relative
S
includes both useful information about the chemical shift to the object field of view, and if the rate map z 1 r 2 is zero (i.e.,
effect as well as the undesirable variations due to field inho- we ignore relaxation and field inhomogeneity), and if we consid-
mogeneity [16]. er just a single coil 1 L 5 1 2 and treat the sensitivity pattern as
S
For the linear field gradients (19), substituting (20) into (17) uniform, i.e., c1 1 r 2 5 1 , then the system matrix Al is orthogo-
yields simpler expressions for the system matrix nal. In this special case, no regularization is needed and
A 21 5 1/N Ar and the solution is simply f^ 5 1/N Ary, which
S S S
#S
alij < cl 1 r j 2 e 2z 1r j2 ti e2i2pk 1ti2 r
j
, (22) can be evaluated by an inverse FFT. This is the most common
MR image-reconstruction method. However, if any of these con-
S
where we define the “rate map” z 1 r 2 by combining the relax- ditions do not hold, then typically the system matrix A is not
ation and field maps well conditioned, and the unregularized LS solution can lead to
undesirable noise amplification. To avoid this problem, some
S S S
z 1 r 2 ! 1/T *2 1 r 2 1 iDv 0 1 r 2 . (23) form of regularization is needed.

When this rate map is assumed to be zero, i.e., if relaxation REGULARIZATION

and off resonance are ignored, then alij is the Sproduct of a An open problem in most image reconstruction problems,
#S
Fourier encoding matrix having elements e 2i2pk 1ti2 rj with a including MRI, is how to best choose the regularizer R 1 f 2 . If
S
diagonal sensitivity encoding matrix having elements cl 1 r j 2 . this term is not included, then the image estimate f^ will suffer
S
If the k-space sample locations k 1 ti 2 lie on an appropriate from noise and artifacts for under-sampled and/or non-Cartesian
subset of a Cartesian grid, then FFT operations provide effi- data, because this inverse problem is ill conditioned. The
cient multiplication by A and its transpose. If non-Cartesian approach for iterative reconstruction that has been adopted in
k-space sampling is used, then a nonuniform FFT (NUFFT) is commercial positron emission tomography scanners is to use
needed [17]. an unregularized algorithm, initialize it with a uniform image,
S
When z 1 r 2 in (22) is nonzero, then the elements (22) no lon- stop iterating just as the image gets unacceptably noisy, and
ger correspond to a standard Fourier transform. Approximations then perhaps apply a bit of post-filtering to reduce the noise.
are needed to provide fast computation of matrix-vector products. One could adopt a similar approach for MR imaging. However,
In particular, often one can approximate the exponentials in (17) introducing regularization can ensure that the iterative algo-
using an additively separable form rithm converges to a stable image and can enforce prior infor-
mation that improves image quality particularly for
S
e 2z1 r j 2ti < a bik ckj under-sampled data.
k

IEEE SIGNAL PROCESSING MAGAZINE [85] JULY 2010

 The simplest choice is Tikhonov regularization R 1 f 2 5 || f ||2 ALGORITHMS
or R 1 f 2 5 || f 2 f ||2 , where f is some prior or reference image Iterative algorithms are needed to minimize (24). For differen-
(possibly zero). The disadvantage of this choice is that it biases tiable regularizers such as (25), the conjugate gradient algo-
the estimate towards the reference image f . In particular, if the rithm is a natural choice [10]. For nondifferentiable regularizers
reference image is zero, then all pixel values in f^ are dimin- like (26), more sophisticated algorithms are needed and this is
ished towards zero, possibly reducing contrast. an active research area [20].
Another choice is a quadratic roughness penalty function,
which in one-dimensional (1-D) would be written RECONSTRUCTION CHALLENGES
N Although a variety of useful problems can be solved in MRI
R 1 f 2 5 a | fj 2 fj21 |2. (25) using the formulation (24), there are numerous challenges that
j52
provide research opportunities.
This choice biases the reconstruction towards a smooth image
where neighboring pixel values are similar. It is convenient for REGULARIZATION PARAMETER SELECTION
minimization [10], but it has the drawback of smoothing image A practical challenge with regularized methods is selection of
edges, particularly if the regularization parameter b in (24) is the regularization parameter b in (24). For quadratic regular-
too large. One can prove that ization, there is a well-devel-
using (25) guarantees that the oped theory for choosing b in
cost function (24) has a unique RECENTLY IT HAS BECOME VERY terms of the desired spatial res-
minimizer. POPULAR TO EXPRESS PRIOR olution properties of the recon-
More recently, total varia- INFORMATION IN TERMS OF SOME TYPE structed image [21]. This theory
tion methods have been OF SPARSITY OF THE OBJECT. extends readily to MR imaging
investigated for MR image with reasonably well sampled
reconstruction [19]. In 1-D, these methods replace the trajectories (and to parallel imaging with reasonable accelera-
squared differences between neighboring pixels above with tion factors) for which the point spread function (PSF) of the
absolute differences reconstructed image is relatively close to a Kronecker impulse
so that simple measures like full width at half maximum
N
(FWHM) are reasonable resolution metrics. For highly under-
R 1 f 2 5 a | fj 2 fj21 |. (26)
j52 sampled trajectories, the PSF can have “heavy tails” due to alias-
ing effects, and more investigation is needed to extend the above
In 2-D continuous space, the analogous functional is methods to MR applications.
For nonquadratic regularization such as the total variation
method (26), the analysis in [21] is inapplicable so one must
S ' S 2 ' S 2
3 00 =f 00 d r 5 33 ` f 1 r 2 ` 1 ` f 1 r 2 ` dx dy. resort to other methods for choosing b. Statisticians often use
Å 'x 'y
cross validation for choosing regularization parameters, with a
goal of finding the parameter that minimizes the mean-squared
The advantage of this type of regularization is that it biases the error (MSE) between f^ and the unknown f . However, MSE is
reconstructed image towards a piecewise smooth image, instead the sum of variance and bias squared, and where bias is related
of a globally smooth image, thereby better preserving image to spatial resolution and artifacts, and it is unclear whether an
edges. However it is harder to minimize and can lead to the equal weighting of noise variance and bias (squared) is optimal
appearance of “blocky” texture in images. Numerous alterna- from an image-quality perspective in medical imaging.
tives of the form Another method for choosing b is the “L-curve” method.
This method is expensive because it requires evaluating f^ for
N
several values of b, and it has some theoretical deficiencies [22].
R 1 f 2 5 a c 1 fj 2 fj21 2
j52 In summary, choosing b for nonquadratic regularization
remains a nontrivial issue in most ill-posed imaging problems
for various choices of the “potential function” c 1 # 2 have been including MRI, and remains an active research area [23].
proposed in the imaging literature. Many of these compromise
between the quadratic case (25) and the absolute difference case PARTIAL K-SPACE METHODS
S
(26), for example the hyperbola If the object f 1 r 2 were real, then its Fourier transform would
be Hermitian symmetric so in principle only half of k-space
c 1 t 2 5 "1 1 | t/d |2 2 1 (27) would need to be sampled. In practice, the magnetization (3) is
complex due to a variety of physical effects. However, in many
S
is approximately quadratic near zero, which aids noise reduc- cases the phase of M 1 r , t 2 can be assumed to be a smooth func-
tion, yet approximately linear away from zero, which helps tion. This property has led to a variety of partial k-space meth-
preserve edges. ods where one samples a bit more than half of k-space, then

IEEE SIGNAL PROCESSING MAGAZINE [86] JULY 2010

estimates the phase from the central portion of k-space this approach include choosing the sparsifying transform C and
(corresponding to low spatial frequencies), and then uses this regularization parameters b and d appropriately. Furthermore,
estimated phase to reconstruct the entire image [24]. Such when d is small, the regularizer (28) has very high curvature
methods are used routinely in many types of MR scans. near zero, which can slow convergence.

UNDER-SAMPLED K-SPACE DATA NONLINEAR RECONSTRUCTION PROBLEMS

The need for some type of regularization is essential when the The linear image reconstruction problem (24) is just one of
k-space data is under sampled, i.e., when the number of mea- many estimation problems of interest in MRI. Returning to the
surements Lnd is less than the number of unknown voxels N . elements of the system matrix (22), there has been research on
In MRI, the scan time is roughly proportional to the number of estimating essentially every component therein, as summa-
measurements, so collecting fewer samples can reduce scan rized below.
time, which is particularly desirable in dynamic imaging.
In the broader field of tomographic image reconstruction, FIELD MAP ESTIMATION
there is a long history of using prior information, such as For scans with long readout times, the effect of field inhomo-
assuming objects are piecewise smooth, to reconstruct images geneity Dv 0 in (22) is important. In practice, the field map
S
from an under-sampled set of v 1 r 2 is not known a priori but
projection views, e.g., [25]. rather it must be estimated
Many of these methods involve
THE LINEAR IMAGE RECONSTRUCTION from noisy MR scans. One can
cost functions of the form (24) PROBLEM IS JUST ONE OF MANY examine the phase differences
with a suitable system matrix ESTIMATION PROBLEMS OF between two scans having dif-
A for the application and INTEREST IN MRI. ferent echo times to determine
appropriate regularizers R 1 f 2 Dv 0 . If these two scans have
that capture prior information about the object. short readouts, then there are simple image-domain methods
Recently it has become very popular to express prior infor- for estimating Dv 0 , which is known as B0 field mapping [14].
mation in terms of some type of sparsity of the object. Notions Errors in the field map estimates may cause artifacts in recon-
of sparsity have deep roots in statistical signal processing [26]. structed images that are based on models like (22).
Sparsity is especially apparent in MR angiography. The moniker In addition, object motion that occurs between the field
of compressed sensing or compressive sampling has become map scans and subsequent scans of interest, e.g., in fMRI, will
widespread for such techniques, and recently entire sessions at lead to an inconsistency between the actual scan data and the
MR conferences have been devoted to this topic [20]. Some assumed model (22) used by the reconstruction algorithm.
compressed sensing formulations ignore the noise in the data. This possibility has motivated the development of dynamic
In the presence of noise, a typical formulation is field mapping methods that estimate the field map separately
for each frame in a dynamic study, e.g., [15]. For scans with
arg min || Cf || 1 s.t. || y 2 Af ||2 # P , long readout durations, the appearance of Dv 0 in a complex
f
exponential in (22) makes this a somewhat complicated non-
where C transforms the image f into a domain (such as wavelet linear estimation problem.
coefficients) where one postulates that the signal is sparse.
Often this optimization problem is solved using a Lagrange RELAXATION MAP ESTIMATION
multiplier approach In some MR applications, it is useful to estimate tissue relax-
ation parameters, particularly T2 or T *2 , on a pixel-by-pixel
arg min 7 y 2 Af 7 22 1 b 7 Cf 7 1, basis. One approach to measuring such relaxation parameters
f
is to acquire a “baseline” scan of the object and then acquire
which corresponds to a particular regularizer in (24). Rarely is one or more additional scans having different echo times. One
the ,1 norm implemented exactly; in practice usually a continu- then reconstructs images from each of those scans and then
ously differentiable approximation is used, such as performs linear regression on a voxel-by-voxel basis using the
logarithm of the image voxel values. This approach can be ade-
00 v 00 1 < a 1 "| vi |2 1 d2 2 d 2 (28) quate if the readout durations are sufficiently small. But for
i
acquisitions with long readouts, the effect of time ti in the
S
for some small value of d . 0. This approximation is equivalent e 2z1 rj 2ti in (22) should be considered, i.e., we should account
to the hyperbola (27) used frequently for edge-preserving image for relaxation during the signal readout. This requires meth-
reconstruction. Nonconvex methods that enforce sparsity even ods that estimate the relaxation map directly from the k-space
more strongly are also under investigation. In the usual case data. These are more challenging nonlinear estimation prob-
where A corresponds to an under-sampled discrete Fourier lems because T *2 appears in an exponent in (22). Several meth-
S
transform (DFT), a variety of algorithms are available that have ods for jointly estimating T *2 , Dv 0 , and f 1 r 2 have been
numerous potential applications in MR [20]. Challenges with investigated [27].

IEEE SIGNAL PROCESSING MAGAZINE [87] JULY 2010

SENSITIVITY MAP ESTIMATION functional magnetic resonance imaging (fMRI) based on the
S
The coil sensitivity patterns cl 1 r 2 in (22) also must be deter- BOLD effect [29]. Accurate reconstruction of signals near air-
mined for parallel imaging based on sensitivity encoding. tissue interfaces requires compensation for these within-voxel
Normally this is done by acquiring well-sampled data both gradients, which complicates the reconstruction method [30].
with local receive coils and with a reference body coil and
dividing the two [7]. Acquiring the extra reference data can EXAMPLE
be inconvenient, so normalizing by the square root of the To illustrate the capabilities of model-based image reconstruc-
sum of squares of the local receive coils is also used. A variety tion methods for MRI, we simulated k-space data for a four-
of other estimation methods have been proposed, including shot EPI sequence with matrix size 128 × 128 and 5 ms
S
methods that jointly estimate the sensitivity maps 5 cl 1 r 2 6 sampling so the readout duration was 27.3 ms per shot. The
S S
and the image f 1 r 2 [28]. Note field map D B0 1 r 2 appears in
S
that if f 1 r 2 were known, then Figure 2 of [14] and is based
DESPITE OVER THREE DECADES
the problem of estimating on a brain slice above the
S OF MR RESEARCH, THERE REMAIN
cl 1 r 2 would be a linear esti- sinuses and ear canals where
mation problem because cl 1 r 2
S CHALLENGING AND INTRIGUING susceptibility effects occur.
appears as a linear scaling in PROBLEMS IN MR IMAGE Figure 1 shows the true image
S
(22). But when both f 1 r 2 and RECONSTRUCTION. used in the simulations and
S
cl 1 r 2 are to be estimated, the images from three different
S S
model is bilinear because f 1 r 2 and cl 1 r 2 appear as a product reconstruction methods. The “uncorrected” reconstruction
in (10). This complicates joint estimation. simply uses an inverse 2-D FFT, with no consideration of field
inhomogeneity. The field inhomogeneity causes spatial distor-
TRAJECTORY MAPPING tion in the read-out (vertical) direction (that increases NRMSE
S
The k-space trajectory k 1 ti 2 , defined as an integral of the gradi- dramatically), as well as significant intensity artifacts above the
ent waveforms in (21), should be calibrated carefully to ensure ears and sinuses where the susceptibility effects are largest.
that the system model (22) is accurate. In practice, the field The classical conjugate phase reconstruction method, which
gradients induced by the gradient coils in the scanner are not corresponds to Ary in this single-coil case, reduces the spatial
exactly proportional to the waveforms applied to those coils due distortion but the intensity artifacts persist. Applying 15 itera-
to eddy currents. Therefore the physical k-space trajectory real- tions of a conjugate gradient algorithm with a monotonic line
ized in the system can depart somewhat from the desired k- search [11] to the cost function (24) with the edge-preserving
space trajectory. These differences can degrade the hyperbola (27) yields the right-most image in Figure 1. This
reconstructed image, particularly for non-Cartesian trajectories model-based image reconstruction method yields the lowest
with long readout durations. Therefore, a variety of techniques RMS error, but it requires about 30 times more computation
have been developed for mapping the actual k-space trajectory than the noniterative conjugate phase method [11] because
experimentally. each iteration requires multiplication by A and Ar. The soft-
ware that generated this figure is available online [18].
WITHIN-VOXEL GRADIENTS
The model (23) treats the field inhomogeneity within each voxel SUMMARY
as being a constant, ignoring within-voxel gradients of the off- Image reconstruction is not a single problem in MRI but rath-
resonance map. However, these gradients can be significant in er is a wide family of problems depending on what physical

True Uncorrected Conjugate Phase MBIR 8

NRMSE = 143% NRMSE = 13% NRMSE = 5% 0

[FIG1] Comparison of model-based image reconstruction with convention methods.

IEEE SIGNAL PROCESSING MAGAZINE [88] JULY 2010

effects are included in the signal model. The most widely stud- [6] P. B. Roemer, W. A. Edelstein, C. E. Hayes, S. P. Souza, and O. M. Mueller,
“The NMR phased array,” Magn. Reson. Med., vol. 16, no. 2, pp. 192–225, Nov.
ied case, particularly in the signal processing community, is 1990.
when nearly all physical effects are disregarded and the system [7] K. P. Pruessmann, M. Weiger, M. B. Scheidegger, and P. Boe-siger, “SENSE:
Sensitivity encoding for fast MRI,” Magn. Reson. Med., vol. 42, no. 5, pp. 952–962,
model consists solely of sampled of the Fourier transform of Nov. 1999.
the object. This basic model is amenable to familiar signal pro- [8] A. Macovski, “Noise in MRI,” Magn. Reson. Med., vol. 36, no. 3, pp. 494–497,
cessing tools and is applicable to many MR scans. But there Sept. 1996.
are also many interesting applications where other physical [9] Y. Censor, “Finite series expansion reconstruction methods,” Proc. IEEE,
vol. 71, no. 3, pp. 409–419, Mar. 1983.
effects are relevant, and model-based methods that account for
[10] B. P. Sutton, D. C. Noll, and J. A. Fessler, “Fast, iterative image reconstruction
those effects are proving to be beneficial for improving image for MRI in the presence of field inhomogeneities,” IEEE Trans. Med. Imag., vol. 22,
no. 2, pp. 178–188, Feb. 2003.
quality. Model-based methods themselves depend on estimates
[11] J. A. Fessler, S. Lee, V. T. Olafsson, H. R. Shi, and D. C. Noll, “Toeplitz-based
of a variety of model parameters, leading to interesting prob- iterative image reconstruction for MRI with correction for magnetic field inho-
lems where those parameters are determined either by sepa- mogeneity,” IEEE Trans. Signal Processing, vol. 53, no. 9, pp. 3393–3402, Sept.
2005.
rate calibration scans or by jointly estimating the image and [12] J. Hennig, A. M. Welz, G. Schultz, J. Korvink, Z. Liu, O. Speck, and M. Zai-
those parameters. Despite over three decades of MR research, tsev, “Parallel imaging in non-bijective, curvilinear magnetic field gradients: A
concept study,” Magn. Reson. Mater. Biol. Phys. Med., vol. 21, no. 1–2, pp. 5–14,
there remain challenging and intriguing problems in MR Mar. 2008.
image reconstruction. [13] J. Stockmann, P. Ciris, and R. T. Constable, “Efficient “O-space” parallel
imaging with higher-order encoding gradients and no phase encoding,” in Proc.
Int. Soc. Magnetic Resonance in Medicine, 2009, p. 761.
ACKNOWLEDGMENT [14] A. K. Funai, J. A. Fessler, D. T. B. Yeo, V. T. Olafsson, and D. C. Noll, “Regular-
This work was supported in part by NIH grants CA87634 and ized field map estimation in MRI,” IEEE Trans. Med. Imag., vol. 27, no. 10, pp.
1484–1494, Oct. 2008.
NS058576.
[15] B. P. Sutton, D. C. Noll, and J. A. Fessler, “Dynamic field map estimation using
a spiral-in/spiral-out acquisition,” Magn. Reson. Med., vol. 51, no. 6, pp. 1194–
1204, June 2004.
AUTHOR
[16] W. Dixon, “Simple proton spectroscopic imaging,” Radiology, vol. 153, no. 1,
Jeffrey A. Fessler ([email protected]) received the B.S.E.E. pp. 189–194, Oct. 1984.
degree from Purdue University in 1985, the M.S.E.E. degree [17] J. A. Fessler and B. P. Sutton, “Nonuniform fast Fourier transforms using min-
from Stanford University in 1986, and the M.S. degree in sta- max interpolation,” IEEE Trans. Signal Processing, vol. 51, no. 2, pp. 560–574,
Feb. 2003.
tistics from Stanford University in 1989. From 1985 to 1988,
[18] J. A. Fessler. MATLAB tomography toolbox, 2004 [Online]. Available:
he was a National Science Foundation Graduate Fellow at http://www.eecs.umich.edu/~fessler
Stanford, where he earned a Ph.D. degree in electrical engi- [19] K. T. Block, M. Uecker, and J. Frahm, “Undersampled radial MRI with mul-
tiple coils. Iterative image reconstruction using a total variation constraint,” Magn.
neering in 1990. He has worked at the University of Michigan Reson. Med., vol. 57, no. 6, pp. 1086–1098, June 2007.
since thenwhere he is now a professor in the Departments of [20] M. Lustig, D. L. Donoho, J. M. Santos, and J. M. Pauly, “Compressed sensing
Electrical Engineering and Computer Science, Radiology, and MRI,” IEEE Signal Processing Mag., vol. 25, no. 2, pp. 72–82, Mar. 2008.

Biomedical Engineering. He is a Fellow of the IEEE. He [21] J. A. Fessler and W. L. Rogers, “Spatial resolution properties of penalized-like-
lihood image reconstruction methods: Space-invariant tomographs,” IEEE Trans.
received the Francois Erbsmann Award. He is an associate edi- Image Processing, vol. 5, no. 9, pp. 1346–1358, Sept. 1996.
tor for IEEE Transactions on Medical Imaging and was an [22] C. R. Vogel, “Non-convergence of the L-curve regularization parameter selec-
tion method,” Inverse Probl., vol. 12, no. 4, pp. 535–547, Aug. 1996.
associate editor for IEEE Transactions on Image Processing
[23] S. Ahn and R. M. Leahy, “Analysis of resolution and noise properties of non-
and IEEE Signal Processing Letters. He was cochair of the quadratically regularized image reconstruction methods for PET,” IEEE Trans.
Med. Imag., vol. 27, no. 3, pp. 413–424, Mar. 2008.
1997 SPIE Conference on Image Reconstruction and
[24] D. C. Noll, D. G. Nishimura, and A. Macovski, “Homodyne detection in mag-
Restoration, technical program cochair of the 2002 IEEE netic resonance imaging,” IEEE Trans. Med. Imag., vol. 10, no. 2, pp. 154–163,
International Symposium on Biomedical Imaging (ISBI), and June 1991.
general chair of ISBI 2007. His research interests are in statis- [25] A. H. Delaney and Y. Bresler, “Globally convergent edge-preserving regular-
ized reconstruction: An application to limited-angle tomography,” IEEE Trans. Im-
tical aspects of imaging problems. age Processing, vol. 7, no. 2, pp. 204–221, Feb. 1998.
[26] G. Harikumar and Y. Bresler, “A new algorithm for computing sparse solu-
tions to linear inverse problems,” in Proc. IEEE Conf. Acoustics, Speech, and
REFERENCES Signal Processing, 1996, vol. 3, pp. 1331–1334.
[1] M. Bydder, A. A. Samsonov, and J. Du, “Evaluation of optimal density weight-
ing for regridding,” Magn. Reson. Imag., vol. 25, no. 5, pp. 695–702, June [27] V. T. Olafsson, D. C. Noll, and J. A. Fessler, “Fast joint reconstruction of dy-
2007. namic R2* and field maps in functional MRI,” IEEE Trans. Med. Imag., vol. 27, no.
9, pp. 1177–1188, Sept. 2008.
[2] G. A. Wright, “Magnetic resonance imaging,” IEEE Signal Processing Mag.,
vol. 14, no. 1, pp. 56–66, Jan. 1997. [28] L. Ying and J. Sheng, “Joint image reconstruction and sensitivity estimation in
SENSE (JSENSE),” Magn. Reson. Med., vol. 57, no. 6, pp. 1196–1202, June 2007.
[3] Z.-P. Liang and P. C. Lauterber, Principles of Magnetic Resonance Imaging.
New York: IEEE Press, 2000. [29] D. C. Noll, J. A. Fessler, and B. P. Sutton, “Conjugate phase MRI reconstruc-
tion with spatially variant sample density correction,” IEEE Trans. Med. Imag., vol.
[4] W. A. Grissom, D. Xu, A. B. Kerr, J. A. Fessler, and D. C. Noll, “Fast large-tip- 24, no. 3, pp. 325–336, Mar. 2005.
angle multidimensional and parallel RF pulse design in MRI,” IEEE Trans. Med.
Imag., vol. 28, no. 10, pp. 1548–1559, Oct. 2009. [30] J. A. Fessler and D. C. Noll, “Model-based MR image reconstruction with
compensation for through-plane field inhomogeneity,” in Proc. IEEE Int. Symp.
[5] P. Mansfield and I. L. Pykett, “Biological and medical imaging by NMR,” Biomedical Imaging, 2007, pp. 920–923.
J. Magn. Reson., vol. 29, no. 2, pp. 355–373, Feb. 1978. [SP]

IEEE SIGNAL PROCESSING MAGAZINE [89] JULY 2010