IEEE TRANSACTIONS ON MEDICAL IMAGING, VOL. 34, NO.

2, FEBRUARY 2015

531

Simultaneous Phase Unwrapping and Removal

of Chemical Shift (SPURS) Using Graph Cuts:
Application in Quantitative Susceptibility Mapping
Jianwu Dong, Tian Liu*, Feng Chen*, Dong Zhou, Alexey Dimov, Ashish Raj, Qiang Cheng,
Pascal Spincemaille, and Yi Wang

AbstractQuantitative susceptibility mapping (QSM) is a magnetic resonance imaging technique that reveals tissue magnetic
susceptibility. It relies on having a high quality field map, typically acquired with a relatively long echo spacing and long final
TE. Applications of QSM outside the brain require the removal
of fat contributions to the total signal phase. However, current
water/fat separation methods applied on typical data acquired for
QSM suffer from three issues: inadequacy when using large echo
spacing, over-smoothing of the field maps and high computational
cost. In this paper, the general phase wrap and chemical shift
problem is formulated using a single species fitting and is solved
using graph cuts with conditional jump moves. This method is
referred as simultaneous phase unwrapping and removal of chemical shift (SPURS). The result from SPURS is then used as the
initial guess for a voxel-wise iterative decomposition of water and
fat with echo asymmetric and least-squares estimation (IDEAL).
The estimated 3-D field maps are used to compute QSM in body
regions outside of the brain, such as the liver. Experimental results
show substantial improvements in field map estimation, water/fat
separation and reconstructed QSM compared to two existing
water/fat separation methods on 1.5T and 3T magnetic resonance
human data with long echo spacing and rapid field map variation.
Index TermsField map estimation, magnetic resonance
imaging (MRI), quantitative susceptibility mapping, water/fat
separation.
Manuscript received August 21, 2014; revised September 25, 2014; accepted
September 29, 2014. Date of publication October 08, 2014; date of current version January 30, 2015. This work was supported in part by the National Natural
Science Foundation of China under Grant 61271388 and Grant 61327902, in
part by the Beijing Natural Science Foundation under Grant 4122040, in part by
the Research Project of Tsinghua University under Grant 2012Z01011, in part by
the Specialized Research Fund for the Doctoral Program of Higher Education,
and in part by the U.S. National Institute of Health under Grant R43EB015293,
R01EB013443, and Grant R01CA178007. Asterisk indicates corresponding author.
J. Dong and Q. Cheng are with the Department of Automation, Tsinghua University, Beijing 100084, China.
*T. Liu is with Medimagemetric LLC, New York, NY 10044 USA (e-mail:
tian.liu@medimagemetric.com).
*F. Chen is with the Department of Automation, Tsinghua University, Beijing
100084, China (e-mail: chenfeng@mail.tsinghua.edu.cn).
D. Zhou, A. Raj, and P. Spincemaille are with the Department of Radiology,
Weill Cornell Medical College, New York, NY 10021 USA.
A. Dimov is with the Department of Radiology, Weill Cornell Medical College, New York, NY 10021 USA, and also with the Department of Biomedical
Engineering, Cornell University, Ithaca, NY 14853 USA.
Y. Wang is with the Department of Radiology, Weill Cornell Medical College, New York, NY 10021 USA, and also with the Department of Biomedical
Engineering, Cornell University, Ithaca, NY 14853 USA, and also with the Department of Biomedical Engineering, Kyung Hee University, Seoul 130-701,
Korea.
Digital Object Identifier 10.1109/TMI.2014.2361764

I. INTRODUCTION

UANTITATIVE susceptibility mapping (QSM) is a

magnetic resonance imaging technique that reveals
tissue magnetic susceptibility by deconvolving a 3-D
field map with a dipole kernel [1][13]. Because the quality of
QSM hinges upon the quality of the field map, a multi-echo
readout with a long echo spacing (
ms) and a long last TE
(
ms at 3T) has generally been adopted to accrue sufficient
phase signal. The multi-echo signal is then fitted with a single
species model to estimate the field map. This model is adequate
for typical QSM applications in the brain, as the contribution
of fat to the MRI signal is minimal. In order to apply QSM to
other body regions that contain fat, water/fat separation with a
nonlinear signal model is usually needed.
The problem of field map estimation in the presence of both
water and fat is commonly solved as an energy
minimization problem, with unknowns including a spatially varying
water component
, a fat component
with a chemical shift
of
, a susceptibility induced field inhomogeneity , and a
signal decay rate
necessary when TE is long [14][23]

(1)
where
is the signal acquired at echo time
, and
is the number of echoes. Here, all the unknowns are scalar
fields with the spatial index omitted. There are multiple unknowns for each voxel: two complex-valued unknowns
and
and two real unknowns
and , so this is a high-dimensional optimization problem. One solution is a gradient descent
based method, such as
-Iterative Decomposition of water
and fat with Echo Asymmetric and Least-squares estimation
( -IDEAL) proposed in [14]. However, this method is susceptible to finding local minima close to the initial guess instead
of identifying the global minimum [17]. Fundamentally, current
solutions to (1) suffer from the three issues: inadequacy when
using large echo spacing, over-smoothing of the field maps, and
high computational cost.
1) Large Echo Spacing Challenge: It has been demonstrated that the optimal echo spacing
is
for minimizing noise propagation [24]. For a
minimum
at 3T, the longest optimal echo spacing

0278-0062 2014 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.

532

is 0.75 ms, which would require either a very high receiver

bandwidth consequently leading to poor SNR, or multiple
scans that increase total acquisition time. In most existing QSM
protocols, multiple TEs are sampled within the same repetition
time (TR) using an echo spacing of around 5 ms. This long echo
spacing poses a challenge to field map estimation: the phase
accrual in one echo spacing from both the chemical shift and
the susceptibility inhomogeneity field may exceed , which
is beyond the range where signal phase can unambiguously
represent the magnetic field [25].
2) Over-Smoothing: In current methods to enforce the continuity of the field map [20], [26], [27], a regularization term
is typically added to the data fitting term to penalize abrupt
changes in the field map [17], [20], [28]. This regularization
formulation gives a maximum a posteriori estimation and has
been successfully applied in numerous applications in medical
imaging. However, the choice of regularization parameter is still
an open question and is usually determined heuristically, and
the regularization may introduce over-smoothing on the field
map that causes errors in the subsequent QSM [29]. Another approach is to formulate the continuity condition as a combinatorial optimization problem. Region growing (RG) based methods
have been proposed to efficiently connect the local minima identified from VARiable PROjection (VARPRO), as in [25]. However, RG is a greedy algorithm that does not guarantee a global
optimum and its result is dependent on the RG path, i.e., an error
in a single step may propagate to the whole image. Loopy belief propagation (LBP) algorithm with additional swap moves
[27] has also been suggested for solving this combinatorial optimization problem. However, field maps obtained with LBP are
prone to having piecewise smooth segments with abrupt discontinuities. This may be caused by the fact that LBP is suited for a
tree-structure Markov random field (MRF) but not a grid-structure MRF [27].
3) Computational Cost: Current attempts to find the
minimum in a global sense start with candidate generation.
VARPRO is widely used in various methods [16], [17], [20],
[25], [27], [28], [30]. VARPRO uniformly discretizes field map
values in a given range, scans through all possible values and
identifies the local minima as candidates. This step is time consuming for a 3-D dataset, especially when a fine discretization
is used for accuracy or when
is also subject to optimization. The candidate solutions may also be found through a
golden-section search, as implemented in Multi-Resolution
field map estimation using Golden Section search method
(MRGS) [25], but this method does not account for the
term
in (1) that is necessary for long echo data in QSM.
We propose here to formulate the removal of the echo-spacing
induced phase discontinuities and fat chemical shifts as a
discrete problem using an approximate signal model for multiple species (fat and water). We demonstrate that the discrete
problem can be efficiently and effectively solved using graph
cuts with conditional jump moves. This Simultaneous Phase
Unwrapping and Removal of chemical Shift (SPURS) method
can provide a faithful initial guess without any smoothing for
IDEAL to obtain the water and fat maps [14]. In this work,
we show that the SPURS method outperforms two existing
methods on several 1.5T and 3T human datasets with long echo

IEEE TRANSACTIONS ON MEDICAL IMAGING, VOL. 34, NO. 2, FEBRUARY 2015

Fig. 1. Flowchart of the SPURS method.

spacing and rapid field map variation, and we show promising

QSM results in liver imaging.
II. METHODS
It has been shown in signal spectral analysis [31] that if a
single species model is used to fit a multiple species signal,
the fitted frequency is close to the frequency of the dominant
species. Thus if a single species fit is used for the signal
,
the fitted frequency is close to
if the dominant species is
water, or to
if the dominant species is fat, allowing
the determination of the dominant species when there are multiple species present in a single voxel [31]. Therefore, a fitted
frequency map has discontinuities due to 1) limited dynamic
range dictated by echo spacing (phase wraps) and 2) fat chemical shift. Characterizing the spatial continuity of the actual field
map by the
norm of its gradient, the discontinuities can be
addressed by solving the following combinatorial optimization
problem over integer scalar fields and a Boolean scalar fields

(2)
where
denotes the gradient operator,
corrects for
the field map wrap due to limited dynamic range with
, and
corrects for the fat chemical shift in each
voxel with
. This joint optimization over and
can be efficiently solved using graph cuts with conditional jump
moves.
Fig. 1 illustrates the flowchart of the SPURS method in four
steps: 1) formulate phase wrap and chemical shift problem
using a single species fitting to obtain an approximate field map
estimate; 2) use graph cuts to solve the joint phase unwrapping
and fat chemical shift removal problem; 3) refine the fitted frequency by performing a voxel-wise -IDEAL; 4) reconstruct
QSM using a modified version of the Morphology Enabled
Dipole Inversion method (MEDI) [6].
A. Formulation of Phase Wrap and Chemical Shift Problem
Using Single Species Fitting
A vector representation of the signal is shown in Fig. 2. If
, the complex signal generated by the two species

DONG et al.: SPURS USING GRAPH CUTS: APPLICATION IN QUANTITATIVE SUSCEPTIBILITY MAPPING

533

if
or
. That is, if the voxels consist of only water
or only fat, the model error of the single species fitting is zero.
The true model error introduced in the single species fitting
will calculated numerically in the simulation section below.
B. Simultaneous Phase Unwrapping and Removal of Chemical
Shift via Graph Cuts With Conditional Jump Moves

Fig. 2. Vector representation of the signal consisting of both water and fat.
Geometric relationship between these vectors is used to provide an intuition of
the error introduced by the single species fitting.

can be formulated as
is

the

magnitude

, where
of

and

is the phase of
.
We use a single species model (frequency ) to fit the two
species signal by minimizing the following nonlinear least
squares fit [6]:

Equation (2) is solved in this step using a graph cuts algorithm. In the following, the field map is changed from Hz to
radian by multiplying
for notational convenience. It is
assumed that the field map
is continuous after
summing
, where
is the effective fat chemical
shift derived by wrapping
to the
interval. Let
, we have
(7)
in the folWithout loss of generality, we assume
lowing and denote
.
A discrete energy function is generalized from (2)

(8)
(3)

, subscript denotes a spatial location,

where
represents any of the three Cartesian coordinate axes,
is
a
weighting
that
is
equal
to
the
sum
of
the
magnitude
value
Equation (3) is solved using the Gauss-Newton method. To
is the finite
analyze model errors introduced by the single species fit, an of the voxel and its neighbor along , and
difference
operator
along
.
The
magnitude
weighting
term
approximate analytical solution is obtained (detailed derivation
introduced
in
the
energy
function
serves
as
a
quality
map.
in Appendix A)
Smaller weights are given to voxels with weaker signal in
order to enhance the robustness against phase noise [32]. The
smaller weights also allow discontinuity in the field map where
the corresponding magnitude has low SNR [33]. This energy
function is defined on a 3-D first order MRF, in which every
(4) node is locally dependent on its six neighboring nodes. Miniwhere
mizing the above energy function is a multi-label optimization
. problem, which can be approximately solved by changing it
Therefore, the true field map,
, can be written as the to a sequence of binary optimization problem in which each
summation of the estimated field map, , the error term due to binary problem is solved by graph cuts [34][36].
limited dynamic range, the error term due to chemical shift,
We start with an initial labeling in which each voxel is asand the model error
signed label
. In each iteration, every voxel's label
can
jump to
or remain unchanged, where is the step size and
(5) this jump move is called an -jump move. Every jump move iteration decreases or preserves the function energy. In this study,
where
and model error
is approxwas set equal to 2 (called the 2-jump move) until no further
imately
decrease in energy was observed. After the energy ceased to decrease using
, the iterative process was repeated with
(called the 1-jump move). Fig. 3 shows an example that illustrates the evolution of labels in the jump move optimization.
For 2-jump move optimization, the jump lengths are uniformly
2 for all voxels, so it is equivalent to traditional multi(6)
label optimization problem with integer labels. However, the
Note that larger echo spacing and more echoes tend to lead construction of the binary energy term for 1-jump move optito a smaller approximate model error because the numerator is mization requires a conditional jump move strategy due to the
Boolean variable , in which a voxel's jump length depends on
a summation of and the denominator is a summation of
its current label. We use superscripts to denote the number of
multiplied by echo spacing. In addition,

534

IEEE TRANSACTIONS ON MEDICAL IMAGING, VOL. 34, NO. 2, FEBRUARY 2015

minimization problem of the binary energy function and the

min-cut/max-flow problem was proven in [37].
C. Voxel-Wise

-IDEAL Water/Fat Separation

After the approximate field map is obtained in step 2, it

is taken as the initial guess for the field inhomogeneity in
-IDEAL, the objective function of which is identical to (1).
This step is a fine tuning step to remove the model error in the
field due to the single species assumption in the previous step
and to obtain the water and fat maps. Note that no additional
spatial smoothing is performed, as conventionally done in
-IDEAL.
Fig. 3. Example illustrating the evolution of voxel labels in the jump move
optimizations. (a) 2-jump move optimization and (b) 1-jump move optimization.
Bold labels indicate voxels that perform jump moves.

iterations, and subscripts to denote the spatial location. For example,

denotes the label of voxel at iteration , and

D. QSM Reconstruction
After the field map is fine-tuned by -IDEAL, it is taken
as the input for QSM reconstruction using a modified version
of MEDI [6] that incorporates background field removal in the
susceptibility estimation [38][40]

(9)
.
is the jump length that depends on
where
the value of
. If
, the possible values of
are
, which can jump to
, respectively with step size 1, and the jump length
is
. If
, the possible values of
are
,
which can jump to
, respectively with step size 1, and
the jump length
is . So we have
(10)
In the
th iteration, the label
can make a binary decision
to either jump to
or remain unchanged
. We use graph cuts to optimize the binary variables
. For every pair of neighboring voxels and , we aim to
minimize their difference in the following:

(13)
where is a regularization parameter empirically determined to
be 1000,
is a data weighting, is a matrix representing the
dipole convolution, and
is a binary weighting term derived
from the anatomical image, with 30% of its voxels equal to zero
to allow changes in
and the rest equal to one [41]. This energy minimization has a data fidelity term that is formulated in
the complex plane to account for Gaussian noise and an norm
regularization term to encode the prior. In contrast to the original nonlinear MEDI, this energy minimization has an additional
Laplacian in the data fidelity term. Since the background field is
a harmonic function, the Laplacian eliminates any background
fields. The Laplacian is implemented as a convolution with a
kernel equal to the Kronecker delta function minus a sphere with
radius 5 mm and unit integral. For efficiency, this convolution
is evaluated in Fourier space.

(11)

III. MRI EXPERIMENTS

which is the binary energy term for graph cuts. For every binary
energy term
, the following inequality exists:

A. Single Species Fitting Error Simulations

(12)
i.e., the energy function satisfies the submodularity condition
and can be efficiently solved by standard graph cuts method
[37]. The proof of the submodularity condition is provided in
Appendix B.
After
the
following
binary
energy
function
is obtained, a weighted
with source node and sink node
directed graph
is constructed by following Kolmogorov and Zabih's paper
[37]. The edge weights of
are assigned according
to the energy function. Afterwards, the original energy
minimization problem is changed to a min-cut/max-flow
problem in
, which has been studied extensively
and can be solved efficiently. The equivalence between the

Numerical phantom simulations were performed to compute

the model error
(6). Phantoms with different
percentages of water and fat were generated. The fat fraction
was varied between 0 and 1 with a uniform
spacing 0.01.
was varied between 2 ms and 10 ms with
a uniform spacing 0.1 ms. The true model error was computed
as
, where
was the true field
map and
was the result of SPURS. It
was assumed in the simulation that the main magnetic field was
3T,
and the number of echoes was four, which
was the same as used in our in vivo scans.
B. Numerical Phantom Simulations
A numerical phantom with a mixture of water and fat was
constructed to test the proposed method. The matrix size was
128 128 64. The single peak fat model with fat chemical

DONG et al.: SPURS USING GRAPH CUTS: APPLICATION IN QUANTITATIVE SUSCEPTIBILITY MAPPING

shift
ppm [42] was used. We assumed that the fat susceptibility was 0.6 ppm with respect to water. The numerical
phantom is a water phantom where most part is made up of
water. In the center, there were six cylinders and two cuboids
of different size, where the cylinders and cuboids had different
fat fractions. The fat fractions in the cylinders and cuboids were
23%, 35%, 47%, 60%, 73%, 86%, 94%, and 100%, respectively. The main magnetic field was 3T and a linear gradient
background field with range
Hz was added to the
main magnetic field. We added complex Gaussian noise to produce a noisy signal with signal-to-noise ratio (SNR) of 30 dB.
The SNR was defined as the ratio of the power of signal to
the power of noise. The scan parameters were assumed to be
ms,
ms, and
mm .
C. In Vivo Validation
Among all the cases provided by the ISMRM Fat-Water Separation workshop (http://ismrm.org/workshops/FatWater12/data.
htm), we identified one 3-D dataset that was acquired using Cartesian sampling and had a final TE larger than 15 ms. The imaging
parameters were as follows: 1.5T, six echoes,
ms,
ms,
mm . A unipolar readout
was used in this experiment to avoid phase variations between
even and odd echoes. To establish a benchmark, we compared
the Hernando et al. method [17] with the proposed method. The
Hernando et al. method combined a VARPRO formulation with a
graph cut solver to solve (1), so it was denoted as VARPRO-GC
[16], [17]. Both methods gave virtually identical water and fat
maps when all six echoes are used. Thus, we took the six echo
results from VARPRO-GC as the benchmark. To test the method
performance with long echo spacing, we discarded all odd echoes
such that the effective
ms. The field map error was
defined as the
norm of the absolute difference map divided
by the
norm of the benchmark field map.
D. MR Acquisition
With Institutional Review Board approval and informed consent, a total of eleven healthy volunteers were recruited for liver
MRI. Six of them were scanned on a 1.5T MRI system (GE Excite HD, Milwaukee, WI, USA), two on a 3T MRI system (GE
Excite HD, Milwaukee, WI, USA), and three at both 1.5T and
3T. An 8-channel cardiac coil and a 4-echo spiral sequence were
used in 1.5T scans. The scan parameters were: 48 spiral leaves,
kHz,
ms,
ms,
ms,
mm . The scans
were finished in the time of a single breath hold (about 45 s).
We used an 8-channel torso coil and a 4-echo 3-D spoiled
gradient echo sequence in 3T scans. The scan parameters were:
kHz,
ms,
ms,
ms,
mm . The scans
were finished in the time of a single breath hold (about 40 s).
The matrix size for both 1.5T and 3T human data was 256 256
26.
E. Image Processing and Analysis
The proposed method was compared with VARPRO-GC.
Additionally, the Lu et al. Multi-Resolution field map esti-

535

mation using Golden Section search method (MRGS) was

also compared [25]. The MRGS method identifies candidate
solutions from the VARPRO formulation and connects the
candidate solutions to form a continuous field map using region
growing. For VARPRO-GC and MRGS, the original codes from
the ISMRM Fat-Water toolbox (http://www.ismrm.org/workshops/FatWater12/) were used. Note that both VARPRO-GC
and MRGS were designed for 2-D MR data, so VARPRO-GC
and MRGS were run on every slice of the 3-D data. We
implemented SPURS in both 2-D and 3-D. For numerical
simulations and in vivo validation, 2-D SPURS was compared
with VARPRO-GC and MRGS to allow a fair comparison.
For the acquired liver MR images, both 2-D SPURS and
3-D SPURS were compared to the literature methods. The
same single fat peak model was used for all methods. The
field map range of VARPRO-GC was set to
Hz and all other parameters were set to the default values in
the code. The final in vivo field maps obtained by each of the
methods were further processed using the modified version
of MEDI [6] to generate quantitative susceptibility maps.
The tissue masks were created in an automated process using
magnitude signal thresh-holding: voxels whose magnitude
values were less than 5% of the maximum magnitude were
masked out. All the experiments were run on a personal computer with i7-2600 CPU and 8 GB memory. Running time
for water/fat separation was recorded. The performance of
the water/fat separation for 14 abdomen cases was evaluated
by two experienced MR researchers on a 3-point scale, with
no water/fat swap
pockets of water/fat swap, and
substantial water/fat swap. A performance score was
obtained by taking the mean of the grades assigned to each
of the axial slices for each method. Quantitative analysis was
performed on the reconstructed QSM images from the 14
liver cases. Regions of interest (ROIs) were chosen by placing
a circle of radius 10 mm in the liver, a circle of radius 7.5
mm in the back muscle, and a circle of radius 3 mm in the
subcutaneous fat regions in the central slice of the 3-D volume.
Relative susceptibility values with respect to the back muscle
were recorded for each of the subjects. Inter-subject mean and
standard deviation were calculated.
IV. RESULTS
The error due to single species fit with different fat fractions
and different TEs is shown in Fig. 4. It is noted that the model
error peaked when there was an equal amount of water and fat
in a voxel and reduced to zero when only a single species was
present. The error is not a monotonic function of echo spacing,
but the peak error across fat fraction decreases with increasing
echo spacing, as can be seen from (6). For example, when the
echo spacing is larger than 4.4 ms, all model errors are smaller
than 14.97 Hz. The average size of the error is 2.67 Hz, which
corresponds to 0.02 ppm at 3T or 0.04 ppm at 1.5T, and it can
be further refined by running voxel-wise -IDEAL.
Results of numerical validation are shown in Fig. 5 and
Table I. The field map errors shown in Table I are defined as
the ratio of
norm of the difference map and the
norm of
the true map. As shown in the difference maps, 2-D SPURS

536

IEEE TRANSACTIONS ON MEDICAL IMAGING, VOL. 34, NO. 2, FEBRUARY 2015

TABLE I
COMPARISON OF DIFFERENT METHODS

Fig. 4. Error of using single species model to fit the signal with different fat
fractions and different TEs. If the voxel consists of only water or only fat, the
fitting error is zero. Average size of fitting error for different parameter is 2.67
HZ.

Fig. 5. Comparison of 2-D SPURS with VARPRO-GC and MRGS method on

a numerical phantom with different fat fraction. First column is the true values,
the second column is the difference maps of 2-D SPURS and the true value.
Third column is the difference maps of VARPRO-GC and the true value. Fourth
column is the difference maps of MRGS and the true value. Arrows point out
errors in the estimation.

gives accurate field map estimation and water/fat separation

results across various fat fractions. The field maps of both
VARPRO-GC and MRGS have moderate error at the boundary
of cylinders and cuboids. The MRGS result still suffers from
wraps in which the field is rapidly varying. For water/fat separation, VARPRO-GC and MRGS also have moderate error at
the boundary of cylinders and cuboids.
Results of the in vivo validation are shown in Fig. 6 and
Table I. As the figure shows, 2-D SPURS on the 6-echo data produces a result that is virtually identical to the benchmark. When
only 3-echo data was used, all three methods showed noticeable differences with respect to the benchmark, but 2-D SPURS
gave the lowest error compared to VARPRO-GC and MRGS.
The results of VARPRO-GC and MRGS have artifactual discontinuities in the estimated field map (arrows), which may be
the cause of the large errors measured. For VARPRO-GC and
MRGS, pockets of water/fat swaps are observed near tissue-air
interfaces.

Fig. 6. Comparison of 2-D SPURS with VARPRO-GC and MRGS in the in

vivo validation. First column is the benchmark results. Second to fifth columns
are the diffrence map between various reconstructions and the benchmarks.
Second column is from 2-D SPURS in 6-echo data. Third column is from
2-D SPURS in 3-echo data. Fourth and fifth column are from the results of
VARPRO-GC and MRGS, respectively. For VARPRO-GC and MRGS, there
are discontinuities in field map and water fat swaps in water/fat map, which
are indicated by white arrows.

Fig. 7 shows the comparison results on a 3T human data.

Both VARPRO-GC and MRGS have water/fat swaps in subcutaneous fat regions. The field maps generated by VARPRO-GC
and MRGS had large artifactual discontinuities within tissue
(white arrows). These resulted in discontinuities and artifacts
in the QSM results (indicated by white arrow). The field map
by 2-D SPURS had a small discontinuity (indicated by the
white arrow). This was no longer present in 3-D SPURS,
which takes the field map continuity along the slice selection direction into account. Since SPURS does not perform
any additional smoothing on the field map, the tissue details are better delineated in the reconstructed QSM result.
The susceptibility values of the liver were
ppm
and the susceptibility values of
the fat region were
ppm.
The average score and running time of the three methods
in 14 abdomen datasets are summarized in Table I. The 3-D
SPURS had the best water/fat separation quality on average and
had the lowest standard deviations. 3-D SPURS outperformed
2-D SPURS in water/fat separation. 2-D SPURS performed
better than VARPRO-GC and MRGS. The running time of 2-D
SPURS was significantly faster than that of VARPRO-GC.

DONG et al.: SPURS USING GRAPH CUTS: APPLICATION IN QUANTITATIVE SUSCEPTIBILITY MAPPING

Fig. 7. Comparison of SPURS with VARPRO-GC and MRGS method. First

column is the results of 3-D SPURS. Second column is the results of 2-D
SPURS. Third column is the results of VARPRO-GC. Fourth column is the
results of MRGS. From top to bottom are: field map, water map, fat map, and
QSM. For VARPRO-GC and MRGS, there are discontinuities in field map,
water fat swaps in water/fat map, and artifacts in QSM, which are indicated by
white arrows.

V. DISCUSSION
In this paper, a 3-D method for Simultaneous Phase Unwrapping and Removal of chemical Shift (SPURS) is described to
estimate the susceptibility inhomogeneity generated field map
from tissues in which fat is present. The solver is implemented
using graph cuts with conditional jump moves. Experimental
results on several datasets with large field inhomogeneity and
large echo spacing acquired at both 1.5T and 3T show that the
SPURS method outperformed two existing methods.
In most current QSM protocols, multiple TEs are sampled in
one repetition time (TR) with a long last echo to order to enhance
the SNR on the estimated field map. Due to receiver bandwidth
and other hardware limits, the echo spacing in such acquisitions
is about 5 ms. This is much longer than the optimal TE spacing
preferred by most existing water/fat separation methods. In this
paper, the field map is first estimated using a single species fitting, and long echo spacing actually reduces the model bias in
this simplified model. Compared with multi-species fitting that
involves solving for six real unknown variables, single species
fitting is more robust because it only has a single unknown being
the off-resonance frequency, and
is not an unknown subject
to optimization [6]. With the single species fitting, the water/fat
ambiguity is reduced to a problem parallel to phase unwrapping,
which has been carefully studied in literature.
The spatial continuity of the field map is enforced using
graph cuts algorithm with conditional jump moves to judiciously connect the candidate solutions. Graph cuts algorithm is
well-suited for combinatorial optimization problems satisfying
the submodular condition, such as the problem formulated in
(8). An additional fine-tuning of the field map is performed
using a voxel-wise
-IDEAL algorithm without spatial regularization. The sole purpose of this step is to correct the
model error introduced in the single species fitting and to
obtain water and fat maps. Because the expected distance
between the initial guess and the truth is the model error, and

537

the model error is small from the theoretical and numerical

analysis, the -IDEAL step converged to the global minimum
in most cases. It should be noted that no additional spatial
smoothing is performed, so SPURS does not have the problem
of over-smoothing that VARPRO-GC has. In some cases, the
-IDEAL fine tuning step may be optional for QSM applications, as the employed long TE in QSM applications effectively
reduces the model error introduced in the single species fitting.
The proposed SPURS method is also faster than VARPRO-GC
and MRGS, partly because the candidate generation step is more
efficient. There are some other efficient water/fat separation
methods that save time in the candidate connection stage, such
as proposed in [20], [28]. However, wraps often remain in the
estimated field map in these methods, as field map continuity is
not strictly enforced. In comparison, SPURS results in a spatially
continuous field map without phase wraps, which is important
for optimal QSM reconstruction.
Because the magnetic dipole kernel is a long-range field in
nature, discontinuities in one location will propagate to other locations and other slices. QSM images from VARPRO-GC and
MRGS still had discontinuous artifacts (white arrows in Fig. 7)
resulting from artifactual discontinuities in their estimated field
maps. The best QSM results were obtained when using 3-D
SPURS for field map estimation. Beyond the reasons mentioned
above, 3-D SPURS further enforces continuity along the third
dimension. Streaking artifacts that were often seen in QSM results were well suppressed due to the spatial regularization employed in MEDI, and the Laplacian also suppresses streaking
from remnant background field.
The following factors need to be considered when choosing
practical TEs to lead to optimal results. 1) If water and fat need
to be separated, the minimum number of echoes for SPURS is
3, as -IDEAL is needed. The choice of TE should not make
the absolute value of
(effective phase of fat chemical shift)
too small. In a theoretical case, where the angle between water
and fat remains exactly the same for all echoes,
and
water and fat cannot be separated. 2) When focusing on the
field map only, acquisition with a long last TE is preferred. This
is evident by analyzing the model error in SPURS. The model
error
is roughly inversely proportional to the
number of echoes multiplied by echo spacing
. A long last
TE also allows adequate phase accrual to improve the SNR on
the field map. The upper limit for the last TE should be determined by the
of the tissue being imaged in order to avoid fitting pure noise. Considering the preference for a long last echo
and a realistic
of tissue, current QSM protocols employing
a last echo ranging from 2050 ms appear to offer a balance between the two competing factors.
In this study, we demonstrated that the field map estimated
using SPURS is useful for liver QSM reconstruction. The liver
is a major iron storage organ in the human body. Excessive iron
deposition in the liver may indicate pathologies, such as those
related to thalassemia. Although
mapping has been used to
estimate iron concentration,
also includes contribution from
, making it less specific to the magnetic field inhomogeneity
caused by iron. In addition, QSM provides an opportunity to
study the oxygen level in the vessels, which may be useful to
investigate the oxygen extraction process in the liver. The use

538

IEEE TRANSACTIONS ON MEDICAL IMAGING, VOL. 34, NO. 2, FEBRUARY 2015

of susceptibility provides a more direct measure of fundamental

physiological tissue parameters [43] that are related to disease
diagnosis, monitoring, and treatment. Conceptually, the SPURS
method can be easily applied to imaging other organs with adipose tissue, such as joint imaging, breast imaging [44], or whole
body mouse imaging.
Several limitations need to be acknowledged. We did not aim
to address the waterfat separation problem when the fat-only
or water-only region is absolutely isolated from other regions
by air. This is because such a scenario is not very common
in human body imaging and most existing water-fat separation
methods, including VARPRO-GC and MRGS, are not designed
for this purpose. If an accurate identification of fat is desirable
in this case, other methods such as short tau inversion recovery
could be employed. From the model error analysis of the single
species fitting, it is found that the largest error occurs in SPURS
when the water and fat ratio in a voxel is 1:1. Fortunately, it
is observed that most of the voxels in an imaging field of view
have one dominant species. We restrict step size to be positive in order to make the algorithm easier to implement. The
restriction of only introduces a constant overall offset of
in the final field map, which does not affect the water fat separation and the QSM reconstruction results. Another limitation
in this study is that a multi-peak fat model is not considered
[45]. Nevertheless, multi-peak IDEAL may be adopted in the
last fine-tuning step [46].

Invoking
uated as

VI. CONCLUSION
In this paper, we proposed a 3-D joint field map estimation
and water/fat separation method. A field map is first estimated
via solving a single species fitting problem. Simultaneous phase
unwrapping and chemical shift removal was then performed
using graph cuts with conditional jump moves. Finally, the field
map was fine-tuned by -IDEAL. Experiments on numerical
phantom and human data showed that the SPURS method provides accurate water/fat separation results in MR datasets with
long echo spacing and in areas where field map varied rapidly.
The 3-D continuous field map without wrapping is well suited
for QSM reconstruction.

, this cost function can be eval-

(A3)
Its minimum is found by setting

(A4)
When the phases in the sine functions are close to
with
as integers depending on echo index , a Taylor expansion can be used

(A5)
Then (A4) becomes

(A6)
(A7)
For ease of explanation, we assume
. This
is not a necessary condition as any additional time constant
in equally spaced TEs can be absorbed by the initial phase of
. With
( an integer), (A7) becomes
(A8)
can be interpreted as a wrap caused by the limited
Here
dynamic range.
If
, the signal equation can be written
as
,
shown
in
Fig.
2(b).
Then

APPENDIX A
SINGLE SPECIES FITTING ERROR ANALYSIS

. Similar to (A1) and (A8), we have

If
. Following (3), we use the following
single species model to fit the two species signal equation:

(A9)
(A10)

(A1)
For any complex-valued number , we have
the above objective function can be written as

, so

(A2)

So, the approximate model error of using single species fitting

is

(A11)

DONG et al.: SPURS USING GRAPH CUTS: APPLICATION IN QUANTITATIVE SUSCEPTIBILITY MAPPING

APPENDIX B
PROOF OF THE SUBMODUALR CONDITION
Theorem 1: For 1-jump move optimization, as long as
is convex, the submodular condition holds for all binary energy
term
.
Proof: From (11),

, where
For any convex function
In our problem,

.
, we have
, where
.
is a convex function. Let

, we have
.
So we have the following inequality from the definition of
convex function:

(A12)
Similarly,

, we have

(A13)
Summing the above two inequalities, we get

(A14)
Multiplying the positive value
to both sides of the
above inequality, we get the inequality (12). So the submodular
condition holds for all binary energy terms in 1-jump move optimization. We introduce the function
to emphasize that
any convex function satisfying the submodular condition can
be used.
ACKNOWLEDGMENT
The authors would like to thank A. J. Luzzi for proofreading. We acknowledge the use of the Fat-Water Toolbox
(http://ismrm.org/workshops/FatWater12/data.htm).
REFERENCES
[1] L. de Rochefort, R. Brown, M. R. Prince, and Y. Wang, Quantitative
MR susceptibility mapping using piece-wise constant regularized
inversion of the magnetic field, Magn. Reson. Med., vol. 60, pp.
10031009, Oct. 2008.
[2] T. Liu, P. Spincemaille, L. de Rochefort, B. Kressler, and Y. Wang,
Calculation of susceptibility through multiple orientation sampling
(COSMOS): A method for conditioning the inverse problem from
measured magnetic field map to susceptibility source image in MRI,
Magn. Reson. Med., vol. 61, pp. 196204, Jan. 2009.
[3] K. Shmueli, J. A. de Zwart, P. van Gelderen, T. Q. Li, S. J. Dodd, and
J. H. Duyn, Magnetic susceptibility mapping of brain tissue in vivo
using MRI phase data, Magn. Reson. Med., vol. 62, pp. 15101522,
Dec. 2009.
[4] L. de Rochefort et al., Quantitative susceptibility map reconstruction
from MR phase data using bayesian regularization: Validation and application to brain imaging, Magn. Reson. Med., vol. 63, pp. 194206,
Jan. 2010.

539

[5] J. Liu et al., Morphology enabled dipole inversion for quantitative

susceptibility mapping using structural consistency between the magnitude image and the susceptibility map, Neuroimage, vol. 59, pp.
25602568, Feb. 1, 2012.
[6] T. Liu, C. Wisnieff, M. Lou, W. Chen, P. Spincemaille, and Y. Wang,
Nonlinear formulation of the magnetic field to source relationship for
robust quantitative susceptibility mapping, Magn. Reson. Med., vol.
69, pp. 467476, Feb. 2013.
[7] E. M. Haacke, J. Tang, J. Neelavalli, and Y. C. Cheng, Susceptibility
mapping as a means to visualize veins and quantify oxygen saturation,
J. Magn. Reson. Imag., vol. 32, pp. 663676, Sep. 2010.
[8] S. Wharton and R. Bowtell, Whole-brain susceptibility mapping at
high field: A comparison of multiple- and single-orientation methods,
Neuroimage, vol. 53, pp. 515525, Nov. 1, 2010.
[9] S. Wharton, A. Schafer, and R. Bowtell, Susceptibility mapping in the
human brain using threshold-based k-space division, Magn. Reson.
Med., vol. 63, pp. 12921304, May 2010.
[10] W. Li, B. Wu, and C. Liu, Quantitative susceptibility mapping of
human brain reflects spatial variation in tissue composition, Neuroimage, vol. 55, pp. 16451656, Apr. 15, 2011.
[11] F. Schweser, A. Deistung, B. W. Lehr, and J. R. Reichenbach, Quantitative imaging of intrinsic magnetic tissue properties using MRI signal
phase: An approach to in vivo brain iron metabolism?, Neuroimage,
vol. 54, pp. 27892807, Feb. 14, 2011.
[12] F. Schweser, K. Sommer, A. Deistung, and J. R. Reichenbach, Quantitative susceptibility mapping for investigating subtle susceptibility
variations in the human brain, Neuroimage, vol. 62, pp. 20832100,
Sep. 2012.
[13] B. Wu, W. Li, A. Guidon, and C. Liu, Whole brain susceptibility
mapping using compressed sensing, Magn. Reson. Med., vol. 67, pp.
137147, Jan. 2012.
[14] H. Yu et al., Multiecho reconstruction for simultaneous water-fat decomposition and T2* estimation, J. Magn. Reson. Imag., vol. 26, pp.
11531561, Oct. 2007.
[15] J. Ma, Dixon techniques for water and fat imaging, J. Magn. Reson.
Imag., vol. 28, pp. 543558, Sep. 2008.
[16] D. Hernando, J. P. Haldar, B. P. Sutton, J. Ma, P. Kellman, and Z. P.
Liang, Joint estimation of water/fat images and field inhomogeneity
map, Magn. Reson. Med., vol. 59, pp. 571580, Mar. 2008.
[17] D. Hernando, P. Kellman, J. P. Haldar, and Z. P. Liang, Robust
water/fat separation in the presence of large field inhomogeneities
using a graph cut algorithm, Magn. Reson. Med., vol. 63, pp. 7990,
Jan. 2010.
[18] A. J. Madhuranthakam et al., T(2)-weighted 3-D fast spin echo
imaging with water-fat separation in a single acquisition, J. Magn.
Reson. Imag., vol. 32, pp. 745751, Sept. 2010.
[19] H. Yu, S. B. Reeder, A. Shimakawa, C. A. McKenzie, and J. H. Brittain,
Robust multipoint water-fat separation using fat likelihood analysis,
Magn. Reson. Med., vol. 67, pp. 10651076, Apr. 2012.
[20] J. Berglund and J. Kullberg, Three-dimensional water/fat separation
and T2* estimation based on whole-image optimizationApplication
in breathhold liver imaging at 1.5 T, Magn. Reson. Med., vol. 67, pp.
16841693, June 2012.
[21] J. Tsao and Y. Jiang, Hierarchical IDEAL: Fast, robust, and multiresolution separation of multiple chemical species from multiple echo
times, Magn. Reson. Med., vol. 70, pp. 155159, Jul. 2013.
[22] Q. S. Xiang and L. An, Water-fat imaging with direct phase encoding,
J. Magn. Reson. Imag., vol. 7, pp. 10021015, Nov.Dec. 1997.
[23] W. T. Dixon, Simple proton spectroscopic imaging, Radiology, vol.
153, pp. 189194, Oct. 1984.
[24] A. R. Pineda, S. B. Reeder, Z. Wen, and N. J. Pelc, Cramer-Rao
bounds for three-point decomposition of water and fat, Magn. Reson.
Med., vol. 54, pp. 625635, Sep. 2005.
[25] W. Lu and B. A. Hargreaves, Multiresolution field map estimation
using golden section search for water-fat separation, Magn. Reson.
Med., vol. 60, pp. 236244, Jul. 2008.
[26] J. Berglund, L. Johansson, H. Ahlstrom, and J. Kullberg, Three-point
Dixon method enables whole-body water and fat imaging of obese subjects, Magn. Reson. Med., vol. 63, pp. 16591568, June 2010.
[27] W. Lu and Y. Lu, JIGSAW: Joint inhomogeneity estimation via global
segment assembly for water-fat separation, IEEE Trans. Med. Imag.,
vol. 30, pp. 14171426, Jul. 2011.
[28] A. S. Soliman, J. Yuan, K. K. Vigen, J. A. White, T. M. Peters, and C.
A. McKenzie, Max-IDEAL: A max-flow based approach for IDEAL
water/fat separation, Magn. Reson. Med., vol. 72, pp. 510521, Aug.
2014.

540

IEEE TRANSACTIONS ON MEDICAL IMAGING, VOL. 34, NO. 2, FEBRUARY 2015

[29] Y. Wang and T. Liu, Quantitative susceptibility mapping (QSM): Decoding MRI data for a tissue magnetic biomarker, Magn. Reson. Med.,
Jul. 17, 2014.
[30] F. Huang, S. Narayan, D. Wilson, D. Johnson, and G. Q. Zhang, A
fast iterated conditional modes algorithm for water-fat decomposition
in MRI, IEEE Trans. Med. Imag., vol. 30, no. 8, pp. 14801492, Aug.
2011.
[31] P. Stoica and R. Moses, Spectral Analysis of Signals. Upper Saddle
River, NJ: Pearson/Prentice Hall, 2005, ch. 4, pp. 151155.
[32] H. Gudbjartsson and S. Patz, The Rician distribution of noisy MRI
data, Magn. Reson. Med., vol. 34, pp. 910914, Dec. 1995.
[33] L. Ying, Z. P. Liang, D. C. Munson, Jr., R. Koetter, and B. J. Frey, Unwrapping of MR phase images using a Markov random field model,
IEEE Trans. Med. Imag., vol. 25, no. 1, pp. 128136, Jan. 2006.
[34] O. Veksler, Efficient graph-based energy minimization methods in
computer vision, Ph.D. dissertation, Cornell Univ., Ithaca, NY, 1999.
[35] Y. Boykov and V. Kolmogorov, An experimental comparison of
min-cut/max-flow algorithms for energy minimization in vision,
IEEE Trans. Pattern Anal. Mach. Intell., vol. 26, no. 9, pp. 11241137,
Sep. 2004.
[36] J. M. Bioucas-Dias and G. Valadao, Phase unwrapping via graph
cuts, IEEE Trans. Image Process., vol. 16, no. 3, pp. 698709, Mar.
2007.
[37] V. Kolmogorov and R. Zabih, What energy functions can be minimized via graph cuts?, IEEE Trans. Pattern Anal. Mach. Intell., vol.
26, no. 2, pp. 147159, Feb. 2004.
[38] L. Li and J. S. Leigh, Quantifying arbitrary magnetic susceptibility
distributions with MR, Magn. Reson. Med., vol. 51, pp. 10771082,
May 2004.

[39] F. Schweser, A. Deistung, B. W. Lehr, and J. R. Reichenbach, Differentiation between diamagnetic and paramagnetic cerebral lesions
based on magnetic susceptibility mapping, Med. Phys., vol. 37, pp.
51655178, Oct. 2010.
[40] T. Liu, D. Zhou, P. Spincemaille, and Y. Wang, Differential approach
to quantitative susceptibility mapping without background field removal, in Proc. 22nd Annu. Meet. ISMRM, Milan, Italy, 2014, p.
0597.
[41] T. Liu, W. Xu, P. Spincemaille, A. S. Avestimehr, and Y. Wang, Accuracy of the morphology enabled dipole inversion (MEDI) algorithm for
quantitative susceptibility mapping in MRI, IEEE Trans. Med. Imag.,
vol. 31, no. 3, pp. 816824, Mar. 2012.
[42] S. B. Reeder et al., Multicoil Dixon chemical species separation with
an iterative least-squares estimation method, Magn. Reson. Med., vol.
51, pp. 3545, Jan. 2004.
[43] D. Hernando, R. J. Cook, C. Diamond, and S. B. Reeder, Magnetic
susceptibility as a B0 field strength independent MRI biomarker of liver
iron overload, Magn. Reson. Med., vol. 70, pp. 648656, June 25,
2013.
[44] A. V. Dimov et al., Joint estimation of chemical shift and quantitative
susceptibility mapping (chemical QSM), Magn. Reson. Med., Jun. 19,
2014.
[45] D. Hernando, Z. P. Liang, and P. Kellman, Chemical shift-based
water/fat separation: A comparison of signal models, Magn. Reson.
Med., vol. 64, pp. 811822, Sep. 2010.
[46] R. Kijowski et al., Improved fat suppression using multipeak reconstruction for IDEAL chemical shift fat-water separation: Application
with fast spin echo imaging, J. Magn. Reson. Imag., vol. 29, pp.
436442, Feb. 2009.