improved optimization strategies for autofocusing motion compensation in mri via the analysis of...
TRANSCRIPT
Magnetic Resonance Im
Improved optimization strategies for autofocusing motion compensation in
MRI via the analysis of image metric maps
Wei Lin, Hee Kwon Song4
Department of Radiology, University of Pennsylvania Medical Center, Philadelphia, PA 19104, USA
Received 11 May 2005; accepted 14 February 2006
Abstract
Autofocusing is a postprocessing technique for motion correction, which optimizes an image quality metric against various trial motions.
In this work, image metric maps, which are measures of image quality plotted as a function of in-plane 2-D trial translations, are
systematically studied to develop improved autofocusing motion correction algorithms. It is shown that determining object motion with
autofocusing is equivalent to an image metric map optimization problem. These maps provide insights into the motion compensation
process and help improve several aspects of the correction algorithm, including the selection of the image metric and motion search
strategy. A highly efficient and robust 2-D global optimization method is devised, exploiting the properties of the metric map pattern. The
improved algorithm is used to correct phantom and clinical MR images with in-plane 2-D translational motion and is shown to be more
effective than existing methods.
D 2006 Elsevier Inc. All rights reserved.
Keywords: Motion correction; Autofocusing; Autocorrection; Image metric
1. Introduction
From the early years of MR imaging, the acquisition of
high-quality MR images in vivo was often hampered by
motion-related artifacts, and various prospective and retro-
spective techniques have been developed for their compen-
sation. Among these are physiologic gating, data reordering
and spatial presaturation [1–3]. Although effective, these
techniques are limited to known periodic physiologic
movements and cannot be used for arbitrary motion. One
of the more effective means of compensation for transla-
tional motion in use today is the method of bnavigatorechoesQ [4], which requires the acquisition of additional data
during the scan to extract the motion information. In some
high-resolution imaging sequences, however, it may not be
desirable or feasible to obtain additional data since the
minimum sequence repetition time (TR) or the total scan
time can become prolonged. The use of navigator echoes
may also undesirably affect the steady state.
Several postprocessing techniques have also been pro-
posed for correction of general translational motion. The
0730-725X/$ – see front matter D 2006 Elsevier Inc. All rights reserved.
doi:10.1016/j.mri.2006.02.003
4 Corresponding author. Tel.: +1 215 662 6265; fax: +1 215 349 5925.
E-mail address: [email protected] (H.K. Song).
method of generalized projections attempts to correct for
motion by estimating the phase errors in the acquired data
[5], while the method of Zoroofi et al. [6] uses a
combination of both edge detection and phase error
estimation. Although shown to be successful in simulation
studies and simple phantom experiments, these techniques
have not been tested successfully in in vivo imaging. They
are dependent on the existence of significant ghosting
artifacts or bright object boundaries, which are often not
present in high-resolution images of objects undergoing
subtle, nonperiodic motion, and it is likely that these
methods will fail in typical in vivo situations.
One of the postprocessing techniques devised to correct
motion artifacts is autofocusing (also known as autocorrec-
tion), which was first introduced for application in MR
imaging by Atkinson et al. [7] and has been improved since
[8–11]. In autofocusing, motion is estimated by trial and
error but in a systematic fashion. For each k-space segment,
which consists of one or more contiguous phase-encoding
views, a range of possible object displacements is assumed
and the corresponding phase shifts in k-space are applied.
The data set that yields a minimum (or a maximum) in an
image metric, a measure of image sharpness and quality, is
retained and assumed to be motion compensated. The
aging 24 (2006) 751–760
Fig. 1. k-Space regions used for metric map computation. The segment S, to
which the trial motions are applied, and the background B do not overlap,
and the combined area does not necessarily cover the entire k-space.
W. Lin, H.K. Song / Magnetic Resonance Imaging 24 (2006) 751–760752
method assumes that there is little or no motion during
acquisition of each segment and that those data are acquired
contiguously in time. Autofocusing does not require
additional data and has been shown to successfully
compensate for motion, often performing as well as the
navigator echo technique in cooperative volunteers [7,8].
Both 1-D and 2-D autofocusing have also been demonstrat-
ed in several clinical examples [10,11]. However, high
computational costs and the lack of robustness still impede
wide clinical applicability of the method.
In this work, we utilize the patterns observed in the
metric maps, which are 2-D images whose individual pixel
intensity corresponds to the metric value for a given 2-D
trial translation, for developing improved autofocusing
algorithms. From a systematic analysis of the metric map
patterns, improvements in both speed and accuracy of the
motion correction process are developed. Both the choice of
an image metric and an efficient search strategy could be
determined based on the behavior of these maps. It should
be emphasized here that these maps are used solely to
develop improved autofocusing strategies; they are not
computed during actual motion compensation. The im-
proved algorithm is used to correct for translational motion
in phantom and clinical MR images of the wrist, and the
results compare with previous autofocusing schemes, as
well as with navigator correction techniques.
2. Theory
A bulk rigid-body in-plane object translation of (Dx, Dy)
causes a complex phase factor to be multiplied to the
measured MR signal:
F Dx;Dyð Þ kx; ky� �
¼ F 0;0ð Þ kx; ky� �
e�j2p kxDxþkyDyð Þ: ð1Þ
Here, superscripts indicate the amount of the translation.
Throughout this article, it is assumed that frequency
encoding is along the x-axis and phase encoding is along
y. In a typical spin- or gradient-echo sequence, the duration
of the phase encoding and readout gradients are much
shorter than the TR period between successive views (i.e.,
different ky values). Therefore, intra-view motion can be
considered negligible compared with inter-view motion, and
both Dx and Dy can be taken as only dependent on ky. As a
result, translational motion correction is essentially estimat-
ing a total of NPE 2-D vectors T, where NPE is the number of
phase-encoding views:
T ky� �
¼ Dx ky� �
;Dy ky� �� �
;
kya � NPE=2; N ; 0; N ;NPE=2� 1f g: ð2Þ
The autofocusing technique attempts to successively recover
each component of T by searching for the trial translation
(�Dx, �Dy) that would minimize (or maximize) the metric
for the reconstructed image I.
Let us define S, the k-space data segment in which the
trial motion is applied, and B, the botherQ ky lines includedin the current analysis, which may contain only a portion of
the remaining data (Fig. 1). We define the metric map for
translational motion MS |B(dx,dy) as the computed image
metric versus translation (dx,dy) applied to the k-space
segment S, while the image is reconstructed from the
combined k-space data of S and B. Formally,
MSjB dx; dyð ÞuRðFT�1ðF dx;dyð ÞSjB ÞÞ
¼ RðFT�1ðF dx;dyð ÞS þ F
0;0ð ÞB ÞÞ: ð3Þ
Here, R is the metric operation, the subscript denotes the
k-space coverage of the data F and the superscript is the
translation applied.
If autofocusing were to estimate motion correctly, a
global extremum should appear at the origin of the metric
map computed on a motion-free image. Assuming that the
metric minimum is sought, we have
arg mindx;dyð Þ
Mf
SjB dx; dyð Þ ¼ 0; 0ð Þ: ð4Þ
Here, the superscript f indicates that metric map is computed
over a motion-free image. For a motion-corrupted image
where a translation of (Dx,Dy) occurs when acquiring data
contained in segment S, the metric map is (assuming that
background B is already corrected and therefore motion
free):
McSjB dx; dyð Þ ¼ M
f
SjB dxþ Dx; dyþ Dyð Þ: ð5Þ
Here, the superscript c indicates that metric map is
computed over a motion-corrupted image. From Eqs. (4)
and (5), we have
arg mindx;dyð Þ
McSjB dx; dyð Þ ¼ � Dx; � Dyð Þ: ð6Þ
Fundamentally, autofocusing is a global optimization
problem [12] that searches the variable (dx ,dy) that
W. Lin, H.K. Song / Magnetic Resonance Imaging 24 (2006) 751–760 753
minimizes (or maximizes) the metric map function
MS |Bc(dx,dy) for various S and B configurations. The
characteristic of the metric map is therefore of great interest
since it dictates how to perform the optimization process.
Metric maps computed from a motion-free brain MR data
set are shown in Fig. 2. As expected, a minimum always
occurs at the center of the maps, which corresponds to a
motion-free state. The maps also show an oscillation along
the y-direction (phase encoding), confined in a narrow band
along the dx=0 line, with multiple minima at
arg mindx;dyð Þ
MSjB dx; dyð Þi 0; nDy
� �; Dy ¼ 1=kSy : ð7Þ
Here, kyS is the average spatial frequency of segment
S. Segments close to the k-space center have smaller kyS
values and larger oscillation periods. The periodicity
along the phase-encoding direction on the metric map is
expected: a shift of Dy while acquiring one view with
a spatial frequency of ky causes a phase factor of
Uy=exp[�j2pkyDy(ky)] to be multiplied to the acquired
data. One cannot distinguish between shifts of Dy and
Dy+nDy, where n is an arbitrary integer. Eq. (7) is exact
when S is a single phase-encoding line. When segment S
contains more lines, however, only the minimum at (0,0) is
global while the other minima become local, as shown in
Fig. 2B. This is due to the existence of a distribution of Dy
values within segment S; that is, each line within S has
Fig. 2. The metric maps. (A) The motionless image used for the metric comp
consecutive phase-encoding lines S ={(kx,ky)|kya[k1,k1+n�1]}. First row: n =1 (o
locations: k1=8, 16, 32, 64 and 96 are shown from left to right. The NGS metric (E
in number of pixel shifts, and the intensity of each map is scaled individually fo
slightly different periodicity, resulting in diminished oscil-
lation amplitudes away from the true object position.
3. Methods
3.1. Image metrics
Two types of image metrics were primarily studied based
on their successful outcomes in previous reports: entropy
[7,8] and normalized gradient squared (NGS) [9]. Entropy is
an image metric used for autofocusing in inverse synthetic
aperture radar imaging [13] from which the idea of
autofocusing MR images was derived and was shown to
successfully compensate for motion in MR studies of
volunteers [7,8]. It is defined as
E ¼ �Xi;j
Bi;j=Bmax
� �ln Bi;j=Bmax
� �;
Bmax ¼ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiXi;j
B2i;j
sð8Þ
where Bi,j is the pixel intensity at coordinate (i,j). The term
Bmax can be interpreted as the pixel brightness if all the
energy was contained in a single pixel. Since translational
motion only introduces a phase factor to each k-space
data point, the total image energy and Bmax are invariant.
NGS was among the best-performing metrics in a study
of 24 different metrics (including entropy) [9] and was
utation, SNR=25. (B) Metric maps computed for segment S containing
ne phase-encoding line), second row: n =32. Maps for segments at different
q. (9)) and Background Selection Method 1 (Fig. 3) were used. The axes are
r display purposes.
W. Lin, H.K. Song / Magnetic Resonance Imaging 24 (2006) 751–760754
therefore also studied. Although another metric, the gradient
entropy, slightly outperformed NGS, the latter was used in
this study due to its lower computation cost. It is defined as
NGS ¼Xi;j
jGi;jj2� �
=� X
i;j
jGi;jj�2
: ð9Þ
|Gi,j|, the gradient magnitude at pixel coordinate (i,j), was
computed by taking the square root of the sum of the
squares of x and y gradients. Both directional gradients were
computed using a simple filter of [�1,1]. Other filters such
as the steerable Gaussian first-derivative basis filters [14],
which has a proven optimality for edge detection and allows
trade-off between noise filtering and fine-scale structure
detection, can be used, if desired. However, the computation
cost by using such a kernel would be significantly higher,
and it is therefore not considered in this study. For a sharper
image, entropy is minimized while NGS is maximized. To
simplify the discussion, we always minimized metrics in
this work. Therefore, for NGS, the metric used was:
M ¼ � NGS: ð10Þ3.2. k-Space segment and background selection
Different k-space segment S and background B selection
methods have been used in previous autofocusing algo-
rithms [7–11]. However, to the best of the authors’
knowledge, no methodical comparison of the various
possible techniques has been reported. Manduca et al. [10]
proposed a bmultiresolutionQ scheme that optimizes the
k-space in several iterations, first using large segment sizes
(e.g., 64 lines) followed by successively smaller ones. In
their scheme, as well as in the originally proposed method
[7], the k-space data were always used in their entirety
during motion compensation. In contrast, Atkinson et al. [8]
proposed excluding k-space lines that have not yet been
corrected, proceeding in a center-out fashion in k-space.
This method has the advantage in that all data in
background B is already motion compensated, improving
the likelihood of correctly determining the motion present in
the uncorrected segment S.
Fig. 3. Four different k-space background selection methods. The black
area indicates segment S, and the gray area indicates background k-space
region B.
To evaluate the performance of these differing strategies,
we studied the metric maps of four k-space background
selection methods, as shown in Fig. 3. In Method 1, only the
k-space regions more central to segment S are included. The
ky lines that are conjugate of S are not included even if those
lines have been previously motion compensated. In Method
2, background B contains not only the more central regions
as in Method 1 but also the region that is the conjugate of S.
Atkinson et al.’s [8] center-out method, including only
motion-compensated background region, essentially alter-
nates between Methods 1 and 2, with the exception that their
processing is performed on a reduced matrix that only
includes the corrected data and current segment (B and S
shown in Fig. 3), without zero filling to the original matrix
size. Method 3 is the conventional ball otherQ k-space data
technique. Finally, in Method 4, the conjugate of segment S
is always excluded from the background used in Method 3.
k-space regions not included in S or B are zero filled to the
original size to maintain consistent image resolution.
For each of the background selection schemes, the effects
of different locations and sizes of the segment S were
investigated systematically by analyzing the metric maps.
Several motion-free phantom and clinical MR images were
used to compute the metric map MS |Bf(dx,dy). For each
configuration of k-space segment S and background B, the
metric map was computed for
dx; dyð Þj dxa � Nx;Nx½ ; dya � Ny;Ny
� �: ð11Þ
Here, Nx and Ny set the limits of the trial motion. The
segment S contains data in consecutive ky lines:
S ¼ kx; ky� �
jkya k1; k1F n� 1ð Þ½ �
: ð12Þ
The positive sign was used for k1N0, and the negative sign
was used for k1b0. The segments studied contain n=1, 2, 4,
8, 16, 32 and 64 ky lines. The locations of the segments,
measured by the distance from the k-space center, were
|k1|=8, 16, 32, 64 and 96. To test the metric map behavior at
different SNRs, we added Gaussian noise with zero mean
and various standard deviations to the k-space data prior to
the metric map computation.
3.3. Application of trial motions in k-space and image space
Trial motions can be applied either in k-space [7–10] or
in image space [11] to compute the image metrics. In the
original k-space approach, trial motions involve a phase
factor to be multiplied to the k-space segment S:
Fdx;dyð Þ
S ¼ F0;0ð ÞS e�j2p kxdxþkydyð Þ: ð13Þ
Then, an inverse FFT is taken over the combined k-space
data S+B to obtain the image:
ISjB dx; dyð Þ ¼ FT�1 Fdx;dyð Þ
S þ F0;0ð ÞB
�:
�ð14Þ
Autofocusing can also be performed in image space. Due
to the linearity of the Fourier transform, the reconstructed
W. Lin, H.K. Song / Magnetic Resonance Imaging 24 (2006) 751–760 755
MR image I is a summation of many bcomponent imagesQIi, each of which contains contribution from a spatial
frequency component Fi:
I ¼ FT�1ðXi
FiÞ ¼Xi
Ii; Ii ¼ FT�1 Fið Þ: ð15Þ
From this viewpoint, a translation occurring while acquiring
one phase-encoding line simply causes the same amount of
the translation to be applied to the corresponding component
image Ii. In the image-space autofocusing approach, two
complex images are first obtained by separate inverse FFTs,
the bsegment imageQ IS for S and the bbackground imageQ IBfor B. Then, translational shifts are applied to IS before it is
added to IB:
ISjB dx; dyð Þ ¼ IS dx; dyð Þ þ IB 0; 0ð Þ: ð16Þ
The k-space data have to be zero filled to the full prescribed
resolution before transforming to image space to maintain
consistent image sizes and to permit single pixel shifts.
In this study, translations were applied in the image space
for its speed advantage. In the image-space approach, only
two FFTs need to be performed for each k-space segment,
while Ntrial FFTs are required for the k-space approach
where Ntrial is the number of trial motions. Since, typically,
NtrialH2 and image shifting and summation involve less
computation than FFT, image-space approach has consider-
able speed advantage, between 30% and 50% [11].
Furthermore, if the image metric is computed only within
a small user-defined region of interest, pixels only in that
region need to be shifted and summed, further reducing the
computational cost. Fractional pixel shifts can be performed,
by first zero filling the segment S to a larger matrix before
the FFT. For example, if the segment S contains 8 phase-
encoding lines in a 256�256 original data set, the 8�256
data can be zero filled to a matrix size of 512�512 before
the FFT, which subsequently allows a translational precision
of 1/2 pixel. This method of subpixel shifting is mathemat-
ically equivalent to applying a 1/2 pixel shift in k-space
using the appropriate phase.
3.4. Optimization strategy
The task of motion detection for a given configuration of
segment S and background B by a measure of image metric is
a 2-D global optimization problem. In the presence of 2-D
translation, the position of the global minimum (motion-free
state) is shifted accordingly. Various searching strategies to
locate this point have been used in the past, including
exhaustive 2-D [7], successive 1-D [8], golden section [10]
and simplex [11]. An exhaustive 2-D search is equivalent to
computing the entire metric map. Although robust, its
computational cost becomes prohibitively high when number
of trial motions becomes large. Successive 1-D, golden
section and simplex are local optimization techniques [12],
whose success relies on the assumption that the initial choice
of the translation vector lie in the local convergence region of
the global minimum. Since there could be multiple local
minima on the metric map, the algorithm may not always
converge to the true global location.
The patterns that are observed in the metric maps allow
us to design a more efficient optimization strategy. The new
strategy proposed here is composed of two distinct steps. In
the first step, the displacement of the metric minimum
along the x-direction is first located. Since the periodic
pattern that appear along dx=0 occurs throughout the entire
length of the metric map, the search could be facilitated by
first locating this periodic band. In the NGS metric map,
the vertical band is approximately 1 pixel wide. Therefore,
a small step size of 1 pixel should be used along x. Along
the y-axis, a search range of one oscillation period (Dy in
Eq. (7)) and a step size of a quarter of this period are
sufficient in locating this band. Therefore, the number of
metric evaluations performed in this step is 4Nx, where Nx is
the search range along x. To reiterate, the goal of this first
step is not to locate the absolute minimum but to locate the
x-axis shift.
In the second step, various local optimization techniques
previously proposed can be applied. In this work, a
successive 1-D searching strategy is used. Starting from
the minimum detected in Step 1, a y-axis search is first
performed, alternating with an x-axis search until the
location of the minimum no longer changes in two
consecutive searches. However, because of the existing
periodicity of the local minima along the y-axis, the
alternating 1-D local search is performed not only from
the minimum detected in Step 1 but also in the regions of
periodically occurring local minima, whose positions are
solely determined by the location of segment S, ensuring
that the global minimum could be found.
Our proposed strategy is different from previous techni-
ques in that a coarse 2-D search is used first to detect x-axis
shift, followed by successive 1-D iterations. Compared to
just a successive 1-D approach, the technique is more robust
in locating the desired minimum. It is worth repeating here
that the entire metric map does not need to be computed
during actual motion compensation; the patterns observed in
these metric maps were only used to develop the best means
to locate the global minimum.
3.5. Image acquisition and analysis
Motion-corrupted phantom and high-resolution in vivo
wrist data were acquired. Translational motion was manu-
ally applied during the phantom scan, while in one of the in
vivo scans, the subject was informed to move voluntarily
several times during the imaging session. Two other in vivo
data sets were from a high-resolution wrist imaging study in
which the subjects are advised to lie still [15]. In the
phantom experiment, a 2-D spin-echo sequence with the
following parameters was used: 24 cm field of view (FOV),
5 mm thickness, 256�256 matrix, 500 ms TR. A piecewise
constant translational motion was applied manually eight
times during the scan. For the in vivo experiments, a 3-D
Fig. 4. The NGS metric maps at low SNR. Background Selection Method 1 was used. Noise was added to the data for the image shown in Fig. 2A and the
resulting SNR=4. NGS maps were computed for segments S ={(kx,ky)|kya[k1,k1+n�1]}. First row: n = 1 (one phase-encoding line), second row: n = 32.
W. Lin, H.K. Song / Magnetic Resonance Imaging 24 (2006) 751–760756
FLASE sequence [15] with the following parameters was
used: TR/TE, 80/7.8 ms; 7.0�3.5 cm FOV; 0.5 mm section
thickness; 512�256�32 matrix (137�137�500 Am3 reso-
lution); 1208 excitation flip angle; F16 kHz bandwidth;
10.9 min scan time. An alternating navigator echo was
incorporated at the end of each readout to recover both
readout and phase-encoding direction translations [16],
allowing the comparison of our autofocusing techniques
with motion compensation using the navigators. For auto-
Fig. 5. The effects of background and metric selections on the metric map pattern
ky =k1 line. Two step shifts were applied on each half of the k-space for data
(Dx,Dy)= (�3,0) pixels for kya[�128,�k1] and kya[k1,127], respectively. Meth
focusing, an inverse FFT along the z-direction was first
performed to separate the slices, and data from 8 central
(out of 32) slices were used for 2-D in-plane motion
compensation by summing up their NGS values. Since
TR=80 ms and since 32 slices were acquired, the temporal
resolution of motion detection was effectively 2.5 s. In
addition to these acquisitions, several additional motion-free
scans in vivo were collected to help evaluate the properties
of the metric maps.
. The image used is the one shown in Fig. 2A. The segment S contains the
that have not been corrected. The resulting motion is (Dx,Dy)= (4,4) and
ods 1–4 are those shown in Fig. 3. (A) Entropy maps. (B) NGS maps.
Fig. 6. Comparison of image quality of different autofocusing schemes in a
phantom experiment with manually applied translations. (A) Motion-free
control image. (B) Motion-corrupted image. (C) Entropy, BG1 (Back-
ground Selection Method 1), proposed two-step search. (D) NGS, BG1,
1-D successive search. (E) NGS, BG3, two-step search. (F) NGS, BG1,
two-step search. (G) Detected displacements for Panel F. In Panel G, ad-
justments (additional shifts) were made to the detected Dy at some kylocations, such that while the phases applied to those lines were unchanged,
the detected displacements were more consistent with the expected
piecewise motion. This causes Dy to appear sloped at those locations.
W. Lin, H.K. Song / Magnetic Resonance Imaging 24 (2006) 751–760 757
On a 3.0-GHz Pentium computer (with 1 GB RAM), the
total processing time for the phantom experiment was about
100 s using the following parameters: 4 phase encoding lines
per segment near k-space center, 8 lines per segment near
k-space edge, (dx,dy) search range of (F10,F10) pixels and
precision of (1/4,1/4) pixel. Larger segment sizes were used
in the outer k-space regions due to their lower signal content
to improve the robustness. For the in vivo experiments, the
total processing times were about 5 min using these
parameters: 4 or 8 lines per segment near k-space center,
16 lines per segment near k-space edge, (dx,dy) search rangeof (F4,F4) pixel and precision of (1/4,1/4) pixel. NGS was
used as the image metric.
The effectiveness of the autofocusing algorithm was
evaluated in three ways. First, the motion-corrected images
were inspected visually and compared with the motion-
corrupted images and either a motionless image (phantom)
or images corrected using navigator echoes (in vivo).
Second, a measure of image improvement (relative to the
navigator technique) for the in vivo study was computed
according to the following equation:
Maf �Mcorrupted
Mnavigator �Mcorrupted
: ð17Þ
Here, Mnavigator, Mcorrupted and Maf are the metric values for
navigator-corrected, corrupted and autofocusing-corrected
images, respectively, evaluated over the eight central slices
where autofocusing was carried out. Finally, the derived
motion trajectory was compared with the navigator-
detected motion.
4. Results
4.1. The metric map pattern
Metric maps were computed for several motion-free
phantom and in vivo MR images. Despite variations in the
image features and metrics used, all metric maps share
similar patterns as in Fig. 2. It was found that metric maps of
segments that are near the k-space center or which consists
of more views, both of which have more image energy, have
a higher absolute range of signal intensity (peak minus
valley) and are, therefore, less susceptible to noise. The
latter phenomenon can be seen in Fig. 2, in which the
regions outside the dx=0 line are more homogeneous (less
noisy) for larger n and smaller k1.
One unexpected observation in the metric maps is the
presence of high-intensity maxima between the local
minima. This phenomenon could be explained by consid-
ering the interaction between the segment image IS and the
background image IB. Because IS contains finite spatial
frequencies, ringing will appear at the object’s edges along
the y-direction. The frequency of this modulation corre-
sponds to that of the missing spatial frequencies of the
background image IB. Local minima appear when the two
complex images are aligned such that the object’s edges are
sharpened, while, simultaneously, the ringing is reduced by
destructive interference. In between the minima, however,
the phases become aligned such that edges become blurred,
while ringing from two component images adds construc-
tively, resulting in local maxima.
At high SNR, a minimum occurs at the metric map’s
center, which corresponds to a motion-free state (Fig. 2,
SNR=25).When noise is added to reduce the image SNR and
when the segment used contains only a small number of
phase-encoding lines, one can no longer rely on ametricmini-
mum for proper motion compensation (Fig. 4A). Including
more phase-encoding lines to the segment S reduces the
undesired intensity fluctuations in the image metric map and
yields a more predictable pattern from which the proper
minimum can be extracted (Fig. 4B). This example visually
demonstrates the necessity of using larger segment sizes in
low SNR conditions, particularly in the outer k-space regions.
4.2. Background selection and image metric
Metric map patterns, as shown in Fig. 5, reveal that
the background selection method can have a significant
impact on the effectiveness of autofocusing algorithms. If
Table 1
Comparison of metric values for different autofocusing schemes in the phantom experiment
Metric used for autofocusing Optimization strategy Background selection method (Fig. 3) Image Quality measure
Entropy NGS
Motion free Fig. 6A 204.5 �19.72
Motion corrupted Fig. 6B 335.9 �7.37
Autofocused Entropy Proposed two-step search 1 Fig. 6C 327.4 �7.52
NGS 1-D successive search 1 Fig. 6D 332.6 �7.93
NGS Proposed two-step search 3 Fig. 6E 335.4 �7.86
NGS Proposed two-step search 1 Fig. 6F 323.0 �10.13
Lower values (more negative) indicate better motion compensation.
W. Lin, H.K. Song / Magnetic Resonance Imaging 24 (2006) 751–760758
autofocusing was to proceed in a center-out fashion in
k-space, among the four different background selection
methods, only Method 1 uses exclusively previously
motion-corrected data in background B. To simulate the
effects when B contains uncorrected motion, we applied
two step shifts on each half of the k-space for data that
have not been corrected. Since segment S contains motion
(Dx,Dy)=(�3,0), a distinct oscillating band should ideally
appear at dx=�Dx=3. However, for both entropy and
NGS, Methods 2–4 all generate additional oscillating
bands that may interfere with the correct motion estima-
tion. The problem is likely to worsen with reduced image
SNR. Overall, NGS maps with Method 1 produce the most
distinct oscillatory pattern along dx=3, as well as the
smallest intensity variations outside this band. Based on
these findings, the background selection scheme of Method
1 in conjunction with NGS image metric was chosen for
our experiments. This proposed scheme, proceeding in a
center-out fashion and alternating between positive and
negative ky, is similar to that proposed by Atkinson et al.
[8] for entropy, with the exception that the conjugate
region of S is always excluded and k-space is always zero
filled to the full matrix size. These results underscore the
potential usefulness of the metric maps in helping to
develop an improved combination of image metric and
motion search procedure.
Fig. 7. Motion-corrupted (A), navigator-corrected (B) and autofocusing-corrected
detail. (D–F) A portion of the image with �2 magnification. The arrows in Panels
D due to motion corruption.
4.3. Phantom and in vivo experiments
The results of the phantom experiment comparing our
proposed technique with others are shown in Fig. 6. Only
the image corrected with our proposed global optimiza-
tion procedure is of acceptable quality. The motion trajec-
tories determined from the proposed method are plotted in
Fig. 6G, showing several step motions along both x- and y-
directions, consistent with the actual motion applied during
the experiment. Table 1 shows the metric values of the
corrected images. Both entropy and NGS were computed
from the final corrected image, although only NGS was
used for the correction itself (except Fig. 6C).
The results from one of the three in vivo scans are shown
in Fig. 7. The image quality of the autofocused and
navigator-corrected images is similar and has significantly
improved compared with that of the motion-corrupted
images. The finer details of the trabecular structure are
more clearly visualized after motion correction. Such high
image quality is demanded in applications measuring bone
parameters, such as trabecular bone volume, thickness and
topology [15]. Table 2 shows the computed metric values of
the corrected images. Using the navigator correction as a
gold standard and based on the computed entropy and NGS
values, the relative image quality improvement after
autofocusing correction ranged from 0.72 to 0.90 in the
three subjects. There was no visually discernible difference
(C) images of the distal radius of Subject 1. Images were cropped to show
E and F show a trabecular element recovered, which was not visible in Panel
Table 2
Comparison of motion correction by autofocusing and navigator echoes for three high-resolution in vivo trabecular bone MR images
Subject Motion Image quality measure Motion corrupted Navigator corrected Autofocused Relative image quality improvement
1 Involuntary Entropy 1430.3 1418.2 1419.8 0.872
NGS �1.637 �1.671 �1.655 0.823
2 Involuntary Entropy 1613.9 1600.4 1602.9 0.818
NGS �1.521 �1.545 �1.539 0.756
3 Voluntary Entropy 1592.4 1584.0 1584.8 0.903
NGS �1.578 �1.596 �1.591 0.720
The image quality improvement relative to the navigator technique is defined by Eq. (17). Autofocusing was performed using the average NGS of eight slices
of the 3-D data set.
W. Lin, H.K. Song / Magnetic Resonance Imaging 24 (2006) 751–760 759
between the images corrected with either technique. These
results show that autofocusing is a very sensitive technique
able to compensate for motion in in vivo micro-MR
applications where SNR can be limited (8–10 in our
examples), yielding image quality improvements similar
to that of navigator echoes but without requiring addi-
tional data.
Motion trajectory recovered from the in vivo data is
compared with that determined from the navigator data in
Fig. 8. The trajectories from the two methods agree well,
although the errors tend to increase near the edges of
k-space due to lower SNR. It is also possible that the navi-
gator correction itself contains errors, as the acquired
navigator projections are gradient echoes with a long echo
time and are therefore noisy [16]. Potential through-plane
and rotational motion are additional sources of error not
addressed by the techniques implemented.
5. Discussion and conclusions
Although this work concentrates on entropy and NGS,
other metrics (such as correlation and gradient entropy [9])
were found to produce similar metric map patterns and,
therefore, could also be studied using the methodology
established here.
Fig. 8. Motion trajectories recovered from autofocusing applied on eight
slices of the data set and navigator echoes in Subject 1.
The computation time required by the proposed global
optimization strategy is highly predictable. In the first step
(the coarse 2-D search), the number of metric evaluations is
predetermined by the motion search range and the segment
position. The subsequent 1-D successive search typically
requires only two to three iterations since the global
minimum is already in proximity after the first step. In
contrast, a local searching approach might require a long
convergence time, depending on the position of the initial
trial motion vector on the entire metric map. In addition,
local searches may become trapped in a local minimum, as
evidenced by the presence of many local minima in the
metric maps, as shown in Figs. 2, 4 and 5. As previously
mentioned, high computational costs and the lack of
robustness are the two main drawbacks of current algo-
rithms. The improvements in both speed and robustness
presented in this work may help achieve the ultimate goal of
clinical applicability.
As discussed previously, it is necessary to use larger
segment sizes when SNR is low since otherwise noisy
metric maps would prevent reliable detection of the global
ig. 9. The effects of intrasegment motion on the metric map pattern. The
age used is the one shown in Fig. 2A. NGS maps (Background Selection
ethod 1) were computed for k -space segments S = {(k x ,k y )|
ya[k1,k1+7]}, with k1=8, 32 and 96. (A) A linear shift, with no motion
t ky =k1 and a maximal shift of (Dx,Dy)= (4,4) pixels at k =k1+7. (B) A
F
im
M
k
a
step shift of (Dx,Dy)= (4,4) pixels for kya[k1+4,k1+7].
W. Lin, H.K. Song / Magnetic Resonance Imaging 24 (2006) 751–760760
minimum (see Fig. 4). However, with larger segment size,
bintrasegmentQ motion may occur. Fig. 9 shows the effects
of such motion on the metric map pattern. In the first
example, a linear motion is present within segment S. The
maxima/minima oscillations (Fig. 9A) become less well
behaved, and the global minimum may potentially appear at
incorrect locations. If instead a sudden shift occurs in
segment S, a second set of local minima may appear
(Fig. 9B), particularly for large k1. These results show that it
is important to include a minimal number of ky lines in each
segment, as allowed by the SNR, in order to minimize the
influence of intrasegment motion.
Correction for combined in-plane rotation/translation has
been previously demonstrated [8], which involves applying
various trial rotations, in addition to translations. Generally
speaking, if motion other than in-plane translation occurs
during data acquisition and if only 2-D translation
correction was applied, then the resulting metric map
may not yield the expected patterns, leading to incorrect
results with the proposed strategies. However, if a
rotational motion is first compensated by applying a trial
rotation to the k-space segment, then it is anticipated that
the metric map would yield the expected pattern, permitting
the detection of the metric minimum that corresponds to the
motion-free state.
Like most existing techniques, the proposed autofocus-
ing motion compensation scheme does not address through-
plane motion or nonrigid deformations. Since through-plane
motion affects the excitation slice, it is not possible to fully
compensate retrospectively in a 2-D data set. Correction for
nonrigid deformations, on the other hand, may be possible,
but a more complex process beyond simple k-space phase
correction would be required [17]. For certain non-
Cartesian sampled data, it may be possible to mitigate the
effects of motion by considering data consistency [18].
Prospective gating techniques may also be used to mitigate
through-plane and nonrigid body motion [19], although
real-time signal processing and longer scan times are
often required.
This work introduces the concept of an image metric map
for the autofocusing optimization process. The metric map
pattern aids in the formulation of several significant
improvements to the existing autofocusing algorithm. The
maps were useful not only for determining the most relevant
image metric for motion correction and strategies to reduce
the overall processing time but also for developing a more
robust algorithm that is less susceptible to local minima.
Phantom experiments demonstrated that the proposed
algorithm is superior to previous methods. In clinical
high-resolution in vivo data corrupted with 2-D translational
motion, the improved autofocusing technique achieved
motion correction comparable with that of navigator echoes.
Acknowledgments
This work is supported by the NSF Grant BES-0302251.
The authors also wish to acknowledge Dr. Felix W. Wehrli
for useful discussions and for providing the in vivo images.
References
[1] Runge VM, Clanton JA, Partain CL, James Jr AE. Respiratory gating
in magnetic resonance imaging at 0.5 Tesla. Radiology 1984;
151:521–3.
[2] Bailes DR, Gilderdale DJ, Bydder GM, Collins AG, Firmin DN.
Respiratory ordered phase encoding (ROPE): a method for reducing
respiratory motion artifacts in MR imaging. J Comput Assist Tomogr
1985;9:835–8.
[3] Felmlee JP, Ehman RL. Spatial presaturation: a method for suppress-
ing flow artifacts and improving depiction of vascular anatomy in MR
imaging. Radiology 1987;164:559–64.
[4] Ehman RL, Felmlee JP. Adaptive technique for high-definition MR
imaging of moving structures. Radiology 1989;173:255–63.
[5] HedleyM,YanH, RosenfeldD.Motion artifact correction inMRI using
generalized projections. IEEE Trans Med Imaging 1991;10:40–6.
[6] Zoroofi RA, Sato Y, Tamura S, Naito H, Tang L. An improved method
for MRI artifact correction due to translational motion in the imaging
plane. IEEE Trans Med Imaging 1995;14:471–9.
[7] Atkinson D, Hill DL, Stoyle PN, Summers PE, Keevil SF. Automatic
correction of motion artifacts in magnetic resonance images using an
entropy focus criterion. IEEE Trans Med Imag 1997;16:903–10.
[8] Atkinson D, Hill DL, Stoyle PN, Summers PE, Clare S, Bowtell R,
et al. Automatic compensation of motion artifacts in MRI. Magn
Reson Med 1999;41:163–70.
[9] McGee KP, Manduca A, Felmlee JP, Riederer SJ, Ehman RL. Image
metrics-based correction (autocorrection) of motion effects: analysis
of image metrics. J Magn Reson Imaging 2000;11:174–81.
[10] Manduca A, McGee KP, Welch EB, Felmlee JP, Grimm RC, Ehman
RL, et al. Autocorrection in MR imaging: adaptive motion correction
without navigator echoes. Radiology 2000;215:904–9.
[11] Manduca A, Kiessel LM, Lake DS, Ehman RL. Automatic retrospec-
tive translational motion correction in image space. Proc ISMRM
2003;11:1058.
[12] PressWH, Teukolsky SA,VetterlingWT, FlanneryBP.Numerical recipes
in C, 2nd ed. NewYork7 Cambridge University Press; 1992. p. 394–455.
[13] Gull SF, Daniell GJ. Image reconstruction from incomplete and noisy
data. Nature 1978;272:686–90.
[14] Elder JH, Zucker SW. Local scale control for edge detection and blur
estimation. IEEE Trans Pattern Anal Mach Intell 1998;20:699–716.
[15] Wehrli FW, Hwang SN, Ma J, Song HK, Ford JC, Haddad JG.
Cancellous bone volume and structure in the forearm: noninvasive
assessment with MR micro-imaging and image processing. Radiology
1998;206:347–58.
[16] Song HK, Wehrli FW. In vivo micro-imaging using alternating
navigator echoes with applications to cancellous bone structural
analysis. Magn Reson Med 1999;41:947–53.
[17] Batchelor PG, Atkinson D, Irarrazaval P, Hill DL, Hajnal J, Larkman
J. Matrix description of general motion correction applied to multishot
images. Magn Reson Med 2005;54:1273–80.
[18] Pipe JG. Motion correction with PROPELLER MRI: application to
head motion and free-breathing cardiac imaging. Magn Reson Med
1999;42:963–9.
[19] Ward HA, Riederer SJ, Grimm RC, Ehman RL, Felmlee JP, Jack Jr
CR. Prospective multiaxial motion correction for fMRI. Magn Reson
Med 2000;43:459–69.