JOURNAL OF MAGNETIC RESONANCE IMAGING 24:444 – 450 (2006)
Technical Note
Accelerated Volumetric MRI With a
SENSE/GRAPPA Combination
Martin Blaimer, MS,1,4* Felix A. Breuer, MS,1 Nicole Seiberlich, BS,1
Matthias F. Mueller, MS,1 Robin M. Heidemann, MS,1 Vladimir Jellus, PhD,2
Graham Wiggins, PhD,3 Lawrence L. Wald, PhD,3 Mark A. Griswold, PhD,1,4
and Peter M. Jakob, PhD1
Purpose: To combine the specific advantages of the gener-
alized autocalibrating partially parallel acquisitions
(GRAPPA) technique and sensitivity encoding (SENSE) with
two-dimensional (2D) undersampling.
Materials and Methods: By splitting the 2D reconstruction
process into multiple one-dimensional (1D) reconstructions,
the normal 1D GRAPPA method can be used for image recon-
struction. Due to this data-handling process, a GRAPPA re-
construction is performed along the phase-encoding (PE) di-
rection and effectively a SENSE reconstruction is performed
along the partition-encoding (PAE) direction.
Results: In vivo experiments demonstrate the successful im-
plementation of the SENSE/GRAPPA combination. Experi-
mental results with up to 9.6-fold acceleration using a proto-
type 32-channel receiver head coil array are presented.
Conclusion: The proposed SENSE/GRAPPA combination
for 3D imaging allows the GRAPPA method to be applied in
combination with 2D undersampling. Because the SENSE/
GRAPPA combination is not based on knowledge of spatial
coil sensitivities, it should be the method of choice when-
ever it is difficult to extract the sensitivity information.
Key Words: 3D MRI; phased array; parallel imaging;
SENSE; GRAPPA
J. Magn. Reson. Imaging 2006;24:444 – 450.
© 2006 Wiley-Liss, Inc.
PARALLEL MRI (pMRI) (1–12) allows a significant re-
duction in scan time in an elegant way by using the
1
Department of Experimental Physics 5, University of Würzburg, Würz-
burg, Germany.
2
Siemens Medical Solutions, Erlangen, Germany.
3
Department of Radiology, Harvard Medical School, Boston, Massachu-
setts, USA.
4
Department of Radiology, University Hospitals of Cleveland and Case
Western Reserve University, Cleveland, Ohio, USA.
Contract grant sponsor: Deutsche Forschungsgemeinschaft; Contract
grant number: DFG JA 827/4.
*Address reprint requests to: M.B., Department of Radiology, Case
Western Reserve University, Bolwell Building Room 140, 11100 Euclid
Ave., Cleveland, OH 44106. E-mail: martin.blaimer@case.edu
Received June 23, 2005; Accepted April 13, 2006.
DOI 10.1002/jmri.20632
Published online 19 June 2006 in Wiley InterScience (www.interscience.
wiley.com).
© 2006 Wiley-Liss, Inc. 444
spatial information inherent in a multiple receiver coil
array in combination with dedicated reconstruction al-
gorithms. In the last few years two pMRI techniques—
sensitivity encoding (SENSE (10)) and generalized au-
tocalibrating partially parallel acquisitions (GRAPPA
(11))— have become commercially available. Today
many clinical applications benefit from accelerated ac-
quisitions due to pMRI. However, the performance of
pMRI is limited by the number of radiofrequency (RF)
coils in the array and the decrease in the signal-to-noise
ratio (SNR) due to undersampling. In general, the SNR
is reduced by at least the square root of the acceleration
factor. An additional decrease in SNR can be expected
by the geometry factor (10) of the underlying coil array.
The geometry factor analytically describes the spatially
varying noise amplification due to the SENSE recon-
struction, and is a measure of how easily the separation
of aliased pixels can be performed with a specific coil
array. Because the SNR is the limiting factor in pMRI,
applications with an already intrinsically low SNR are
not suited for acceleration with pMRI.
Combining parallel imaging with three-dimensional
(3D) Fourier encoding is advantageous for several rea-
sons. First, the intrinsic SNR in 3D encoding is higher
than in 2D imaging because a larger volume contrib-
utes to the signal. Second, the acquisition time of a fully
encoded 3D scan is long compared to a 2D scan. For
many applications a scan time reduction is desired.
Third, the reduction of phase-encoding (PE) steps can
be performed in two dimensions, thereby exploiting coil
sensitivity variations in two spatial dimensions (13).
This results in improved encoding efficiency and hence
higher scan time reductions.
In this study a combination of SENSE and GRAPPA
for 3D imaging is introduced. The computational ap-
proach presented here is based on a GRAPPA algorithm
for image reconstruction of data sets that are under-
sampled along two dimensions. The combination with a
SENSE-like reconstruction along the partition-encod-
ing (PAE) direction reduces the complexity of the com-
putation and allows one to use a standard GRAPPA
algorithm for 2D imaging. The SENSE/GRAPPA combi-
nation is not based on accurate knowledge of coil sen-
sitivity maps and is therefore beneficial whenever it is
3D MRI With a SENSE/GRAPPA Combination
Figure 1. a: Multiple equidistant partitions
from within the excited slab are superimposed
due to undersampling in the PAE direction. For
example, using a reduction factor RPAE ⫽ 2 in
this dimension, two partitions with distance
equal to half the slab thickness are superim-
posed. b: Additional in-plane aliasing occurs
due to the undersampling in PE direction (re-
duction factor RPE ⫽ 2 in this example) result-
ing in a reduced FOV. The total acceleration
factor in this example is R ⫽ RPAE ⫻ RPE ⫽ 4.
difficult to extract the coil sensitivity information, such
as in lung or abdomen imaging.
MATERIALS AND METHODS
Image Reconstructions
In 3D MRI phase encoding is performed along two spa-
tial directions: the PE direction and the PAE direction.
Therefore, undersampling can be carried out along both
dimensions. A reduction in the sampling density under
the Nyquist criterion along the PE direction results in
well known aliasing artifacts, which can be removed by
using pMRI methods.
Throughout this work we define the frequency-encod-
ing direction to be along the x-axis, the PE direction to
be along the y-axis, and the PAE direction to be along
the z-axis. Undersampling along the y-axis (PE direc-
tion) by a factor of RPE ⫽ 2 results in a reduced field of
view (FOV) and therefore in-plane aliasing artifacts,
which are well known from 2D imaging. Equivalently,
undersampling along the z-axis (PAE direction) by a
factor of RPAE ⫽ 2 results in aliasing artifacts along the
z-axis. Therefore, one partition of the undersampled
data set corresponds to a superposition of two parti-
tions located at positions z1 ⫽ z and z2 ⫽ z ⫹ ⌬S/2 from
within the excited slab of thickness ⌬S, as illustrated in
Fig. 1a. If the data are undersampled in both directions
(PE and PAE), one partition is composed of two super-
imposed partitions with additional in-plane aliasing ar-
tifacts due to the reduced FOV along the PE direction
(see Fig. 1b). The total acceleration factor is R ⫽ RPE ⫻
RPAE ⫽ 4.
445
The aliased partitions can be unfolded by using fully
Fourier-encoded low-resolution reference images to cal-
culate the reconstruction parameters. The reference
images can be acquired in an additional experiment
before or after the accelerated scan or during the scan
by using a variable density (VD) acquisition scheme (9).
For each coil the corresponding low-resolution refer-
ence images ␳1 and ␳2 at different positions z1 and z2 are
combined to form an extended 2D image matrix ␳⬘. The
low-resolution reference image of the first partition is
placed in the upper half of the extended matrix,
reordering
␳ 1 共x,y,z 1 兲O¡ ␳⬘(x,y),
and the low-resolution reference image (consisting of ny
k-space lines) of the second partition is placed in the
lower half of the extended image matrix, which is illus-
trated on the left side of Fig. 2a:
reordering
␳ 2 共x,y,z 2 兲O¡ ␳⬘(x,y⫹ny),
The extended 2D image matrix ␳⬘ contains informa-
tion from three dimensions because the PAE direction
is transformed into the PE direction.
The extended matrix is Fourier-transformed into k-
space in order to calculate the reconstruction parame-
ters using a conventional GRAPPA algorithm for 2D
imaging.
446
After the GRAPPA reconstruction and inverse Fourier
transformation, single-coil images containing the sep-
arated partitions are available (shown on the right side
of Fig. 2b). To obtain the final images, the single-coil
images can be combined, for example by using a sum-
of-squares reconstruction (14). After the reconstruction
process, the individual partitions are reordered to their
corresponding position.
The reconstruction procedure described above must
be applied to all aliased partitions in the undersampled
data set. Although GRAPPA is applied for image recon-
struction, effectively the reconstruction algorithm is
SENSE along the partition direction. In general, the
SENSE reconstruction can be characterized as an im-
age domain unfolding algorithm. Aliased pixels are un-
folded using the knowledge of the complex coil sensitiv-
ities at the corresponding positions. In the case of the
SENSE/GRAPPA combination, the partition direction is
in the image domain. The aliased partitions are un-
folded partition-wise using a GRAPPA reconstruction
algorithm and then reordered to their corresponding
position. Therefore, the reconstruction can be seen as
GRAPPA in the PE direction and SENSE in the partition
direction.
In Vivo Experiments
Several in vivo imaging experiments were performed on
a 3.0 T whole-body MR scanner equipped with 32-chan-
nel receiver technology (Siemens Medical Solutions, Er-
Blaimer et al.
Figure 2. Reconstruction scheme for a data
set reduced by a factor RPAE ⫽ 2 in the PAE
direction and RPE ⫽ 2 in the PE direction. a: To
calculate the coil weighting factors necessary
for the GRAPPA reconstruction, the corre-
sponding low-resolution reference images are
reordered into an extended matrix and then a
2D fast Fourier-transform (FT) is applied to
this matrix. b: Using the coil weighting factors
from image a, the superimposed partitions are
separated by a standard GRAPPA algorithm for
2D imaging. The individual partitions can then
be stored at their corresponding positions.
langen, Germany). Healthy volunteers were examined,
and informed consent was obtained before each study.
To investigate the performance of the SENSE/
GRAPPA combination, a fully-sampled 3D data set was
acquired using an eight-channel head coil array (MRI
Devices, Waukesha, WI, USA) with a T1-weighted mag-
netization prepared rapid gradient-echo sequence (TR/
TE/FA ⫽ 1630 msec/4.4 msec/15°, FOV ⫽ 220 ⫻ 220
mm2, slab thickness ⫽ 144 mm, 72 partitions, matrix ⫽
256 ⫻ 256). Image reconstruction was performed offline
using Matlab (The Mathworks, Natick, MA, USA). The
fully-sampled data set was reduced by a factor of RPE in
the PE direction (anterior–posterior) and by a factor of
RPAE in the partition direction (left–right), resulting in a
total reduction factor of R ⫽ RPE ⫻ RPAE. Low-resolution
reference images were obtained from the fully-sampled
data by extracting a matrix block, typically 48 parti-
tions with matrix size ⫽ 32 ⫻ 256, from the central
k-space. The reference data were only used to calculate
the reconstruction parameters and were not included
in the final image reconstruction.
Additionally, a 2D-SENSE reconstruction (13) was
performed for comparison. The coil sensitivity maps
were calculated by using coil array correlation statistics
(15). To ensure accurate coil sensitivity maps, the size
of the reference data was chosen to be 64 ⫻ 256, 64
partitions.
An additional accelerated experiment was performed
using a prototype 32-channel receiver head coil array
3D MRI With a SENSE/GRAPPA Combination
Figure 3. Results from the 3.0 T whole-body system equipped with an eight-element head-coil array. Two representative
partitions from the fully-sampled data set are shown (left side). The same partitions reconstructed with the SENSE/GRAPPA
combination with acceleration factor R ⫽ RPAE ⫻ RPE ⫽ 2 ⫻ 2 ⫽ 4 (middle) and acceleration factor R ⫽ RPAE ⫻ RPE ⫽ 2 ⫻ 3 ⫽ 6
(right side) are shown.
(16) and a T1-weighted gradient-echo sequence (TR/
TE/FA ⫽ 12 msec/4.7 msec/15°, FOV ⫽ 256 ⫻ 256
mm2, slab thickness ⫽ 160 mm, 80 partitions, matrix ⫽
256 ⫻ 256). The acquisition was accelerated in the PE
direction by a factor of RPE ⫽ 3.2 and in the PAE direc-
tion by a factor of RPAE ⫽ 3, resulting in a total acceler-
ation factor of R ⫽ 9.6. For calibration of the recon-
struction process, a low-resolution reference data set is
necessary. In this case, a fast spin-density-weighted 3D
fast low-angle shot (FLASH) experiment (matrix ⫽ 48 ⫻
64, 36 partitions, acquisition time ⫽ 6 seconds) was
performed prior to the accelerated experiment to ac-
quire the reference data (17). The total experiment time
for this examination, including the acquisition of the
reference data, was 31 seconds. The reader should note
that different contrasts were successfully used in the
actual accelerated experiment and the low-resolution
reference data.
For all experiments a modified GRAPPA algorithm was
used for image reconstruction (18,19). In our implemen-
tation a kernel of four neighboring points along the PE
direction and three neighboring points along the frequen-
cy-encoding direction from all coils were combined to re-
construct a missing data point in a single coil.
RESULTS
In Fig. 3 two representative partitions after pMRI recon-
struction using the SENSE/GRAPPA combination are
447
shown. Images with total acceleration factors of R ⫽ 4
(Fig. 3, middle) and R ⫽ 6 (Fig. 3, right side) are pre-
sented. On the visual scale the images are of good qual-
ity. No artifacts were introduced by the reconstruction.
For comparison, the corresponding fully Fourier-en-
coded images are shown (Fig. 3, left side).
A comparison of the SENSE/GRAPPA combination
with 2D-SENSE at six-fold acceleration is shown in Fig.
4. The 2D-SENSE approach (Fig. 4a) is based on SENSE
and therefore results in the best possible reconstruc-
tion with optimized SNR (10). The SENSE/GRAPPA
combination (Fig. 4b) is based on the GRAPPA approach
and is a close approximation of the optimum recon-
struction. In this example, on the visual scale, the im-
ages obtained with the SENSE/GRAPPA combination
are of similar quality compared to the results from the
2D-SENSE approach. Figure 4c provides a subtraction
of the 2D-SENSE result from the SENSE/GRAPPA re-
sult for the same partition in order to make differences
visible. For better visualization the subtraction image
was multiplied by a factor of 5. The g-factor map for this
partition (Fig. 4d) indicates that a six-fold acceleration
is marginal for this eight-channel coil array since the
SNR is decreased by a factor of 7 in some regions within
the head.
In Fig. 5 images from a T1-weighted gradient-echo
experiment are shown in sagittal, transverse, and coro-
nal views. Using the prototype 32-channel head-coil
array, an acceleration factor of R ⫽ 9.6 was achieved.
448 Blaimer et al.

Figure 4. Comparison of 2D-SENSE (a)

with the SENSE/GRAPPA combination
(b) at six-fold acceleration using an
eight-channel head-coil array. The re-
sult from the SENSE/GRAPPA combina-
tion (b) is a close approximation to the
result using the 2D-SENSE approach
that provides the ideal reconstruction
with optimized SNR. On the visual scale
both methods show similar image qual-
ity. c: Subtraction of the 2D-SENSE re-
sult from the SENSE/GRAPPA result.
For better visualization the subtraction
image is multiplied by a factor of 5. d:
The g-factor map for this partition.

This example demonstrates that it is feasible to use the
SENSE/GRAPPA combination with high acceleration
factors.
DISCUSSION
In this work a combination of the SENSE and
GRAPPA algorithms is presented. Similarly to 2D-
SENSE, undersampling is performed in two dimen-
sions in order to improve the reconstruction quality of
3D imaging experiments. The computational ap-
proach presented here splits the 2D reconstruction
process into multiple 1D reconstruction processes.
To every partition of the undersampled data set a
standard 1D-GRAPPA algorithm is applied in order to
unfold the aliased partitions. Finally, the individual
partitions are reordered at their corresponding posi-
tions. Effectively, the reconstruction is GRAPPA along
the PE direction and SENSE along the partition di-
rection, because aliased pixels are resolved partition
by partition. As in normal SENSE, the acceleration
factor in PAE direction can be a non-integer number.

Figure 5. Experimental results obtained with a prototype 32-channel head-coil array. Transverse (left), sagittal (middle), and
coronal (right) views of a T1-weighted 3D gradient-echo experiment are shown. The experiment was accelerated in the PE
direction (anterior–posterior) by a factor of RPE ⫽ 3.2, and the PAE direction (left–right) by a factor of RPAE ⫽ 3, resulting in a total
acceleration factor of R ⫽ 9.6. Low-resolution spin-density-weighted reference images for calculating the reconstruction
parameters were acquired in an additional experiment prior to the actual experiment. The total experiment time was 31 seconds,
including the acquisition of the low-resolution reference images.
3D MRI With a SENSE/GRAPPA Combination
For example, a 1.5-fold acceleration in this direction
results in a superposition of two partitions with a
distance of 0.75 䡠 ⌬S from each other, with the slab
thickness ⌬S. The superimposed partitions can be
separated if corresponding reference images are
available. In this example, only a fraction of the data
would have to be unfolded, since not all partitions
suffer from aliasing.
The SENSE/GRAPPA combination is based on the
GRAPPA method and allows one to take advantage of its
specific benefits. The reconstruction does not require
coil sensitivity maps and therefore the SENSE/
GRAPPA-combination is advantageous for applications
in which it is difficult to obtain sensitivity maps. Espe-
cially in inhomogeneous regions with low spin density,
such as the lung or abdomen, and in EPI experiments in
regions with severe distortions, the GRAPPA method
has been shown to be quite successful (see for example
Ref. 20). Therefore, the SENSE/GRAPPA-combination
should be the method of choice whenever acceleration
in two dimensions is desired for 3D extensions of these
applications.
In this work we present the results from head im-
aging experiments. Using a 32-channel head-coil ar-
ray, the scan time for whole-brain coverage with an
in-plane resolution of 1 ⫻ 1 mm2 and 2 mm slice
thickness was reduced from about four minutes with-
out acceleration by pMRI down to 31 seconds with
parallel imaging, including the acquisition of the ref-
erence data.
As in all pMRI methods, reference measurements
for calibration of the reconstruction process are nec-
essary. In this work an additional low-resolution 3D
prescan was used to acquire the reference data. In
principle, a prescan with any arbitrary imaging con-
trast can be used for the GRAPPA reconstruction of
pMRI experiments with totally different imaging con-
trasts (17). In this work we chose a fast spin-density-
weighted 3D gradient-echo sequence for the acquisi-
tion of the reference data. Thus, the acquisition time
for the prescan was minimized and did not become
comparable to the scan time of the accelerated exper-
iment. The resolution of the prescan data should be
large enough to appropriately register any coil sensi-
tivity variations along the PAE direction and to avoid
Gibbs’ artifacts introduced by a Fourier transform of
the low-resolution reference data. In our experience,
32– 48 partitions with a matrix size of 32 ⫻ 64 are
typically sufficient to achieve good results. However,
especially for coil arrays with many small elements
and high acceleration factors, the resolution of the
reference data may need to be increased.
Another approach for obtaining the reference data
is an extended variable density (VD) acquisition
scheme. Here the idea is to acquire the reference data
during the accelerated experiment. In a manner
equivalent to VD-AUTO-SMASH (9), the central k-
space is fully sampled while the outer parts of k-
space are undersampled along both PE directions.
The fully sampled k-space center is extracted to ob-
tain the low-resolution reference data, which are
used for calibration of the reconstruction process.
Furthermore, the reference data may be included in
449
the final reconstruction, which will improve the image
quality by reducing the residual artifact power and
increasing the SNR. A drawback of this strategy is the
decrease in data acquisition efficiency because the
total acquisition time is increased.
The SENSE/GRAPPA combination is easy to imple-
ment and allows the use of existing reconstruction
algorithms for 2D imaging. In principle, any other
current pMRI technique, such as SENSE or SMASH,
can be applied instead of the GRAPPA algorithm,
which makes this computational approach very flex-
ible. Since a standard GRAPPA algorithm for 2D im-
aging is used in our implementation, the computation
time for a 3D imaging experiment can be roughly
estimated as the GRAPPA reconstruction time for a
2D image multiplied by the number of partitions. The
computation time for reordering the reference images
is negligible.
In conclusion, the presented SENSE/GRAPPA com-
bination for 3D imaging allows one to take advantage
of the specific benefits of the normal GRAPPA method
in combination with 2D undersampling. The combi-
nation with a SENSE-like reconstruction along the
PAE direction reduces the complexity of the compu-
tation.
REFERENCES
1. Carlson JW. An algorithm for NMR imaging reconstruction based
on multiple RF receiver coils. J Magn Reson 1987;74:376 –380.
2. Hutchinson M, Raff U. Fast MRI data acquisition using multiple
detectors. Magn Reson Med 1988;6:87–91.
3. Kelton JR, Magin RL, Wright SM. An algorithm for rapid image
acquisition using multiple receiver coils. In: Proceedings of the 8th
Annual Meeting of SMRM, Amsterdam, The Netherlands, 1989.
Abstract 1172.
4. Carlson JW, Minemura T. Imaging time reduction through multiple
receiver coil data acquisition and image reconstruction. Magn Re-
son Med 1993;29:681– 687.
5. Ra JB, Rim CY. Fast imaging using subencoding data sets from
multiple detectors. Magn Reson Med 1993;30:142–145.
6. Kwiat D, Einav S. Preliminary experimental evaluation of an
inverse source imaging procedure using a decoupled coil detector
array in magnetic resonance imaging. Med Eng Phys 1995;17:
257–263.
7. Sodickson DK, Manning WJ. Simultaneous acquisition of spatial
harmonics (SMASH): fast imaging with radiofrequency coil arrays.
Magn Reson Med 1997;38:591– 603.
8. Jakob PM, Griswold MA, Edelman RR, Sodickson DK. AUTO-
SMASH: a self-calibrating technique for SMASH imaging. Simul-
taneous acquisition of spatial harmonics. MAGMA 1998;7:42–
54.
9. Heidemann RM, Griswold MA, Haase A, Jakob PM. VD-AUTO-
SMASH imaging. Magn Reson Med 2001;45:1066 –1074.
10. Pruessmann KP, Weiger M, Scheidegger MB, Boesiger P. SENSE:
sensitivity encoding for fast MRI. Magn Reson Med 1999;42:952–
962.
11. Griswold MA, Jakob PM, Heidemann RM, et al. Generalized auto-
calibrating partially parallel acquisitions (GRAPPA). Magn Reson
Med 2002;47:1202–1210.
12. Kyriakos WE, Panych LP, Kacher DF, et al. Sensitivity profiles from
an array of coils for encoding and reconstruction in parallel (SPACE
RIP). Magn Reson Med 2000;44:301–308.
13. Weiger M, Pruessmann KP, Boesiger P. 2D SENSE for faster 3D
MRI. MAGMA 2002;14:10 –19.
14. Roemer PB, Edelstein WA, Hayes CE, Souza SP, Mueller OM. The
NMR phased array. Magn Reson Med 1990;16:192–225.
450
15. Walsh DO, Gmitro AF, Marcellin MW. Adaptive reconstruction of
phased array MR imagery. Magn Reson Med 2000;43:682– 690.
16. Wiggins GC, Triantafyllou C, Potthast A, Reykowski A, Nittka M,
Wald LL. A 32 channel receive-only phased array head coil for 3T
with novel geodesic tiling geometry. In: Proceedings of the 13th
Annual Meeting of ISMRM, Miami Beach, FL, USA, 2005. Abstract
679.
17. Breuer F, Blaimer M, Mueller M, Heidemann RM, Griswold M,
Jakob P. Autocalibrated parallel imaging with GRAPPA using a
single prescan as reference data. In: Proceedings of the 21st
Blaimer et al.
Annual Meeting of ESMRMB, Copenhagen, Denmark, 2004. Ab-
stract 398.
18. Griswold MA. Advanced k-space techniques. In: Proceedings of the
2nd International Workshop on Parallel MRI, Zurich, Switzerland,
2004. Abstract 16.
19. Wang Z, Wang J, Detre JA. Improved data reconstruction method
for GRAPPA. Magn Reson Med 2005;54:738 –742.
20. Heidemann RM, Griswold MA, Kiefer B, et al. Resolution enhance-
ment in lung 1H imaging using parallel imaging methods. Magn
Reson Med 2003;49:391–394.