Seon Young Shin1, JungHyun Song1, Yeji Han1, and Jun-Young Chung1
1Gachon Advanced Institute of Health Sciences and Technology, Gachon University, Incheon, Korea, Republic of
Synopsis
To compare the volumetric GRAPPA algorithms in the presence of
physiological artifacts, five different algorithms were used, i.e.,
2D-GRAPPA-OP, 3D-GRAPPA, EX-3D-GRAPPA and SK-3D-GRAPPA. The performance of
algorithms were compared using the root mean squared error (RMSE) of the image reconstructed
from fully acquired 3D in-vivo k-space data and the image reconstructed using
different reconstruction algorithms from undersampled dataset. Introduction
When
3D MR data is acquired with undersampling in the phase-encoding and the
partition-encoding directions, the image can be reconstructed using volumetric
generalized auto-calibrating partially parallel acquisitions (GRAPPA)
algorithms such as 2D-GRAPPA-Operator (OP), 3D-GRAPPA, extension (EX)-3D-GRAPPA
and single kernel (SK)-3D-GRAPPA [1-4]. As demonstrated in the previous works, 3D-GRAPPA
methods generally showed better performance than 2D-GRAPPA-OP with regards to
aliasing artifacts [5-6]. However, the verification was performed for the
phantom data only, which do not include physiological artifacts such as motion
and blood flow, thereby limiting the scope of the verification. In this study,
investigation was carried out to show the performance of volumetric GRAPPA
algorithms in the presence of physiological artifacts.
Methods
Schematic diagrams for each
reconstruction algorithm are presented in Fig. 1. The 2D-GRAPPA-OP algorithm,
utilizing two 2D-kernels, is considered as two different reconstruction methods,
i.e., 2D-GRAPPA-OP-YZ and 2D-GRAPP-OP-ZY, according to the order of unaliasing
directions, i.e. in which direction the image is unaliased first (Fig. 1(a) and
(b)). 3D-GRAPPA uses three different kernels, consisting of two 2D- and one 3D-kernels
(Fig. 1(c)). EX-3D-GRAPPA uses three different 3D kernels (Fig. 1(d)).
SK-3D-GRAPPA utilizes only one 3D-kernel (Fig. 1(e)).
To demonstrate the
performance of these reconstruction algorithms, phantom and in-vivo human data
were obtained with a 3D gradient echo sequence from a 3T MRI scanner (Verio,
Siemens) equipped with a 12-channel head coil using the following parameters:
FOV = 210 × 210 mm2, slice thickness = 0.8mm, matrix size = 512 × 256 × 208, TR = 20ms, TE = 14ms, flip angle = 25º.
From the fully acquired 3D
k-space data, undersampling was performed with reduction factors of two for
both phase-encoding and partition-encoding directions and 24 auto-calibration
signal (ACS) lines were selected for each reconstruction algorithm. To
calculate the coefficients for estimation of the not-acquired k-space data, the
kernels were then selected as illustrated in the schematics of Fig.1 as
follows: 2D-GRAPPA-OP (3 × 4 × 1, 3 × 1 × 4), 3D-GRAPPA (3 × 2 × 2, 3 × 4 × 1,
3 × 1 × 4), EX-3D-GRAPPA (3 × 2 × 2, 3 × 2 × 3, 3 × 3 × 2) and SK-3D-GRAPPA (3 × 2 × 2). By utilizing the estimated k-space data, images were generated using
different reconstruction algorithms.
Results
In Figs. 2 and 3, the
images reconstructed from the phantom data and the in-vivo data are
respectively presented. The upper rows show the reconstructed images and the
bottom rows show the difference images. The difference images were calculated by
subtracting the images reconstructed from the undersampled data from the true
image reconstructed from the fully acquired 3D k-space data and. The contrast
of the difference images was adjusted for viewing purposes. As demonstrated by Figs. 2(a, b) and 3(a, b), the directions
of the dominant N/2 ghost varied according to the direction of initial
unaliasing in 2D-GRAPPA-OP. All 3D GRAPPA algorithms generally showed better
performance than 2D-GRAPPA-OP, with more improvements provided by EX-3D-GRAPPA
and SK-3D-GRAPPA. In Fig. 4, root mean squared error (RMSE) values were
plotted. For phantom data, the RMSE of EX-3D-GRAPPA and SK-3D-GRAPPA was
reduced by 0.0274 and 0.0166, respectively, compared to the RMSE of 2D-GRPPA-OP-YZ. For in-vivo data, the RMSE was reduced by 0.0348 and 0.0245. In general,
the 3D-GRAPPA algorithms showed larger amount of improvements with respect to
the noise figures in human data.
Discussion & Conclusions
For
both phantom and human experiments, 3D-GRAPPA reconstructions were generally
more effective than the 2D-GRAPPA-OP because 2D-GRAPPA algorithm cannot prevent
the accumulation of errors, which is inevitable as the second unaliasing is
performed by utilizing the k-space data estimated in the first unalising step.
More specifically, EX-3D- and SK-3D-GRAPPA algorithms showed better performance
for noise reduction because they use acquired data only by using 3D kernels for
estimation of k-space data. In case of the human data, which have additional
physiological effects, the results showed similar tendency across different
reconstruction algorithms as in the phantom data. However, as the RMSE plotted
in Fig. 4 demonstrated, the degree of improvements in 3D-GRAPPA compared to
2D-GRAPPA-OP was greater for human data, while EX-3D-GRAPPA exhibited the
lowest values of RMSE. This is partly because artifacts due to the
physiological effects were introduced in the in-vivo data and they influence
the image quality due to the accumulation of errors. Thus, it is suggested to
use EX-3D-GRAPPA for volumetric GRAPPA reconstruction of in-vivo images.
Acknowledgements
No acknowledgement found.References
[1] Breuer et al., MRM, 2006;56:1359-1364
[2] Griswold et al., MRM 2002;47:1202-10
[3] Wang et al., MRM 2005; 54:738-742
[4] F.Breuer et al, ISMRM (2006)
[5] Chung et al, ESMRMB, 2012(#763)
[6] Chung et al, ESMRMB, 2012(#765)
[7] S.Bauer et al, MRM 2011;66:402-409