Synopsis
Adaptive reconstruction (AR) can be used to
combine multichannel images without acquiring coil sensitivity information. It
can improve the SNRs of combined magnitude and phase images. But the
reconstructed phase images may suffer from open-ended fringe artifacts when
coil-sensitivity maps vary a lot. We propose a two-step adaptive reconstruction
method to combine multichannel images. This method is shown to be more robust
than the traditional AR method and can still maintain the high SNR.Purpose
Channel combination without acquiring
coil sensitivity maps can be challenging for phase images. Since different
coils are subject to different sensitivities, the phase images acquired by different
channels vary a lot. A direct Weighted Mean (WM) combination
1 ($$$\theta_{combined}=\sum_kM_k^2exp(i\theta_k)/\sum_kM_k^2$$$,
$$$k$$$ represents channel index, $$$M_k$$$ and $$$\theta_k$$$ are magnitude
and phase of channel $$$k$$$) may lead to signal cancellation and decrease the SNRs of combined
images. Adaptive reconstruction
2 (AR) can improve SNR, but it is sensitive
to patterns and variations of sensitivity maps, which may lead to open-ended
fringe artifacts (Fig. 1f). In this study, we demonstrate a two-step way to estimate
and eliminate the phase terms introduced by coil sensitivities and improve the
robustness of traditional AR method.
Methods
Signals acquired from different channels can
be modeled as $$$C_k=M\cdot S_k\cdot
exp(i(\theta+\phi_k+\eta_k))$$$,
$$$k$$$ represents the
$$$k^{th}$$$ channel, $$$C_k$$$: acquired complex images, $$$S_k$$$ and
$$$\phi_k$$$: magnitude and phase of coil sensitivity, $$$M$$$ and
$$$\theta$$$: magnitude and phase of underlying signals, $$$\eta_k$$$:
noise term. The huge variations of $$$\phi_k$$$ and the singularities contained in $$$\phi_k$$$
can both influence the combined phase image potentially. So $$$\phi_k$$$ should be removed
before AR.
We combine the different channel data in
two steps:
Step 1: Estimate $$$\phi_k$$$. a) The phase images from the original
data are high-pass filtered to remove $$$\phi_k$$$ ($$$\phi_k$$$ is
supposed to be a low-frequency component) and combined using the WM
method to generate a rough combined phase map $$$\widehat{\theta}$$$ 3. b) The original
multichannel phase images are subtracted by $$$\widehat{\theta}$$$ to generate rough $$$\phi_k$$$ maps.
c) $$$\phi_k$$$ maps are smoothed to suppress noise.
Step 2: Remove $$$\phi_k$$$ from the original
data, and then perform an AR operation.
Phantom and in-vivo studies were conducted on a Philips 3.0T Achieva TX MRI
scanner (Philips Healthcare, Best, The Netherlands). A 32-channel head coil was
used, with smart selection on. Human studies were performed under IRB approval
from our institution. Phantom study: 2D multiecho gradient echo, 4 echoes with TR/TE1/△TE = 447/3.6/5.0ms, flip angle = 18°, FOV =
230×230mm2, Matrix = 256×256, slice thickness = 2mm, slice number =
20; coil elements used = 28. In-vivo
study: 2D multiecho gradient echo, 4 echoes with TR/TE1/△TE = 895/3.6/5.0ms, flip angle = 18°, FOV =
230×230mm2, Matrix = 256×256, slice thickness = 2mm, slice number =
40, coil elements used = 30.
These two datasets were processed using three
methods: WM, the traditional AR and our proposed method, to test the robustness
and SNR performance of these algorithms.
Results
Fig. 1 shows the results for the phantom
study. The singularities shown in Fig. 1a exist in all the phase images acquired by this channel, indicate that these singularities are
introduced by coil-dependent phase term $$$\phi_k$$$. $$$\phi_k$$$ maps (Fig. 1c) are
estimated from the original phase maps (Fig. 1b) and removed away (Fig. 1d).
It is obvious that there is open-ended
fringe artifact in the result of AR (circle in Fig. 1f). It has been eliminated
by our method (Fig. 1g). The standard deviations of the same regions (yellow
boxes) from e to g are 0.147, 0.064 and 0.044, respectively, indicate the high
SNR performance of the traditional and our proposed methods. The different standard deviations between f and g may arise from the
elimination of low-frequency components in Step 1.
Fig. 2 shows the results for the in-vivo study. The WM method provides
results with low SNR (Fig. 2a). AR can get a relative high SNR compared with the
WM method. Both the WM and AR methods generate a phase singularity shown in
the circled areas. These singularities lead to the errors in the unwrapped
phase and high-pass filtered phase images (Figs. 2d, 2e, 2g and 2h).
In comparison, our proposed method can eliminate this singularity and maintain
the high SNR (Figs. 2c, 2f and 2i).
Discussion
AR processing of phase images may suffer
from the variations and singularities of sensitivity phase maps and lead to open-ended fringe artifacts. These artifacts may hinder the subsequent phase unwrapping
and phase usage. Our two-step way estimates the phase maps of coil sensitivities firstly and eliminates these phase terms thereafter. Then a repeated AR
can combine phase images with a high SNR and free of singularity.
Conclusion
The two-step way can improve the robustness
of traditional AR method to combine phased-array signals while maintaining high
SNRs of the combined results.
Acknowledgements
The code for adaptive reconstruction was
provided by ISMRM 2012 workshop on fat-water separation and thanks for their contribution.References
1. Bernstein MA, Grgic M, Brosnan TJ, Pelc
NJ. Reconstructions of phase contrast, phased array multicoil data. Magn Reson
Med. 1994; 32:330–334.
2. Walsh DO, Gmitro AF, Marcellin MW.
Adaptive reconstruction of phased array MR imagery. Magn Reson Med.
2000;43:682–690.
3. Koopmans PJ, Manniesing R, Niessen WJ,
Viergever MA, Barth M.MR venography of the human brain using susceptibility
weighted imaging at very high field strength. Magn Reson Mater Phy 2008;21:149–158.
4. Ghiglia DC, Pritt MD, Two-Dimensional
Phase Unwrapping: Theory, Algorithms and Software. New York: Wiley
Interscience; 1998.