Mirco Grosser1,2 and Tobias Knopp1,2
1Section for Biomedical Imaging, University Medical Center Hamburg-Eppendorf, Hamburg, Germany, 2Institute for Biomedical Imaging, Technical University Hamburg, Hamburg, Germany
Synopsis
Reconstruction
methods for multi-contrast MRI often employ a low-rank constraint to
reconstruct images from highly undersampled data. For multi-echo
gradient echo sequences such a constraint is hard to impose due
off-resonance-induced oscillations of the time signals. In this work
we show that spatio-temporal correlations can be exploited
efficiently using locally low rank regularization. Based on this
observation we develop a locally-low rank regularized reconstruction
scheme and test it on a multi-echo gradient echo dataset of a human
brain. In our tests the proposed method shows significantly improved
image quality compared to regular compressed sensing reconstructions.
Introduction
Multi-echo gradient echo
(MEGRE) sequences are a common tool in MRI with a variety of
applications such as R2*-
and quantitative susceptibility mapping (QSM). For other
multi-contrast MRI sequences, strong acceleration can be achieved by
exploiting correlations between the temporal signals for different
tissue parameters. This translates into a low-rank constraint for the
Casorati matrix of the images to be determined [1]. As a matter of
fact, aforementioned low rank-property does not hold for MEGRE images
due to the B0-induced
time evolution of the image phase. Here we demonstrate that
spatio-temporal correlations can be exploited when considering local
image patches. As a result, we formulate image reconstruction as a
locally low rank regularized (LLR) inverse problem [2,3] and show its
effectiveness in comparison to standard compressed-sensing
reconstructions.Methods
In MEGRE MRI , the magnitude
signal in a given voxel shows an approximately exponential decay.
However, B0-inhomogeneity
leads to a temporally oscillating phase, which makes it hard to find
a low rank representation for the different signal evolutions in the
image series. On the other hand, voxels within a small patch
typically experience similar values of inhomogeneity, which makes it
possible to find a low rank representation. To demonstrate this, we
used the in-vivo
brain dataset (measured on a GE scanner) accompanying the MEDI
toolbox [4,5]. Fig. 1 shows the normalized singular values of both
the global Casorati matrix and the Casorati matrix for two image
patches. The former also suggests that a representation with lower
rank can be found for signals similar to the magnitude signal.
Based on those observations,
we formulate the following reconstruction problem
$$ \underset{\mathbf{x}}{\operatorname{argmin}} \parallel \mathbf{y}-\mathbf{E}\mathbf{\Phi}\mathbf{x} \parallel_2^2 + \lambda\sum_{\mathbf{r}} L(R_{\mathbf{r}}\mathbf{x})$$
Here $$$\mathbf{E}$$$ denotes the signal
encoding operator and $$$\mathbf{y}$$$
denotes the measured k-space signal. Analogously
to [1], $$$R_{\mathbf{r}}$$$ extracts blocks around location $$$\mathbf{r}$$$
and reshapes them into the columns of the associated Casorati matrix.
As a surrogate of the rank, $$$L$$$ denotes the logdet penalty of the
singular-values, which was proposed in [4]. Although being non-convex
the latter was observed to show superior results compared to the
convex nuclear norm. Finally, we include the diagonal operator $$$\mathbf{\Phi}$$$ which multiplies each entry in $$$\mathbf{x}$$$
with an estimate of the image phase for the associated voxel and
echo. As discussed before, this leads
to a more compact representation of the spatio-temporal signals. In
summary,
the proposed method starts by solving (1) for $$$\mathbf{\Phi}_0 = \mathbf{1}$$$.
Then
the solution can be refined by updating the phase estimate $$$\mathbf{\Phi}_1$$$ and solving the new form of problem (1).
In order to test the proposed
method, the complex-valued brain
dataset used before was Fourier transformed into k-space. The data
was retrospectively undersampled using a variable density poisson
disk pattern with a fully sampled calibration area of size 24x24.
Finally, the multi contrast image series was reconstructed using the
proposed method with and without updating the phase estimate. For
comparison we also performed a compressed sensing (CS) reconstruction
with ℓ1
regularization in the Wavelet domain.Results
Fig. 2. shows the
reconstructed magnitude images for the first and the last echo of the
image series. The Wavelet-based CS reconstruction shows blurring and
severe artifacts. For the proposed method, the reconstruction quality
is increased. This is also reflected in the normalized root mean
squared deviation (NRMSD), which is shown
in the figure for each contrast. Incorporating the image phase into
the imaging operator, leads to a further reduction in NRMSD and a
reduced amount of perceivable noise.Discussion
Our results demonstrate that
LLR regularization is a powerful prior for the reconstruction of
MRGRE image data. This can be justified by the fact that the
underlying B0
map typically has a smooth structure. As a consequence neighboring
voxels show a similar oscillatory behavior. By incorporating a prior
phase estimate into the imaging operator, the images to reconstruct
show more resemblance to the magnitude signals, which explains the
additional increase in reconstruction quality. Note that the
reconstruction problem (1) can be solved efficiently using ADMM [6].
Moreover, the successive reconstructions can be initialized with the
previous image estimate. Thus, obtaining a phase estimate only leads
to a moderate increase in reconstruction time.Conclusion
A
new image reconstruction was developed, which exploits signal
correlations through LLR regularization. In particular, this method
is suited for MEGRE sequences, where the image signals are no longer
contained in a global low rank subspace. In comparison to standard CS
methods, it allows successful image reconstruction at larger
undersampling factors. Thus, the proposed method has the potential to
further accelerate MRI applications such as R2*-mapping
and QSM.Acknowledgements
No acknowledgement found.References
[1]
Velikina, J., et al., MRM, 70(5), 1263-1273, (2012)
[2]
Zhang, T., et al., MRM, 73(2), 655-661, (2015)
[3]
Tamir, J., et al., MRM, 77(1), 180-195, (2017)
[4]
Wang, Y., & Liu, T. , MRM, 73(1), 82-101, (2015)
[5]
http://pre.weill.cornell.edu/mri/pages/qsmreview.html
[6]
Boyd, S. et al., Foundations and Trends®
in Machine Learning, 3(1), 1-122,(2011)