Zheyuan Yi1,2,3, Yilong Liu1,2, Fei Chen3, and Ed X. Wu1,2
1Department of Electrical and Electronic Engineering, The University of Hong Kong, Hong Kong, China, 2Laboratory of Biomedical Imaging and Signal Processing, The University of Hong Kong, Hong Kong, China, 3Department of Electrical and Electronic Engineering, Southern University of Science and Technology, Shenzhen, China
Synopsis
In conventional compressed sensing (CS) multi-slice
Cartesian 2D imaging, the undersampling is performed along phase-encoding
direction only, leading to coherent 1D aliasing that significantly limits the
effectivess of CS for acceleration. This study proposes a multi-slice CS
reconstruction method to take advantage of extremely augmented sampling
incoherence created by orthogonally alternating phase-encoding directions among
adjacent slices. The multi-slice CS approach was evaluated with single-channel
brain T1W and T2W datasets. The results demonstrate significant improvements
with both pseudo-random and uniform undersampling. This new method will also
greatly augment the existing 2D CS parallel imaging technqiues for very high
acceleration.
Introduction
Compressed
sensing (CS)1 MRI reconstruction exploits the sparsity
of MR images, provides an important alternative to conventional parallel
imaging reconstruction. It requires incoherent random undersampling, which leads
to incoherent artifacts that can be iteratively eliminated by enforcing the
transform sparsity. Multi-slice 2D Cartesian imaging is the most routinely used
acquisition during clinical MRI scans, where undersampling can be only performed
along phase-encoding direction. Such acquisition scheme inherently limits the
image sparsity along the readout direction to be utilized in conventional
single-slice CS. Multi-slice reconstruction via CS has been attempted earlier
to explore the image sparsity (structural similarity) among slices1, where phase-encoding
undersampling pattern was varied between slices but phase-encoding direction
remained the same for all slices. However, this method only offered limited
improvement because severe coherence still and only existed along
phase-encoding direction. In this study, we propose a novel undersampling and
reconstruction strategy. It significantly improves the incoherence of multi-slice
Cartesian CS imaging by orthogonally alternating phase-encoding directions
(APE) among adjacent slices. A new multi-slice CS approach is developed to exploit
such enhanced sparsity along slice dimension. The proposed framework yield
excellent CS reconstruction with undersampling flexibility, including both the
conventional pesduo-random undersampling and the uniform undersampling (which
is widely used in present parallel imaging method). Method
Multi-Slice CS Reconstruction
To take
advantage of the sparsity across slices, conventional 2D CS1 is extended to multi-slice
reconstruction by further applying 1D wavelet transform together with
total-variation (TV) along the slice dimension, as shown in Figure 1. A lower regularization weight
was implemented in 1D wavelet transform, since slice thickness/gap is typically
larger than in-plane resolution. On the contrary, the additional TV is proposed
to have a higher weight to enhance the leakage of artifacts among slices. In
this study, the decomposition level of single-slice and multi-slice CS were set
to 4×4 and 4×4×1 with Daubechies 4 wavelets.
Sampling Schemes
Conventionally,
multi-slice 2D imaging with CS apply undersampling along the same direction for
all slices (Figure 2). Though
different slices can be undersampled differently, the incoherence among
different slices is limited if all slices have the same aliasing direction. In
our proposed approach, three different sampling patterns, including variable/uniform
density random undersampling, and uniform sampling patterns, is augmented by APE.
The point spread function (PSF)1 analysis was performed to evaluate
sampling incoherence of these patterns.
Data Preparation and Image
Reconstruction
Human brain
T1w and T2w data were acquired on a 3T Philips scanner with an 8-channel head
coil. For T1w data, a gradient echo sequence (TR/TE=600/10ms, FOV=240×240mm2,
and slice thickness/gap=4/1mm) For T2w data, a fast spin echo (FSE) sequence
(TR/TE=3000/113ms, FOV=240×240mm2, and slice thickness/gap=4/1mm)
and then retrospectively undersampled for evaluation. The multi-channel images
were combined using virtual body coil method2 to simulate single-channel data
with phase information reserved, and then retrospectively undersampled using
the aforementioned sampling with variable/uniform density random undersampling,
and uniform sampling patterns. All the these pattern were augmented with APE. Multi-slice
CS was performed on all 16 slices and 4 of them are shown.Results
As shown in
Figure 3, the multi-slice CS
reconstructs better image structure and edges corresponding to the incoherency
improvement examined by PSF in Figure 2.
The image details in wavelet domain are corrupted by higher sidelobes of
undersampling artifacts, thus become difficult to be recoverd in single-slice
CS. For random undersampling without densely sampled k-space central region,
multi-slice CS can still yield promising reconstruction because of the
significant augmentation of coherence from APE sampling approach. Note that
such augmentation even be effective when APE is applied with uniform
undersampling pattern (Figure 4).
Single-slice CS completely fails to remove extremely coherent artifacts, which
is corresponding to the same level of coefficient compared to original signals
in wavelet domain (Figure 2). With a
simple APE sampling strategy, multi-slice CS is flexible to handle the
uniformly undersampled single-channel data. Aliasing artifacts, that appeared
only along undersampling direction, are spreaded into other slices and form
pseudo-2D patterns through multi-slice CS (Figure
5). Discussion and Conclusions
This study present
a new acquisition and reconstruction strategy that can be easily applied to
routine clinical multi-slice Cartesian CS imaging. The proposed alternating phase-encoding
(APE) sampling greatly enhances the CS reconstruction by spreading artifacts
among slices and in two directions, thus augementing the transform sparsity. Here
we demonstrate that APE works with a variaty of sampling patterns, including
variable/random and uniform undersampling.
For
simplicity, this strategy is implemented for single-channel multi-slice MR data
only in this study. With multiple channels available in virtually all clinical
MRI scanners, our proposed strategy can be easily incorporated into existing
parallel imaging3 methods, and/or can be used to jointly
reconstruct multi-channel MRI, by enforcing group sparsity for all channels to
achieve very high acceleration. Furthermore, orthogonal APE as proposed in
this study spreads aliasing along two orthogonal directions. However, this
concept can be further extended by rotating the phase-encoding direction with different
angles within spatially or/and temporally adjacent slices, which is
conceptually similar to PROPELLER4 and GRASP5. Acknowledgements
This study
is supported in part by Hong Kong Research Grant Council (C7048-16G and
HKU17115116 to E.X.W.), Guangdong Key Technologies for Treatment of Brain
Disorders (2018B030332001) and Guangdong Key Technologies for Alzheimer's
Disease Diagnosis and Treatment (2018B030336001) to E.X.W.References
[1] Lustig M, Donoho D, Pauly JM. Sparse
MRI: The application of compressed sensing for rapid MR imaging. Magn Reson Med
2007;58(6):1182-1195.
[2] M B, P B, S.
K. Virtual body coil calibration for phased-array imaging. In: Proceedings 17th
ISMRM Scientific Meeting, Honolulu, 2009, p 759.
[3] Lustig M,
Pauly JM. SPIRiT: Iterative self-consistent parallel imaging reconstruction
from arbitrary k-space. Magn Reson Med 2010;64(2):457-471.
[4] Pipe JG.
Motion correction with PROPELLER MRI: application to head motion and
free-breathing cardiac imaging. Magn Reson Med 1999;42(5):963-969.
[5] Feng L, Grimm
R, Block KT et al. Golden-angle
radial sparse parallel MRI: combination of compressed sensing, parallel
imaging, and golden-angle radial sampling for fast and flexible dynamic
volumetric MRI. Magn Reson Med 2014;72(3):707-717.