Dong Wei Lu1, Zhi Hua Ren1, and Shao Ying Huang1
1EPD Pillar, Singapore University of Technology and Design, Singapore, Singapore
Synopsis
To
make a MRI system portable, a practical approach is applying Halbach magnet
array and nonlinear spatial encoding strategy. Here, the rotation of a magnet
array for imaging is replaced by electrically forming RF receive coil
aggregates with phase delay. For the resultant system with a new encoding
matrix, Truncated-Singular-Value-Decomposition with an
optimal regularization parameter is proposed which reconstructs images with
good quality. An accelerated L-curve method is proposed to obtain the optimal
regularization parameter. Results show that the proposed strategy provides
considerable improvement of the image quality compared to existing method, e.g.
Kaczmarz iteration, without rotating the magnet array.
INTRODUCTION
Current clinical Magnetic
Resonance Imaging (MRI) system is bulky, thus researchers are looking for
imaging approaches with significant size reduction to
let MRI possible in a wider range of working conditions. One realistic approach
of a portable MRI system is based on a rotating Halbach magnet array and
nonlinear spatial encoding strategy1,2. In this abstract, the rotation
of a magnet array for imaging is replaced by electrically forming RF receive coils aggregates with phase delay. The physical
setup leads to a new encoding matrix. Truncated Singular Value Decomposition
(TSVD) with an optimal regularization parameter is proposed to solve the
system, which reconstructs images with good quality. The optimal regularization
parameter is selected by the L-curve method with accelerated calculations. With
the same signal-to-noise (SNR) level, the proposed approach provides a significant improvement
of the reconstructed image quality.METHODS
Fig. 1(a) shows the inhomogeneous B0
field generated by a Halbach array for imaging. Eight RF loops are used. The
signal array is given as $$$\mathbf{s=Em}$$$ where $$$\mathbf{m}$$$ is the image and $$$\mathbf{E}$$$ is the encoding matrix containing the phase
encoded by B0 and the amplitude encoded by the coil sensitivity. Four coils chosen from the eight coils can be
turned on/off simultaneously and they are called supercoils. The total number
of supercoils is $$$C(8,4)=70$$$. The phase delay of the successive coils is set
to be $$$π/4$$$. The B1 field maps of one of the supercoils
is shown in Fig. 1(b). The encoding matrix for this portable system is
ill-conditioned. The effect of noise can be magnified severely and the image
may not be reconstructed by Least Mean Square (LMS) method. By discarding small
singular values of the encoding matrix which
can magnify the effect of noise, TSVD method is practical to solve such an ill-conditioned
problem. The encoding matrix can be decomposed as3: $$$\mathbf{E=U\sum V^{T}}$$$, $$$\mathbf{\sum} = diag(\sigma_1, \sigma_2,....\sigma_m)$$$. The pseudo-inverse
matrix can be reconstructed as:
$$\mathbf{m = E^+s = \sum_{i=1}^{k} v_iu_i^T}\sigma_i^{-1}, 1\leq k<m$$
The number $$$k$$$ is a regularization
parameter. In TSVD method, we only use $$$\sigma_i(1\leq i\leq k <m)$$$ which are relatively large singular values of the
encoding matrix and represent the main components of MR signal for image reconstruction.
With the system resulted by using supercoils, the regularization parameter can
be chosen by L-curve method4. L-curve describes the relationship
between $$$\mathbf{log||m||}$$$ and $$$ \mathbf{log||Em-s|}|$$$. The L-curve’s turning point can be an optimal
number for the regularization parameter. For obtaining a L-curve, a direct calculation
takes days. Here, the pseudo-inverse matrix equation is applied which shortens the calculation time to be less
than one hour for a matrix size of 5476×35910.
RESULTS
In this numerical
simulation, the sampling number is 513 and the matrix size is 5476×35910. The rotation of magnet array is not needed and
supercoils are used. TSVD, Kaczmarz iteration (KI), and LMS were applied for
the reconstruction. Fig. 2 shows the reconstructed images by different methods
when SNR is 60dB and the L-curve. The normalized root mean square error (NRMSE)
of the reconstructed images using these methods at different SNR’s are tabulated
and compared in Table. 1.DISCUSSION
Fig. 2(a)-(c) shows the results using
supercoils. As shown, at the same SNR, a center blurry area is observed in the
image reconstructed using KI whereas TSVD shows the best image quality. The
results show that LMS method is not suitable for the reconstruction here. Because of large condition number of encoding matrix, random error could be magnified greatly in the solution. Fig. 2 (d) shows
the L-curve. The first turning point of the L-curve represents the singular
value that the error of the solution begins to decrease. The second turning
point should be selected to get a stable solution because at this point, all the
large singular values of the encoding matrix can be included in the image
reconstruction. In Table. 1, for SNR up to 40dB, the NRMSE of TSVD is lower
than that of KI method. It is observed that, the NRMSE of iteration method decrease
more slowly than TSVD as SNR increases. For the supercoil setup, by using TSVD,
the NRMSE is relatively low at the same SNR and meanwhile, the blurred area can
be mitigated. CONCLUSION
It is shown that in a
portable MRI scanner, the image can be reconstructed by electrically forming
receive RF coils aggregates with phase delay and the rotation of Halbach magnet
array is no longer needed. In the resultant system, TSVD with an optimized
regularization parameter (enabled by an accelerated L-curve method) reconstructs images with an improved quality with lowered NRMSE compared to that of Kaczmarz
iteration. Moreover, by TSVD, the central blurred area of reconstructed image is
successfully mitigated. Acknowledgements
No acknowledgement found.References
1. Cooley, C. Z., et al. Two-dimensional imaging in a
lightweight portable MRI scanner without gradient coils. Magnetic Resonance
Medicine. 2015;73(2): 872-883. 2. Cooley, C. Z. Portable low-cost magnetic resonance
imaging. Ph.D. dissertation, USA, Massachusettts Institute of Technology, Sep.
2014.
3. D. W. Lu, S. Y. Huang. A
TSVD-based Approach for Flexible Spatial Encoding Strategy in Portable Magnetic
Resonance Imaging (MRI) System. In 39th PIERS, Nov. 2017.
4. P. C. Hansen. The L-curve and its use in the numerical treatment of inverse
problems. Computational Inverse Problems in Electrocardiology. 1999: 119-142.