Jose E.C. Serralles1,2, Ilias I. Giannakopoulos1,2, and Riccardo Lattanzi1,2,3
1Bernard and Irene Schwartz Center for Biomedical Imaging, Department of Radiology, New York University Grossman School of Medicine, New York, NY, United States, 2Center for Advanced Imaging Innovation and Research (CAI2R), Department of Radiology, New York University Grossman School of Medicine, New York, NY, United States, 3Vilcek Institute of Graduate Biomedical Sciences, New York University Grossman School of Medicine, New York, NY, United States
Synopsis
Keywords: Low-Field MRI, Simulations
Motivation: Accurate and precise simulation of low field MR, as well as simulation of high field MR with arbitrary granularity.
Goal(s): To assess and address the limitations in the Magnetic Resonance Integral Equation (MARIE) suite that prevent successful low field simulation.
Approach: We achieved these goals using a number of numerical analysis techniques, such as Taylor series approximations and Kahan summation.
Results: We successfully identified and addressed the limitations of MARIE, which we attributed to catastrophic loss of numerical precision. In doing so, we enabled low frequency and fine mesh simulation.
Impact: Low field and detailed simulation of coils would enable many applications, such as coil optimization, synthetic generation of MR data for machine learning algorithms, computational pulse sequence optimization, accurate safety assessments in simulation, among many other applications.
Introduction
The
standard of assessment for precision and accuracy of simulation tools is
typically performed on elements whose characteristic
lengths are on the order of $$$\frac{\lambda}{8}$$$, where $$$\lambda$$$ denotes vacuum wavelength. When designing an accurate RF coil simulation tool
suitable for ultra-low to ultra-high field MRI simulations, this practice is
problematic: for example, at a field strength of 3T, $$$\lambda$$$ is
approximately 2.3 meters, meaning that accuracy is guaranteed only when the characteristic length of the discretization is on
the order of 30cm. Because wavelength is inversely proportional to the frequency, the performance of these
simulation tools further degrades at low field and fails completely at ultra-low fields.
In this work, we detail the processes by which we assessed the precision of the
open-source Magnetic Resonance Integral Equation suite MARIE1, and how we
drastically improved its performance in cases where precision was lost catastrophically.Theory & Background
We focus our attention on the Surface Integral Equation (SIE) component
of the MARIE suite2. We quickly summarize the SIE formulation: The underlying
system of equations is the standard Electric Field Integral Equation formulation
(EFIE) discretized using the Galerkin method and Rao-Wilton-Glisson (RWG) basis
functions3. The
condition number of a such a system for a typical coil array is on the order of
105 because the eigenvalues cluster about $$$0$$$. Consequently, the conditioning of this system limits the maximum
accuracy (11 digits for double precision), and amplifies any errors that may be
present in the system of equations itself. This system of equations is given by
$$\mathbf{Z}_{cc}\mathbf{J}_c=\mathbf{F}\mathbf{V}_{\rm{in}}\text{,}$$ where
$$$\mathbf{Z}_{cc}$$$ is the discretized system of equations, $$$\mathbf{J}_c$$$ is the
set of RWG coil currents, $$$\mathbf{F}$$$ is an interpolation matrix, and
$$$\mathbf{V}_{\rm in}$$$ is the input voltage.
Assembling the system involves calculating a large number of singular and non-singular
integrals. These integrals are of the form
$$\iint_{S_i}\iint_{S_j}f_m\!\left(\mathbf{r}\right)\cdot f_n\!\left(\mathbf{r}^\prime\right)G\!\left(\mathbf{r}-\mathbf{r}^\prime\right){\rm d}S^\prime{\rm d}S$$
where $$$f_m$$$ and $$$f_n$$$ denote the testing and basis functions of the
discretization, respectively. The function $$$G\!\left(\mathbf{r}-\mathbf{r}^\prime\right)$$$
is the Dyadic Green’s function. In MARIE, the singular integrals are computed
using the Direct Evaluation Method and can be subdivided
into three cases2: Self-Term (ST), Edge Adjacent (EA), and Vertex Adjacent (VA).
The non-singular (NS) integrals are computed with standard Gauss-Legendre quadrature
schemes.Methods & Results
We assessed the
precision of the assembly of the SIE method by manually varying the order of
the 1D Gauss-Legendre quadrature, and then calculating the Frobenius norm of
the difference of $$$\mathbf{Z}_{cc}$$$, normalized by the norm of the
highest tested order. We made the following observations:
- ST, EA, and VA
terms had a precision of 10-6 for the fixed orders in the original MARIE.
- NS terms had a precision of 10-6 for the fixed order.
- The
maximum attainable precision of singular terms was 10-10 due to the catastrophic
loss of numerical precision.
- At 1.5T and for a loop with 8cm diameter (scaled
down from Fig. 1), the error in the coil currents is on the order of 5%.
The singular
integrals had a maximum precision of 10
-10 due to catastrophic loss
of numerical precision, which can be attributed to the leading terms of the
corresponding Taylor series expansions canceling out (see Fig. 2 for an example).
Manually calculating these expansions using Horner’s Method
4 decreased the error to 10
-15. We separated
NS interactions into several cases based on distance and ascribed higher
quadrature orders to each case to obtain machine precision without compromising
speed. Additionally, we also employed Kahan summation
5 to further boost precision. The end result is an implementation of SIE
with precision (and hence accuracy) in the coil currents on the order of 10
-10.
Additionally, in boosting precision, we addressed low-frequency breakdown
in SIE. Fig. 3 demonstrates the imaginary part of the impedance of the loop versus frequency. The old approach fails catastrophically at 3MHz, whereas the proposed novel approach works for frequencies
as low as 30kHz.
Conclusion & Future Work
We successfully
identified and addressed catastrophic loss of numerical precision issues and
high quadrature errors in MARIE using Taylor series expansions and higher quadrature orders. We believe that this increase in
accuracy is essential to better replicate experiments in simulation, especially
at low field strengths. Such a high degree of precision and hence accuracy
would enable applications like synthetic data generation for machine
learning algorithms, patient safety assessments, and iterative methods used to
approximate ultimate performance bounds6. The achieved high precision will be key for
iterative inverse problems, such as coil optimization7 or electrical property
mapping8 because a systematic error such as the one detected for MARIE could significantly bias the overall simulation’s
accuracy.Acknowledgements
This work was supported in part by NIH R01 EB024536 and was performed under the rubric of the Center for Advanced Imaging Innovation and Research (CAI2R, www.cai2r.net), an NIBIB National Center for Biomedical Imaging and Bioengineering (NIH P41 EB017183).References
- Villena JF, Polimeridis AG, Eryaman Y,
Adalsteinsson E, Wald LL, White JK, Daniel L. Fast electromagnetic
analysis of MRI transmit RF coils based on accelerated integral equation
methods. IEEE Transactions on Biomedical Engineering. 2016 Jan
22;63(11):2250-61.
- Polimeridis AG, Tamayo JM, Rius JM,
Mosig JR. Fast and accurate computation of hypersingular integrals in
Galerkin surface integral equation formulations via the direct
evaluation method. IEEE transactions on antennas and propagation. 2011
Apr 19;59(6):2329-40.
- Rao S, Wilton D, Glisson A. Electromagnetic scattering by surfaces of
arbitrary shape. IEEE Transactions on antennas and propagation. 1982
May;30(3):409-18.
- Burrus CS, Fox JW, Sitton GA, Treitel
S. Horner’s method for evaluating and deflating polynomials. DSP
Software Notes, Rice University, Nov. 2003 Nov 26;26.
- Kahan, W. Further remarks on reducing truncation errors. Communications of the ACM, 8 (1): 40, doi:10.1145/363707.363723, Jan. 1965.
- Georgakis IP, Polimeridis AG, Lattanzi
R. A formalism to investigate the optimal transmit efficiency in
radiofrequency shimming. NMR in Biomedicine. 2020 Nov;33(11):e4383.
- Serralles JEC, Adalsteinsson E, Wald LL, Daniel L. Parametric coil
optimization via global optimization. InProc. ISMRM 2021 (p. 1399).
- Serrallés JEC, Giannakopoulos II, Zhang
B, Ianniello C, Cloos MA, Polimeridis AG, White JK, Sodickson DK,
Daniel L, Lattanzi R. Noninvasive estimation of electrical properties
from magnetic resonance measurements via global Maxwell tomography and
match regularization. IEEE Transactions on Biomedical Engineering. 2019
Mar 25;67(1):3-15.