Synopsis
This work presents a study on the performance of several least-squares
optimization algorithms used for localized in-vivo B0 shimming.
Seven different algorithms were tested in 4 different shim volumes in the brain:
global shimming region, single slice, and single voxels in two different
positions with 3rd order shimming at 7T. Each algorithm's robustness
and convergence were tested against noisy inputs and different starting values.
The results give an interesting overview of the properties of each algorithm
and their applicability. The regularized iterative inversion algorithm proves
to be the best algorithmic approach suited to this problem.Introduction
Magnetic resonance spectroscopy and imaging applications benefit
from an automated B
0 shimming algorithm which is numerically stable,
efficient and applicable to any arbitrary volume of interest. Recently,
different approaches for solving this problem by 3D B
0 mapping
combined with least-squares optimization have been proposed [1-6]; yet there is
still a need for a systematic comparison of these algorithms that would provide
insights for choosing the best approach for B
0 shimming applications. The aim of this work is to evaluate and
compare the performance of 7 different optimization algorithms embedded into
the same general software framework based on their convergence rate, numerical
stability and solution optimality in a range of B
0 shimming
applications in human brain MRI at 7T.
Methods
The B0 shimming problem can be characterized by a system
of linear equations, where each element of b is the value of the B0
field, A represents the sensitivity of each shim coil and x
contains the applied currents. Thus each element of the matrix Ax gives the
field generated by a given shim channel at a given position. The system of
linear equations can then be cast as a linear least-squares optimization
problem:
$$\min_x ||(Ax-b)||^2$$
The maximum shim fields can be included as box-constraints $$lb\leq x\leq ub$$
This problem is quadratic and hence convex. Therefore any solution
is a global minimum. The convexity is still preserved even if A is
ill-conditioned (which is often the case for single-slice and off-center
single-voxel shimming volumes due to the lack of orthogonality of the shim
fields at off-center positions), however, solving the problem becomes tricky
because a small change in the data (e.g. noise in the acquired reference B0 map)
can lead to radically different answers and optimal convergence is not
guaranteed.
Seven different algorithms were implemented in MATLAB™ R2013b:
- nonlinear optimization using three
methods: active-set (nonlin-as) [7], interior-point (nonlin-ip) [8] and
sequential quadratic programming (sqp) [7] using MATLAB™’s fmincon;
- quadratic programming using the
interior-point-convex method (qp-ip) [9] using MATLAB™’s quadprog ;
- iteratively inverting the system and
truncating the values that exceed the limits (invcon);
- the same method except that the
singular values of A are truncated to improve the conditioning (invcon-tsvd);
and
- finally, for comparison purposes, unconstrained
pseudo-inversion algorithm (pinv) using MATLAB™’s pinv, which always
yields the best shim quality achievable (if the solver was not bound by
constraints).
For this study, 33 in-vivo datasets from the brain of
healthy volunteers obtained at a 7T Philips scanner were used. For each
volunteer, the performance of all the algorithms for 3rd order
shimming in 4 different shim volumes were studied (figure 1): a global brain region,
a single slice positioned off-center, a single spectroscopy voxel located in
the frontal cortex, and one in the visual cortex. The shim volumes were defined
in a custom GUI application.
The methods were evaluated according to three criteria: convergence
rate, starting-point sensitivity (i.e. how sensitive the result is based on the
starting-point of the optimization algorithm), and the robustness of the
solution against small perturbations to the input. For the starting-point
sensitivity test, 1000 randomly initialized starting points were used. For the
robustness test, 1000 different noisy B0 vectors generated by adding random white
noise (with a maximum amplitude of 5% of the original signal) were considered.
The amplitude constraints of the shim system are shown in figure 2.
Results/Discussion
Figure 3 shows the algorithms’ sensitivity to different starting
values. Note that since the dedicated quadratic solvers do not require a
starting value input, this test was not applicable for them. Among all the
other algorithms, the interior-point method is the least sensitive.
Figure 4 shows the robustness of the algorithms against a small noise
perturbation in the input data and the invcon-tsvd algorithm proves to be the most
robust.
Figure 5 shows the overall performance of these algorithms in all 4
shimming areas. As can be seen, the invcon-tsvd method consistently
provides the best shim quality even in hard-to-shim areas where the problem
becomes most ill-conditioned.
Finally in figure 6, a comparison of the convergence rate and run-time of these algorithms is shown. The invcon-tsvd algorithm was found to be
the fastest.
Conclusion
For localized
in-vivo B
0 shimming the use of a dedicated
quadratic programming solver instead of a generic nonlinear least-squares solver
is highly recommended. The reason being that the quadratic solvers do not
depend on a starting value and also converge much faster. Among the quadratic
solvers, the regularized iterative inversion method (invcon-tsvd) was found to
be both fastest and the most robust.
Acknowledgements
No acknowledgement found.References
[1] Gruetter R, et al, “Fast, non-iterative shimming of spatially
localized signals”, JMRI 96, 323-334, 1992.
[2] Kim D, et al, “Regularized Higher order in-vivo shimming”, MRM
48:715-722 2002.
[3] Hetherington H. et al, “Robust fully automated shimming of the human brain for
high-field 1H spectroscopic imaging”, MRM 56:26-33, 2006.
[4] M.
Weiger et al, “Gradient shimming with spectrum optimization”, Journal of Magnetic Resonance
Volume 182, Issue 1, September 2006.
[5] Juchem et al, “Dynamic shimming of the human brain at 7 Tesla”,
Concepts of Magnetic Resonance Part B MR Eng.: 116-128, 2010.
[6] Fillmer A, et al, “Constrained Image-BasedB0Shimming Accounting for
“Local Minimum Traps” in the Optimization and Field inhomogeneity outside the
Region of Interest”, MRM 2014 Epub.
[7] Fletcher, R., "Practical Methods of
Optimization," John Wiley and Sons, 1987.
[8] Byrd, R.H., J. C. Gilbert, and J. Nocedal, "A Trust
Region Method Based on Interior Point Techniques for Nonlinear
Programming," Mathematical Programming, Vol 89, No. 1, pp. 149–185,
2000.
[9] Gill, P.E., W. Murray, and M.H. Wright,
"Numerical Linear Algebra and Optimization", Vol. 1, Addison Wesley, 1991.