4.4.7 Iterative Methods

Improved Equipment and Techniques for Dynamic Shimming in High Field MRI

Dr. Michael Poole 2007

4.4.7  Iterative Methods

Optimisation of the homogeneity, efficiency, inductance etc. may also be achieved by iterative techniques. Most optimisation methods work by allowing the positions of discrete wires as free parameters in the optimisation routine. Wong et al. [114] used an iterative technique to design a Z-gradient coil by allowing the positions of the loops of wires to be iteratively adjusted so as to minimise an error function characterising the field error and/or the coil inductance. The gradient of the error function with respect to the free parameters is also calculated for each iteration and the parameter space is searched by conjugate gradient descent (CGD), a first-order optimisation algorithm. There are several other examples of the use of CGD in gradient coil design [115,116, 117, 118].

The most commonly used iterative optimisation algorithm in gradient coil design has been simulated annealing (SA). This is because it is excellent at avoiding local minima in the error function when finding the global minimum. It was first used to adjust the positions of circular loops of wire so as to generate a coil producing a highly uniform Z-gradient [119]. SA was later used to design X-gradient coils by parameterising the positions of the wire elements [120]. Crozieret al. described one quadrant of the X-gradient coil as a series of concentric circles deformed by transformation mapping with much fewer free parameters [121]. The SA algorithm has been used extensively to optimise parameters in the design of various gradient and shim coil[122, 123, 124, 125, 126, 127, 128,129, 130, 131, 132, 133, 134, 98, 135, 136, 137, 138].

Another common stochastic optimisation technique is to mimic Mendelian evolution, and allow better solutions to the problem to "survive", "mutate" and "breed" to form the solutions to the next iteration. This genetic algorithm (GA) has been applied to the design of gradient coils in a similar way to CGD and SA so as to produce more linear and compact cylindrical Z-gradient coils [139], X-gradient coils [140,141], and bi-planar coils [142, 143].

Adamiak et al. [144] used a zeroth-order Powell optimisation (one that only uses the result of the Biot-Savart calculation, and not its derivative) to adjust the positions of wire segments iteratively so as to generate a four-turn X-gradient coil. The Monte Carlo (MC) method has also been applied to gradient and shim coil design [145,146]. Linear [147] and quadratic [148] programming are other options for optimising gradient coil design.

Iterative coil design techniques often have the advantage that they work directly with the wires of the coil design. The magnetic field, and therefore the error function to optimise, is calculated by direct integration of the Biot-Savart Law (Eq. (4.1)) over the wires. When approximating the continuous current density (§ 4.4.2 to § 4.4.6) in terms of a small number of turns of wire, the approximation becomes less valid. Hence, iterative methods, using discrete wires work better for few wires than continuous current density methods. However, iterative methods are considerably more computationally intensive, and tend to produce wire patterns with abrupt changes in directions. These changes can introduce high order spherical harmonics in the magnetic field. These methods also do not need there to be more free coil design coil parameters than field parameters.