Parallel Differential Evolution: Application to 3-D Medical Image Registration

Abstract

A common framework for 3-D image registration consists in minimizing a cost (or energy) function that expresses the pixel or voxel similarity of the images to be aligned. Standard cost functions, based on voxel similarity measures, are highly nonlinear, non-convex, exhibit many local minima and thus yield hard optimization problems. Local, deterministic optimization algorithms are known to be sensitive to local minima. Global optimization methods (like simulated annealing or evolutionary algorithms) yield better solutions often close to the optimal ones, but are time consuming. In this section we consider the parallelization of a general-purpose global optimization algorithm based on random sampling and evolutionary principles: the differential evolution algorithm. The inherent parallelism of evolutionary algorithms is used to devise a data-parallel implementation of differential evolution. The performances of the parallel version are assessed on a 3-D medical image registration problem. Besides yielding accurate registrations, parallel differential evolution exhibits fast convergence and a speedup almost growing linearly with respect to the number of processors.