High performance derivative-free optimization applied to biomedical image registration

Optimization of a similarity metric is an essential component in most medical image registration approaches based on image intensities. In this paper, two new, deterministic, derivative-free optimization algorithms are parallelized and adapted for image registration. DIRECT (dividing rectangles) is a global technique for linearly bounded problems, and the multidirectional search (MOS) is a local method. Unlike many other deterministic optimization techniques, DIRECT and MDS allow coarse-grained parallelism. The performance of DIRECT, MDS, and hybrid methods using a fine-grained parallelization of Powell´s method for local refinement, are compared. Experimental results show that DIRECT and MDS are robust, and can greatly reduce computation time in parallel implementations.