Image registration is an important preliminary task in many image processing problems like reconstruction of 3D information from 2D views or motion estimation from temporal sequences. The purpose of image registration (also called image matching) is to geometrically align one image (the floating or source image) to another one (the reference or target image) so that voxels (or pixels in  2D) representing the same underlying structure may be superimposed. A standard framework for image registration consists in minimizing a cost function or maximizing a "similarity measure" that expresses the pixel or voxel similarity of the images to be aligned. Due to its large variety of sensors, 3D medical imaging is certainly one of the first application field. Applications range from computer-assisted surgery to the analysis of sequences of functional images used for example to follow the evolution of diseases. In fact, it greatly improves both diagnosis and therapy. The registration may concern images from the same modality or images from different modalities (multimodal image matching). To compare these images, many features (edges, surfaces, voxels, etc.) and similarity measures are now available. Besides, transformations range from rigid transformation, with a small number of parameters, to deformable image warps, depending on several thousands or millions of parameters (figure below - T is the transformation). We consider both rigid and deformable image matching problems.

Source image                                                  Target image

Standard cost functions, based on voxel similarity measures, are highly non-linear, non-convex, exhibit many local local minima and thus yield hard optimization problems. On the one hand, for rigid matching of single modal images (Magnetic Resonance Images) we have considered the parallelization of a general purpose global optimization 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 data-parallel algorithm can yield subvoxel accuracy and exhibits an almost linear speedup. On the other hand, for deformable inter-subject matching of 3D MR images (the registration of two 256^3 MRI typically involves the optimization of 200,000 parameters, and takes several hours on a standard workstation), we have proposed a general fully parallel approach based on the simulation of stochastic differential equations. This approach is naturally suited to massively parallel implementations, yielding high quality registrations and excellent relative speedups. To conclude, this work has shown the relevance of the parallelization for medical image registration and the suitability of the data-parallel programming model.