A parallel approach for alignment of multi-modal gridbased data

Multi-modal registration is still a big challenge in image processing. In this article we present a variation of the wellknown fast Fourier transform (fft) accelerated methods for finding the alignment between two datasets, i.e. the rigid transformation consisting of a rotation as well as a translation mapping all regions in both datasets belonging together in an optimal manner. Our method can be applied to such multi-modal registration problems as computer tomography (CT)/positron emission tomography (PET) matching, or matching CT data with data obtained by magnetic resonance imaging (MRI). We reformulate the alignment problem into an optimization problem concerning a metric measure. The particular form of the proposed objective function can be exploited to fft-accelerate the translational part of the alignment-problem. Thus the reduced problem can be solved for the three remaining degrees of freedom of the rotatory part using standard optimizers, such as downhill-simplex or Powell’s method. A further advantage of our approach is the straight forward parallelization of the objective function’s computation. Our implementation on a graphic processing unit (GPU) yielded a speedup factor between 5 and 25 depending on the size of the data. The results show, that the application of a GPU can be highly rewarding for all fft accelerated algorithms.