Parallel Automatic Registration of Large Scale Microscopic Images on Multiprocessor CPUs and GPUs

During the present decade, emerging architectures like multicore CPUs and graphics processing units (GPUs) have steadily gained popularity for their ability to deploy high computational power at a low cost. In this paper, we combine parallelization techniques on a cooperative cluster of multicore CPUs and multisocket GPUs to apply their joint computational power to an automatic image registration algorithm intended for the analysis of high-resolution microscope images. Registration methods pose a computational challenge within the biomedical field due to the large size of microscope image data sets, which typically extend to the Terabyte scale. We analyze this application to identify those parts which are more favorable to the CPU and GPU execution models and decompose the process accordingly. Performance results are presented for two sets of images: mouse placenta (16K x 16K pixels) and mouse mammary tumor (23K x 62K pixels). Execution times are shown on different multi-node, multi-socket and multi-core configurations to provide performance insights about the most effective approach.