Optimized GPU histograms for multi-modal registration

GPU-based systems are used more and more for medical image processing because of their parallel processing power and memory bandwidth. Impressive results have been achieved when registering large volume, however, one of the most-used similarity measures for multi-modal registration - mutual information - is not well suited for the streaming architecture because of its memory access pattern. We present two optimization approaches that improve the performance by a factor of four compared to state-of-the-art GPU algorithms in the latest research papers.