Towards systematic exploration of tradeoffs for medical image registration on heterogeneous platforms

For the past decade, improving the performance and accuracy of medical image registration has been a driving force of innovation in medical imaging. The ultimate goal of accurate, robust, real-time image registration will enhance diagnoses of patients and enable new image-guided intervention techniques. With such a computationally intensive and multifaceted problem, improvements have been found in high performance platforms such as graphics processors (GPUs) and general purpose clusters, but there has yet to be a solution fast enough and effective enough to gain widespread clinical use. In this study, we examine the differences in accuracy and speed of implementations of the same image registration algorithm on a general purpose uniprocessor, a GPU, and a cluster of GPUs. We utilize a novel domain specific framework that allows us to simultaneously exploit parallelism on a heterogeneous platform. Using a set of representative images, we examine implementations with speedups of up to two orders of magnitude and accuracy varying from sub-millimeter to 2.6 millimeters of average error.