A taxonomy for medical image registration acceleration techniques

For the past decade, improving the performance and accuracy of medical image registration has been a driving force of innovation in medical imaging. Fast, accurate image registration enhances diagnoses of patients, accounts for changes in morphology of structures over time, and even combines images from complementary modalities (e.g. computed tomography and positron emission tomography). The ultimate goal of medical image registration research is to create a robust, real time, elastic registration solution that may be used for multiple modalities such that new image-guided intervention techniques can be pioneered. With such a computationally intensive and multifaceted problem, researchers have applied a variety of innovative approaches at different levels of the problem to improve the performance of this solution. This diversity in work presents many opportunities for synergies between complementary techniques, but to date has mostly gone unrealized. This paper samples existing literature on image registration acceleration techniques and proposes a classification of "levels" into which they can be binned. We show experimentally that combining approaches from different levels can lead to over 100x speedup on an eight node heterogeneous cluster.