CUDA-based acceleration and algorithm refinement for volume image registration

Abstract

In this paper, we propose a GPU-based acceleration method to speed up volume image registration using Compute Unified Device Architecture(CUDA). A novel CUDA-based method for joint histogram computation is introduced in this paper, which is also valuable for 2D image registration and other general graphics applications. Additionally, an algorithm refinement is proposed to improve the widely used FMRIB's Linear Image Registration Tool (FLIRT). Although extra time is taken by applying that algorithm refinement, our implementation showed the ability to perform a full 12 DOF (Degrees of Freedom) registration of two brain volume images in nearly 35 seconds, which is about 10 times faster than the native FLIRT implementation.