A GPU accelerated moving mesh correspondence algorithm with applications to RV segmentation

This study proposes a parallel nonrigid registration algorithm to obtain point correspondence between a sequence of images. Several recent studies have shown that computation of point correspondence is an excellent way to delineate organs from a sequence of images, for example, delineation of cardiac right ventricle (RV) from a series of magnetic resonance (MR) images. However, nonrigid registration algorithms involve optimization of similarity functions, and are therefore, computationally expensive. We propose Graphics Processing Unit (GPU) computing to accelerate the algorithm. The proposed approach consists of two parallelization components: 1) parallel Compute Unified Device Architecture (CUDA) version of the non-rigid registration algorithm; and 2) application of an image concatenation approach to further parallelize the algorithm. The proposed approach was evaluated over a data set of 16 subjects and took an average of 4.36 seconds to segment a sequence of 19 MR images, a significant performance improvement over serial image registration approach.