GPU parallel implementation of B-spline non-rigid grid registration using free-form deformations

Image registration is a vital problem in medical imaging. Fast deformable registration methods are needed especially in intra-operative settings characterised by their time-critical applications. In this paper, we present our work on B-spline non-rigid registration using free-form deformations. Our first implementation which was serial on standard desktop machine showed the algorithm is computed intensive and not real time; even multithread implementation in CPU was not real time because of high execution time. We have broken the registration process in threads and blocks with the help of Compute Unified Device Architecture (CUDA) for handling GPU. Our work employs extensive usage of highly multithreaded architecture of multicored GPU. An efficient use of shared memory is required to optimise parallel reduction in CUDA. Experimental results show that parallel implementation achieved an average speed-up of 82x over the serial implementation when run on a GPU GeForce 9500 GT having 32 cores and around 56x over multicore architecture CPU. Registration method is also being evaluated using SSD (Squared Sum of Intensity Differences) and CC (Cross Correlation).