Evaluation of an efficient GPU implementation of digitally reconstructed radiographs in 3D/2D image registration

Abstract

Intensity-based three-dimensional to two-dimensional (3D/2D) X-ray image registration algorithms usually require generating digitally reconstructed radiographs (DRRs) in every iteration during their optimization phase. Thus a large part of the computation time of such registration algorithms is spent in computing these DRRs. In a 3D-to-multiple-2D image registration framework where a sequence of DRRs is calculated, not only the computation but also the memory cost is high. We present an efficient DRR generation method to reduce both costs on a graphical processing units (GPU) implementation. The method relies on integrating a precomputation stage and a narrow-band region-of-interest calculation into the DRR generation. We have demonstrated its benefits on a previously proposed non-rigid 4D-to-multiple-2D image registration framework to estimate cerebral aneurysm wall motion. The two tested algorithms initially required several hours of highly intensive computation that involves generating a large number of DRRs in every iteration. In this paper, results on datasets of digital and physical pulsating cerebral aneurysm phantoms showed a speedup factor of around 50x in the generation of DRRs. In further image registration based wall motion estimation experiments using our implementation, we could obtain estimation results through the whole cardiac cycle within 5 minutes without degrading the overall performance.