Fast 2D-3D registration using GPU-based preprocessing

This paper describes the fast point-based 2-D/3-D registration that will increase the registration speed of intraoperative two-dimensional (2-D) fluoroscopy and preoperative three-dimensional (3-D) CT images using GPU (graphics processing unit)-based DRR´s preprocessing. Rigid 2-D/3-D registration of 2-D fluoroscopy images with 3-D CT images can be used for image-guided surgery or intraoperative navigation. X-ray fluoroscopy images provide real-time visualization. However, in general, resolution of fluoroscopy images is limited, these modalities are only 2-D and features in the front of body overlap one. Because of its drawback, three-dimensional imaging modalities such as computed tomography (CT) and magnetic resonance (MR) imaging are broadly used in clinical diagnostics and treatment planning [1], These have the spatial information and high resolution, but at present their use as interventional imaging modalities has been limited. In this paper, to utilize CT information during interventional procedures, a preoperative CT scan is aligned with an intraoperative X-ray fluoroscopy image. In preprocessing procedure we generate CT-derived DRRs using graphic hardware. This method is over 150 times faster than software rendering. And for registration accuracy and speed, we propose point-based 2D-3Dregistration of phantom dataset. And to reduce computation cost, we apply point-based registration technique. Because this method leads the computation time to about one second, the registration speed is enough to apply to intervention.