GPU-Accelerated D2VR

Abstract

Traditional volume rendering approaches rely on obtaining values of sampled points in volumetric space, typically on a cartesian grid. Often, this cartesian grid is not the original source of the data. For example, in tomographic imaging applications, such as used in diagnostic medical or industrial CT, the primary source of the data is the set of X-ray projections taken around the object. To enable visualization with established volume rendering methods, the volume must first be reconstructed from these projections. Since sampling is involved, this process introduces errors, adversely impacting image quality. Recently a new rendering technique was proposed, named D2VR, which skips the intermediate reconstruction step entirely and samples the projections directly. It was shown that doing so can improve image quality significantly. But despite its great promise, a shortcoming of the method was its comparatively slow rendering speed. Interactive or at least near-interactive speed, however, is critical for clinical deployment of a visualization framework. To address this shortcoming, our paper proposes a GPU-accelerated D2VR, with facilities for occlusion culling and empty space skipping to achieve further speedups.