Fast and accurate volume data curvature determination using GPGPU computation

Abstract

A methodology for fast determination of a key shape feature in volume datasets using a GPU is described. The shape feature, surface curvature, which is a valuable descriptor for structure classification and dataset registration applications, can be time-consuming to determine reliably by conventional serial computing. The techniques here use parallel processing on a commodity GPU to achieve 100-fold (and above) improvements (for moderate-sized datasets) over conventional serial processing for curvature determination. Techniques for one class of curvature determination methods are detailed, including methods well-suited to datasets acquired by medical scanners.