GPU accelerated implementation of kernel regression for freehand 3D ultrasound volume reconstruction

Freehand three-dimensional (3D) ultrasound has recently gained an increasing attention in the medical field as it enables low-cost, safe, and flexible 3D scanning of arbitrary-shaped anatomical structures. During freehand 3D ultrasound imaging, the clinician freely moves a conventional two-dimensional (2D) ultrasound probe to acquire different views of the organ. The acquired 2D ultrasound data, which is often irregularly and sparsely distributed in the 3D space, is processed using a reconstruction algorithm to synthesize a 3D ultrasound volume. One effective reconstruction method is the 3D Kernel Regression (KR) algorithm that enables the generation of high-quality ultrasound volumes. This algorithm performs nonparametric 3D interpolation of the voxel gray-level values. One limitation of the 3D KR algorithm is the high computational complexity that requires long execution times when run on serial computers. To overcome this limitation, this paper presents a parallel implementation of the 3D KR algorithm using graphics processing unit (GPU) technology. The parallel 3D KR algorithm enables the generation of high-quality ultrasound volumes in short execution times. The feasibility of the parallel 3D KR algorithm has been demonstrated by synthesizing an ultrasound volume of a breast tumor. The parallel execution time required to synthesize the ultrasound volume is more than ten times faster than the serial execution time.