high performance computing on graphics processing units: hgpu.org

A miniature 3D tracked ultrasonic probe has been developed to acquire intra-articular cartilage images under arthroscopic surgical conditions. The aim is to detect cartilaginous lesions (arthritis) and quantify their precise sizes and locations to help the clinician in his diagnostic and his therapeutic decision making. The ultrasonic transducer is tracked by an optical sensor, which permits to find location and orientation of each 2D US images in a common 3D spacial reference. Near two thousands images are acquired when scanning a cartilage surface. An interesting tool is to rebuild a 3D isotropic volume (cubic voxels) with those images, allowing further processing. Conventional 3D ultrasound algorithms have low computational complexity but the huge amount of data generated makes it difficult to compute results within reasonable time on classical computers. In this paper we investigate the possibilities of regenerating a 3D isotropic volume with the help of GPGPU (CUDA) by adapting existing algorithms to massive parallelism provided by modern everyday GPUs.