Generation of planar radiographs from 3D anatomical models using the GPU

This paper describes the development of parallel, high performance algorithms for generation of DRRs out of 3D vertebrae models, using the GPU. NVIDIA CUDA platform was chosen for the implementation of these algorithms because it provides flexible and fine-grain control over the hardware when compared with other platforms. This research is motivated by recent work aiming to attain a process to recover the shape of the human spine of scoliosis patients using two planar radiographs. A DRR generation algorithm would allow for 2D/3D registration during the process, for best reconstruction accuracy. However, this requires generation of hundreds of DRRs per second, which is not compatible with a CPU implementation. The top performance registered for the DRR generation of a lumbar vertebra was 828 FPS with an entry-level graphic card, showing that the GPU implementation achieves performances compatible with 2D/3D registration problems while keeping hardware requirements low.