Towards a unified framework for rapid 3D computed tomography on commodity GPUs

The task of reconstructing an object from its projections via tomographic methods is a time-consuming process due to the vast complexity of the data. For this reason, manufacturers of equipment for computed tomography (CT), both medical and industrial, rely mostly on special ASICs to obtain the fast reconstruction times required in clinical, industrial, and security settings. Although modern CPUs have gained enough power in recent years to be competitive for 2D reconstruction, this is not the case for 3D reconstructions, especially not when iterative algorithms must be applied. Incidentally, this has prevented some very effective algorithms to be used in clinical practice, and the need for proprietary reconstruction hardware has also hampered new equipment manufacturers in their effort on entering the market. However, the recent evolution of GPUs has changed the picture in a very dramatic way. We will show how floating point GPUs can be exploited to perform both analytical and iterative reconstruction from X-ray and functional imaging data at clinical rates and good quality. For this purpose, we derive a decomposition of three popular 3D reconstruction algorithms into a common set of base modules. All of these base modules can be executed on the GPU and their output linked internally. The data never leaves the GPU, which eliminates the previous GPU-CPU bottlenecks. Visualization of the reconstructed object is also easily done since the object already resides in the graphics hardware, and one can simply run a visualization module at any time to view the reconstruction results. Our implementation allows speedups at a factor of 20, compared to software implementations, at comparable image quality.