Rapid rabbit: Highly optimized GPU accelerated cone-beam CT reconstruction

Graphical processing units (GPUs) have become widely adopted in the medical imaging community. The parallel SIMD nature of GPUs maps perfectly to many reconstruction algorithms. Because of this, it is relatively straightforward to parallelize common reconstruction algorithms (e.g. FDK backprojection). This means that significant performance improvements must come from careful memory optimizations, exploiting ASICs and a few other tricks to boost instruction throughput. We present optimizations that build off of previous work to optimize a GPU accelerated FDK backprojection implementation using the RabbitCT dataset.