3-D Tomographic Image Reconstruction from Randomly Ordered Lines with CUDA

We present a novel method of computing line-projection operations along sets of randomly oriented lines with CUDA and its application to positron emission tomography (PET) image reconstruction. The new approach addresses challenges that include compute thread divergence and random memory access by exploiting GPU capabilities such as shared memory and atomic operations. The benefits of the CUDA implementation are compared with a reference CPU-based code. When applied to PET image reconstruction, the CUDA implementation is 43X faster, and images are virtually identical. In particular, the deviation between the CUDA and the CPU implementation is less than 0.08% (RMS) after five iterations of the reconstruction algorithm, which is of negligible consequence in typical clinical applications.