X-Ray CT Image Reconstruction on Highly-Parallel Architectures

Model-based image reconstruction (MBIR) methods for X-ray CT use accurate models of the CT acquisition process, the statistics of the noisy measurements, and noise-reducing regularization to produce potentially higher quality images than conventional methods even at reduced X-ray doses. They do this by minimizing a statistically motivated high-dimensional cost function; the high computational cost of numerically minimizing this function has prevented MBIR methods from reaching ubiquity in the clinic. Modern highly-parallel hardware like graphics processing units (GPUs) may offer the computational resources to solve these reconstruction problems quickly, but simply &#34translating&#34 existing algorithms designed for conventional processors to the GPU may not fully exploit the hardware&#39s capabilities.
This thesis proposes GPU-specialized image denoising and image reconstruction algorithms. The proposed image denoising algorithm uses group coordinate descent with carefully structured groups. The algorithm converges very rapidly: in one experiment, it denoises a 65 megapixel image in about 1.5 seconds, while the popular Chambolle-Pock primal-dual algorithm running on the same hardware takes over a minute to reach the same level of accuracy.
For X-ray CT reconstruction, this thesis uses duality and group coordinate ascent to propose an alternative to the popular ordered subsets (OS) method. Similar to OS, the proposed method can use a subset of the data to update the image. Unlike OS, the proposed method is convergent. In one helical CT reconstruction experiment, an implementation of the proposed algorithm using one GPU converges more quickly than a state-of-the-art algorithm converges using four GPUs. Using four GPUs, the proposed algorithm reaches near convergence of a wide-cone axial reconstruction problem with over 220 million voxels in only 11 minutes.