Learning Effective Parameter Settings for Iterative CT Reconstruction Algorithms

—Iterative reconstruction algorithms are preferably used when the projection data are noisy, irregular in acquisition, or limited in number or size. They typically offer a set of parameters that allow some control over the convergence process, both in terms of quality and speed. Examples include relaxation factor, number of subsets, regularization coefficients, and the like. The interactions among these parameters and within the various data conditions can be complex, and thus effective combinations can be difficult to identify, leaving their choice often to educated guesses. We propose a data-driven learning approach to match given data configurations with their most effective reconstruction parameter configurations. We overcome the computational challenges associated with such a data-intensive approach by using commodity high-performance computing hardware (GPUs), which themselves have interacting parameters as well.