Using GPUs to learn effective parameter settings for GPU-accelerated iterative CT reconstruction algorithms

This chapter introduces a framework based on a GPU-accelerated iterative reconstruction platform with regularization to search the optimal parameter settings for the reconstruction process itself satisfying multiple performance objectives. In computed tomography (CT), the filtered backprojection (FBP) algorithm is most widely used for the reconstruction of an object from its X-ray projections. It considers each projection only once and is therefore fast to compute. However, iterative methods are computationally expensive, but they are now becoming feasible owing to the vast computational performance of GPUs. Iterative methods typically offer a diverse set of parameters that allow control over quality and computation speed, often requiring trade-offs. The interactions among these parameters can be complex, and thus effective combinations can be difficult to identify for a given data scenario. The optimization procedure exploits fast GPU-based computation to generate computation products that it then automatically evaluates using a perceptually based fitness function, implementing a simulated human observer, together with the recorded execution time, to steer the optimization. As future work, the machine-learning procedure could further be accelerated in GPUs, and the regularization parameters could also be studied in the same way. © 2011