Distributed iterative image reconstruction for micro-CT with ordered-subsets and focus of attention problem reduction

Abstract

Iterative reconstruction algorithms are well-known to produce high quality images. In this paper, we describe how to make these computationally demanding algorithms feasible in connection with high-resolution, cone-beam x-ray computed tomography (micro-CT) for small animals. First we outline a cost-effective approach that supports distributing the computation across a cluster of inexpensive dual-processor PCs. To a great extent, the limit on the number of cluster nodes that can be used is determined by the overhead associated with the increased number of interprocessor communications. We then introduce ordered-subsets to accelerate convergence of the reconstruction algorithm thereby reducing the number of required iterations. Finally, we use a method called focus of attention to automatically segment the projection and image data into object and background. By subsequently considering only the object data, we reduce the overall cost of the on-the-fly system matrix computation, the forward and backprojection data updates, and the global reduction operation that facilitates the interprocessor communication. We use the SIRT algorithm to illustrate our work but the methods and results presented apply to iterative reconstruction algorithms in general. The experimental data consists of the three-dimensional Shepp-Logan phantom as well as mouse data obtained from a MicroCAT™ scanner.