Framework for iterative cone-beam micro-CT reconstruction

We have developed a computational framework for iterative cone-beam micro-CT reconstruction. When applied to small animal imaging and other high-resolution applications, iterative algorithms pose significant implementation difficulties due to the associated high computational burden. To address this burden, we use threads and message passing (MPI) to facilitate multiprocessor cluster computing. The system matrix can be either precomputed or calculated on-the-fly. Additionally, the system matrix can be based on trilinear interpolation or volumetric intersection methods. In the case of storing the system matrix, we exploit symmetries in the cone-beam geometry to reduce the storage requirements by a factor of nearly eight. We chose simultaneous iterative reconstruction technique (SIRT) to demonstrate the framework, although other algorithms are easily implemented. When precomputing the system matrix, each iteration of SIRT completes several times faster than when using an on-the-fly system matrix computation. The precomputed method remains faster when implementing ordered subsets, but the performance advantage diminishes as the number of subsets increases. We present reconstructions of the three-dimensional (3-D) Shepp-Logan head phantom as well as mouse data acquired by a small animal X-ray CT scanner.