Fast Ordered Subset Convex iterative panel CT reconstruction

Iterative reconstruction has several advantages over analytical cone-beam reconstruction methods for three-dimensional x-ray computed tomography. Iterative methods tend to incorporate both system and statistical models which yield a better signal to noise ratio. Iterative reconstruction also produces adequate image quality at low count, enabling low-dose X-ray CT studies, at the expense of computational overhead. This work presents an efficient implementation of an Ordered Subset Convex (OSC) algorithm. OSC algorithm requires one forward projection, exponential of the forward projection of the current image, and two back projections per iteration which makes it very computational intensive. Optimizing the forward projector starts by solving a 3D line equation with rotation by using the system geometry (symmetry). A voxel weighting method was also applied during forward projection to achieve better accuracy. Data was aligned in the Z,X,Y directions for cache optimization and sixteen voxels were forward projected at the same time using SIMD registers: four in both z and x-y directions. The back projection optimization was achieved by performing both back projections at the same time. Data were first aligned and both back projections were performed in parallel on four projection bins using SIMD. The algorithm was implemented on a small animal flat panel rotating CT from the Inveon System (Siemens Preclinical Solutions, Knoxville, TN.). Phantom data was acquired in ¿rebin by four¿ mode (i.e. 512 × 512 pixels per projection) with 360 projections and reconstructed with both OSC and the standard Feldkamp method. Both reconstructions were performed on a dual X5450 quad core CPUs with 32 GB RAM and 64 bit operating system. Using this fast OSC implementation, a reconstruction time of 210 s was achieved using 18 subsets and 4 iterations (i.e. 52.5 s per iteration).