Statistical Cone-Beam CT Image Reconstruction using the Cell Broadband Engine

CT images can be reconstructed analytically or iteratively. The analytic methods, e.g. filtered backprojection, are known to be computationally inexpensive and highly accurate. Iterative reconstruction seems of high interest since better dose usage is expected. However, iterative methods are computationally extremely expensive and therefore have been applied to modalities with low amounts of data (e.g. PET) only. A promising algorithm for CT is the ordered subset convex (OSC) whose initial design has recently been significantly improved and now achieves high image quality. Recently, a novel general purpose architecture optimized for distributed computing became available: the Cell Broadband Engine (CBE). Its eight synergistic processing elements (SPEs) currently allow for a theoretical performance of 192 GFlops. We aim at maximizing the OSC image reconstruction speed for flatpanel-based cone-beam CT such as micro-CT or C-arm-CT. For this geometry highly optimized perspective CBE-based cone-beam forward-and backprojection algorithms were designed and implemented. Performance was assessed by reconstructing a 512³ volume from 512 cone-beam projections of size 1024². In combination with a preceding Feldkamp-type initialization, four OSC iterations turned out to be sufficient to achieve high image quality. Using both CBEs of our dual Cell-based blade (Mercury Computer Systems) allows to reconstruct the whole volume in about one minute.