Combination of voxel-based and projection-based methods in terms of convergence for CT reconstruction.

Abstract

Recent applications of iterative image reconstruction algorithms to multislice helical CT have shown that iterative reconstruction can significantly improve image quality and reduce artifacts. In this paper, the authors introduce a combination of two different algorithms with different convergence properties: ordered subsets separable paraboloidal surrogates (OS-SPS) and iterative coordinate descent (ICD). The first one updates image voxels simultaneously, slightly changing attenuation values iteration by iteration. The second algorithm updates image voxel by voxel, each time performing full forward and backward projections of the voxel. It has been shown that ICD converges better at high-frequency areas and requires more iterations to reconstruct low-frequency components of the image. In contrast to ICD, SPS requires multiple iterations to reconstruct high-frequency areas. In this paper, the authors introduce an algorithm which leverages the benefits of both ICD and SPS. The idea is to update the entire image with SPS, determine high-frequency components, and focus ICD computations on it using nonhomogeneous ICD update. The authors have successfully implemented OS-SPS, ICD, their hybrid approach, and few variations of ICD based on spatially nonuniform updates. The authors have examined the convergence of different algorithms and found that proposed algorithm converges better than OS-SPS, ICD, as well as various improved variants of ICD.