Fast and high quality model based iterative reconstruction for computed tomography
Abstract
Recently it has been shown that model-based iterative reconstruction (MBIR) can greatly improve the quality of computed tomography. This research focuses on two aspects of MBIR algorithm: reducing the computation of reconstruction and improving reconstructed image quality. In order to make MBIR practical for daily clinical application, it is critical to reduce the total time required for reconstruction. To this end, we have developed several fast reconstruction algorithms. First, we have proposed the functional substitution method that can solve the 1D optimization problem in one step, so that the computation for each voxel update is significantly reduced. Second, we have developed a non-homogeneous iterative coordinate descent (NHICD) algorithm which can speed convergence of the iterative MBIR algorithms by focusing computation on the voxels of greatest importance. Third, we have proposed a multi-resolution framework that can speed targeted reconstructions by allocating computation based on a voxel's contribution to the region of interest. Fourth, we have developed the edge-localized ICD algorithm that can reconstruct high-resolution images at a computational cost similar to a low-resolution reconstruction. The edge-localized ICD works by concentrating computation on the regions of the image containing fine details, such as edges. In addition to speeding reconstruction, we have also developed a variety of methods to improve image quality for MBIR through the introduction of more accurate system and prior models. First, we improved the uniformity and isotropy of MBIR images by developing a novel method to calculate the coefficients of the prior model. Second, we introduced a kinetic model for MBIR which can reduce motion artifacts through the use of a kinetic parameter reconstruction algorithm.
Degree
Ph.D.
Advisors
Bouman, Purdue University.
Subject Area
Electrical engineering
Off-Campus Purdue Users:
To access this dissertation, please log in to our
proxy server.