Passty framework of Chambolle-Pock for Fast and High accuracy CT reconstruction

Abstract

Abstract To realize low-dose CT, many iterative reconstruction methods have been proposed, but many iterations and high computational complexity are required for each scan. It is necessary to reduce the number of iterations while effectively suppressing artifacts and noise caused by reducing radiation dose for practical use. This study aims to accelerate the algorithm by improving the Chambolle-Pock (CP) algorithm, the latest first-order primal-dual. First, we introduce the Passty framework into the CP algorithm. The proposed CP becomes a row-action type algorithm, enabling the incorporation of ordered subsets to accelerate the algorithm further. Second, to step up with the latest trends, the regularization term (nonlocal TV) is designed with a combined model of the first and second-order derivatives to preserve smooth intensity changes. Third, we extend the proposed CP to the L1 and Huber data-fidelity terms for more practical application. Most CT reconstruction studies employ the L2 data-fidelity term because of its good convergence. Even if the L1 data-fidelity term has tremendous potential in noise immunity, it has a fatal drawback of poor convergence and is not adopted at present. In the proposed CP, the L1 and Huber data-fidelity terms have excellent image quality and convergence performance than the L2.