Blended finite element method and its convergence for three-dimensional image reconstruction using \(L^2\)-gradient flow (Q482261)

In computed tomography a very important problem is image reconstruction, i.e., producing a 2D or 3D image from a large number of line-integral projections from different views. In the literature one can find many methods of image reconstruction. Recently, researchers focused on iterative algorithms because they often yield better image quality than the classical methods for sparse-view and high-noise detected data. The authors of the paper present a new iterative algorithm -- the gradient-flow-based blended finite element method. The presented method includes as its special cases the explicit and semi-implicit finite element methods, that were developed in the previous works of the authors. The paper starts with some mathematical preliminaries about image reconstruction. Next, a reconstruction model is introduced obtaining some Euler-Lagrange equation. To solve the equation the authors use the gradient flow to transform the original equation and use the blended finite element method to solve it. After a detailed derivation of the solution, the authors give a numerical algorithm of the proposed method. In the algorithm there is a need to multiply a not sparse matrix of large dimension with a vector, so the authors propose another representation of the multiplication to obtain lesser demand on the memory. Next, the authors analyse the computational complexity of the proposed algorithm and discuss the selection of the parameter's value that occurs in the algorithm. Then, a theoretical analysis of the algorithm is given. Finally, several numerical experiments are presented. In the first experiment, the authors show the effect of regularizer for noisy projection images. The second experiment illustrates that the proposed algorithm is convergent. In the last experiment the authors compare the efficiency of the algorithm with the explicit finite element method and semi-implicit finite element method. For this purpose they use a real dataset.