Nonlinear backprojection for tomographic reconstruction

Abstract

This paper focuses on a tomographic image reconstruction method, which will be referred to as nonlinear backprojection (NBP). Rather than explicitly statistically modeling the forward process and the unknown image, we train an optimal NBP operator that can be implemented noniteratively. Under appropriate assumptions, the method forms its estimate by applying nonlinear filters to sinogram data, followed by conventional backprojection. The nonlinear filters are designed through off-line training. NBP shows promising results relative to both filtered backprojection and maximum a posteriori probability Bayesian methods.