A nonconvex truncated regularization and box-constrained model for CT reconstruction (Q2198001)

The study is position in the realm of low-dose CT reconstruction completed with the use of nonconvex truncated regularization. The imaging model is expressed in the additive form $g =Au+e$ where $A$ stands for the projection matrix, $g$ and $u$ are the observed projection and the original image, respectively whereas $e$ is a Gaussian noise. The proposed recovery procedure of $u$ from $g$ is formulated with the aid of a certain regularization model. Theoretical analysis of the model with respect to the properties of local minimizers is provided followed by an iterative support shrinking with proximal general linearization algorithm. The properties of the method (sufficient decrease property) are presented. Experimental results are included along with a comparative analysis of the results produced by two other methods (TV-based model and the tight frame wavelet-based model) for several noise levels.