Optimization-based image reconstruction with artifact reduction in C-arm CBCT

Abstract

We investigate an optimization-based reconstruction, with an emphasis on image-artifact reduction, from data collected in C-arm cone-beam computed tomography (CBCT) employed in image-guided interventional procedures. In the study, an image to be reconstructed is formulated as a solution to a convex optimization program in which a weighted data divergence is minimized subject to a constraint on the image total variation (TV); a data-derivative fidelity is introduced in the program specifically for effectively suppressing dominant, low-frequency data artifact caused by, e.g., data truncation; and the Chambolle-Pock (CP) algorithm is tailored to reconstruct an image through solving the program. Like any other reconstructions, the optimization-based reconstruction considered depends upon numerous parameters. We elucidate the parameters, illustrate their determination, and demonstrate their impact on the reconstruction. The optimization-based reconstruction, when applied to data collected from swine and patient subjects, yields images with visibly reduced artifacts in contrast to the reference reconstruction, and it also appears to exhibit a high degree of robustness against distinctively different anatomies of imaged subjects and scanning conditions of clinical significance. Knowledge and insights gained in the study may be exploited for aiding in the design of practical reconstructions of truly clinical-application utility.

A Doubly Constrained TV Algorithm for Image Reconstruction

“…In equation 21, ‖ • ‖ ∞ is the ∞-norm of a vector, which selects the largest element in the vector. In equation 22, ProjectOntoℓ 1 Ball a (x) is a projection operator which may project a vector x onto the ℓ 1 ball of radius a [20,22]. 1 I is a "1" vector in space I � R N .…”

“…Substituting equations (18), (20), (22), and 23into Algorithm 1, we get Algorithm 2, the CP algorithm instance of the dcTV model.…”

“…is model parameter is not as sensitive as that in DDcTV, which is why the TVcDM was later proposed as an improved model in 2014 [20] and has since been the subject of extensive studies. e model has been successful in short scan CBCT [21], C-arm CBCT [22], positron emission tomography (PET) [23], and EPRI [24]. In equation (5), there are two constraint terms, the data divergence constraint and the TV constraint.…”

“…ere are currently three main solving algorithms: ASD-POCS, alternating direction method of multipliers (ADMM) [14] (split Bregman [28]), and Chambolle-Pock (CP) algorithm [18,[29][30][31][32][33]. ASD-POCS has been used to solve the DDcTV model; ADMM or split Bregman are usually used to solve the ucTV model; the CP algorithm has been used to solve DDcTV and TVcDM models [18,[20][21][22][23]. e CP algorithm has certain advantages relative to the ADMM or ASD-POCS algorithms: (1) it may solve all the convex optimization models whether or not they are smooth; (2) each subproblem has a closed form; and (3) all of the algorithm parameters in the CP algorithm may be explicitly determined analytically (i.e., via solving equations) rather than empirically selecting suitable values [31].…”

Adversarial Sparse-View CBCT Artifact Reduction

“…However, to our best knowledge, no scheme has been proposed to correct the cone-beam artifacts introduced by sparse-view CBCT reconstruction in a post-processing step. Instead of reducing artifacts from the CBCT images directly, many other systems [11,14] propose to introduce better sparse-view reconstruction methods that yield less artifacts. Although encouraging improvements have been made, the image quality from the current solutions are still not satisfactory when only a small number of views are used.…”