An Image Inpainting Based Surrogate Data Strategy for Metal Artifact Reduction in CT Images

The goal of this work is the reduction of metal artifacts in reconstructed CT images. Mathematically, those artifacts are caused by inconsistencies in the Radon space. The metal-artifact reduction (MAR) algorithm presented here is based upon an idea adapted from image inpainting, a technique to repair damaged films and photographs. Here, it is used to restore the inconsistent projection data in the Radon space in such a way that the gap, caused by ignoring the inconsistencies, is undetectable at the end. This method is compared to a classically used one-dimensional linear interpolation inside one projection under one view and to a directional interpolation taking the flow of the surrounding projection data into account when calculating the artificial sinogram data. The best result is achieved with the two-dimensional PDE-based image inpainting approach. Due to the fact that the repaired sinogram data are still afflicted with residual inconsistencies - depending on the used interpolation strategy - a weighted MLEM algorithm is used to reconstruct the CT images. Here, those artificially generated sinogram projections are weighted less. The proposed MAR method is evaluated on sinogram data from an anthropomorphic torso phantom marked with two steel markers. Also raw data of the same cross section without the markers were acquired, which serve as ground truth in the evaluation of the metal-artifact reduction quality.