Machine-learning-based nonlinear decomposition of CT images for metal artifact reduction

Computed tomography (CT) images containing metallic objects commonly show severe streaking and shadow artifacts. Metal artifacts are caused by nonlinear beam-hardening effects combined with other factors such as scatter and Poisson noise. In this paper, we propose an implant-specific method that extracts beam-hardening artifacts from CT images without affecting the background image. We found that in cases where metal is inserted in the water (or tissue), the generated beam-hardening artifacts can be approximately extracted by subtracting artifacts generated exclusively by metals. We used a deep learning technique to train nonlinear representations of beam-hardening artifacts arising from metals, which appear as shadows and streaking artifacts. The proposed network is not designed to identify ground-truth CT images (i.e., the CT image before its corruption by metal artifacts). Consequently, these images are not required for training. The proposed method was tested on a dataset consisting of real CT scans of pelvises containing simulated hip prostheses. The results demonstrate that the proposed deep learning method successfully extracts both shadowing and streaking artifacts.