Metal artifact reduction in X-ray CT based on CNN using model based reconstruction results

Abstract

Metal artifacts in CT image disturb accurate diagnosis. To date, many iterative and direct reconstruction methods have been developed to reduce metal artifacts. In this thesis, we address convolutional neural networks (CNN) based computed tomography (CT) image reconstruction. We verify that our algorithm can reduce metal artifacts in the image similar to those used in learning. However, if our algorithm encounters a new image that is not in the class of images for learning, it cannot beat the existing model based reconstruction results. Although our algorithm does not work properly when meeting new images, better model based reconstruction input data produces better output image.