Deep Learning Based Metal Inpainting in the Projection Domain: Initial Results

Abstract

During surgical interventions mobile C-arm systems are used in order to evaluate the correct positioning of e.g. inserted implants or screws. Besides 2D X-ray projections, that often do not suffice for a profound evaluation, new C-arm systems provide 3D reconstructions as additional source of information. However, mainly due to metal artifacts, this additional information is limited. Thus, metal artifact reduction methods were developed to resolve these problems, but no generally accepted approaches have been found yet. In this paper, three different network architectures are presented and compared that perform an inpainting of metal corrupted areas in the projection domain in order to tackle the problems of metal artifacts in the 3D reconstructions. All network architectures were trained using real data and thus all observations should hold during inference in real clinical applications. The network architectures show promising inpainting results with smooth transitions with the non-metal areas of the images and thus homogeneous image impressions. Furthermore, this paper shows that providing additional input data to the network, in form of a metal mask, increases the inpainting performance significantly.