Deep residual learning enabled metal artifact reduction in CT

Many clinical scenarios involve the presence of metal objects in the CT scan field-of-view. Metal objects tend to cause severe artifacts in CT images such as shading, streaks, and a loss of tissue visibility adjacent to metal components, which is often the region-of-interest in imaging. Many existing methods depend on synthesized projections and classification of in-vivo materials whose results can sometimes be subject to error and miss details, while other methods require additional information such as an accurate model of metal component prior to reconstruction. Deep learning approaches have advanced rapidly in recent years and achieved tremendous success in many fields. In this work, we develop a deep residual learning framework that trains a deep convolution neural network to detect and correct for metal artifacts from image content. Training sets are generated from simulation that incorporates modeling of physical processes related to metal artifacts. Testing scenarios included the presence of a surgical screw within the transaxial plane and two rod implants in the craniocaudal direction. The proposed network trained by polychromatic simulation data demonstrates the capability to largely reduce or, in some cases, almost entirely remove metal artifacts caused by beam hardening effects. The proposed method also showed largely reduced metal artifacts on data collected from a multi-slice CT system. These findings suggest deep residual learning enabled methods present a new type of promising approaches for reducing metal artifacts and support further development of the method in more clinically realistic scenarios.