Abstract: Metal artifacts adversely affect computed tomography (CT) image quality and diagnostic accuracy. Metal-artifact reduction (MAR) in CT images has long been a major focus of research. In recent years, with the advancement and application of deep-learning technologies, new approaches have emerged for research on MAR algorithms, leading to a wealth of outstanding achievements. In this paper, we first introduce the causes and manifestations of metal artifacts in CT images. We then review recent progress in deep-learning-based MAR methods, categorizing them into three approaches: image, projection, and dual domains. Finally, we summarize these methods and discuss future research prospects for MAR technology.