1. INTRODUCTION
A Mobile C-arm X-ray system is a medical imaging device frequently used in current interventional surgery, which supports the accurate placement of metallic implants or screws. However, some physical effects like photon starvation and beam hardening occur when x-rays pass through metals, resulting the bright and dark streak artifacts, which reduce the quality of the reconstructed image [1]. The conventional metal artifacts reduction (MAR) algorithms tackle the problem by sinogram completion [2], [3] and iterative reconstruction [4], [5]. Deep learning-based methods are also applied in MAR [6], [7], [8], [9], [10], and the models are trained on paired data in a supervised way. For the C-arm system, since only the central slice in cone-beam geometry can be represented as a sinogram and other artifacts like truncation artifacts also exist in reconstructions, image inpainting in the projection domain is the only feasible method for MAR of cone-beam computed tomography (CBCT).