Physics-Constrained Sinogram Inpainting for CT Metal Artifact Reduction

Computed tomography (CT) is extensively employed in medical diagnosis and treatment guidance. However, when patients carry metal implants, the reconstructed CT images often suffer from severe artifacts caused by beam hardening effect. In such cases, the acquired projection data violate Tuy’s data sufficiency condition, making analytical reconstruction infeasible. To address this, a wide range of deep learning–based metal artifact reduction (MAR) methods have been developed, demonstrating impressive performance. Yet, existing approaches are purely data-driven and depend on large collections of reference images for manifold approximation. In this work, we introduce a physics-driven sinogram inpainting approach that leverages the inherent correlations in projection data during CT acquisition for MAR.