LSBMAR: Latent Schrödinger Bridge for Metal Artifact Reduction in CT Scans

Computed Tomography (CT) scans are often degraded by metal artifacts arising from implants or prosthetic devices, which manifest as bright and dark streaks that compromise diagnostic interpretability. Current Metal Artifact Reduction (MAR) methods typically treat this task as sonogram-domain inpainting, using the image domain as a guide for enforcing consistency. Recently, Diffusion Models (DMs) have been employed and show improved performance. However, because these models are applied directly in the high-dimensional spaces, they remain computationally demanding, while the well-established baselines struggle to achieve the same results as diffusion models. Thus, to overcome these limitations, in this work, we propose a novel MAR framework based on Latent Diffusion Models (LDMs) that reinterprets the task as an image-to-image translation problem in latent space. By leveraging the Schrödinger Bridge formulation, our method learns to map corrupted latent representations to their artifact-free counterparts, conditioned directly on the distribution of the degraded images. This design maintains the expressiveness and yields to a more efficient generation process while preserving anatomical fidelity. Experimental results demonstrate that our approach achieves competitive performance compared to traditional MAR baselines, effectively reducing artifacts and enhancing image quality.