Parallel Sinogram And Image Framework With Co-Training Strategy For Metal Artifact Reduction In Tooth Ct Images

Computed Tomography (CT) is widely used in oral treatment planning but metal artifacts caused by high-density materials such as metal implants heavily influence the effectivness of digital tooth models.In existing studies on metal artifact reduction (MAR), the mathematical relationship between spatial domain and projection domain is generally sequentially considered, or metal traces/masks are required as priors. In this paper, we propose a parallel sinogram and image framework (PSIF), aiming to enable MAR in the spatial and projection domains to benefit each other. We formulate this task as an image enhancement problem in the spatial domain and a sinogram completion problem in projection domain using two parallel networks, and propose a co-training strategy with forward-backward projection consistency loss to optimize the model. The experimental results on $10,000$ tooth slices demonstrate that our proposed method can effectively recover tooth outlines and suppress the stripe artifacts.