Structure-Preserving Two-Stage Diffusion Model for CBCT Metal Artifact Reduction.
Authors
Abstract
Cone-beam computed tomography (CBCT) plays a crucial role in dental clinical applications, but metal implants often cause severe artifacts, challenging accurate diagnosis. Most deep learning-based methods attempt to achieve metal artifact reduction (MAR) by training neural networks on paired simulated data. However, they often struggle to preserve anatomical structures around metal implants, and fail to bridge the domain gap between real-world and simulated data, leading to suboptimal performance in practice. To address these issues, we propose a two-stage diffusion framework with a strong emphasis on structure preservation and domain generalization. In Stage I, a structure-aware diffusion model is trained to extract artifact-free clean edge maps from artifact-affected CBCT images. This training is supervised by the tooth contours derived from the fusion of intraoral scan (IOS) data and CBCT images to improve generalization to real-world data. In Stage II, these extracted clean edge maps serve as structural priors to guide the MAR process. Additionally, we introduce a segmentation-guided sampling (SGS) strategy in this stage to further enhance structure preservation during inference. Experiments on both simulated and real-world data demonstrate that our method achieves superior artifact reduction and better preservation of dental structures compared to competing approaches.