CBCT projection domain metal segmentation for metal artifact reduction using hessian-inspired dual-encoding network with guidance from segment anything model.

Metal artifact is a prevailing factor reducing the image quality of cone-beam computed tomography (CBCT), which is a widely used medical imaging method. Existing metal artifact reduction (MAR) methods typically contain two steps: segmentation and interpolation. Recent MAR algorithms pay more attention to the interpolation of the metal traces, but metal segmentation is also challenging, especially for CBCT. Despite the success of deep learning (DL) in image segmentation, the substantial expense associated with annotating metal traces in the projection domain makes most of these approaches impractical for this task. In this paper, we aim to provide a workflow for DL-based metal-trace segmentation without manually delineated ground truth. We propose a Hessian-inspired dual-encoding network (HIDE-Net) for CBCT projection-domain metal segmentation with guidance from the segment anything model. Specifically, a Hessian eigenvalue module is designed to incorporate human knowledge about the target metal objects; a dual encoder is designed to better extract marginal information; and an input enhancement module is proposed to enhance the projection domain input for better segmentation. Finally, a SAM-based label preprocessing module is investigated to obtain the training label automatically. The proposed method has been tested on both digital phantom data and clinical CBCT data. Experiments on both datasets demonstrate the efficacy of the proposed method. HIDE-Net achieves improved metal segmentation accuracy than recent segmentation-oriented CNN models. Compared with existing MAR algorithms, the proposed method improves Dice index in projection domain by 3.2 %$\%$ , and the RMSE in image domain is reduced by 42 %$\%$ . The proposed methods would advance MAR techniques in CBCT and have the potential to push forward the use of intraoperative CBCT in human-handed and robotic-assisted MISS.