PRAISE-Net: Deep Projection-Domain Data-Consistent Learning Network for CBCT Metal Artifact Reduction

Abstract

High-attenuation metal implants can cause metal artifacts in cone-beam computed tomography (CBCT) scanning due to their strong and energy-dependent photon-absorption ability. These artifacts severely degrade image quality in intraoperative radiotherapy and postoperative diagnosis for clinical physicians. The conventional projection-domain metal artifact reduction (MAR) methods for fan-beam geometry are not efficiently applicable to CBCT MAR, because metallic implants in cone-shaped X-ray beam scanning lead to serious data missing at all the projection views. To tackle the aforementioned challenge, we present a novel projection-domain data-consistent learning network, i.e., PRAISE-Net, to suppress CBCT metal artifacts. First, a Low2High strategy which inpaints metal traces at low resolution and restores high-resolution results with a super-resolution reconstruction (SRR) module is proposed to reduce computational costs. Second, a PIG-DDPM module with the prior knowledge is designed for fine-grained projection-domain metal area inpainting. Third, a CBCT domain adaptation (CBCT-DA) is incorporated in the PIG-DDPM to step across the gap between simulated data and clinical CBCT data. The proposed PRAISE-Net is trained and evaluated on a publicly available dataset and a private clinical CBCT dataset, and our results confirm that the proposed method outperforms the state-of-the-art competing methods. This efficient, accurate, and reliable CBCT MAR technique has a great potential for clinical applications.