Author: Shusen Jing, Qihui Lyu, Dan Ruan, Ke Sheng, Qifan Xu 👨🔬
Affiliation: Department of Radiation Oncology, University of California, Los Angeles, University of California San Francisco, Department of Radiation Oncology, University of California, San Francisco 🌍
Purpose: Metallic implants can significantly distort sinograms, leading to severe artifacts in computed tomography (CT) reconstructions. Reconstructing CT images containing metal is fundamentally an ill-conditioned inverse problem that cannot be accurately solved without additional prior information. Traditional approaches enforcing piecewise smoothness through regularization fail to capture intricate anatomical structures contaminated by the metal artifacts. In contrast, machine learning methods, particularly generative models, have demonstrated the potential to accurately model CT image distributions.
Methods: We proposed a mask guided diffusion model for MAR (MaskDM). Specifically, we adopted a pretrained diffusion model on Deeplesion dataset containing 900,000 clinic images with different anatomy to learn the prior distribution of CT images. For each metal-corrupted CT image, we generate a soft error mask from the corrupted sinogram, which indicates the reliability of the pixels in the metal-corrupted CT image. During the denoising process of the diffusion model, the intermediate results are fused with the metal-corrupted CT image based on the soft error mask. Additionally, we incorporate a sinogram alignment technique to ensure that the forward projection of the intermediate results aligns with the non-metal regions of the raw sinogram measurements.
Results: We test the proposed methods with 200 images in Deeplesion dataset with artificial metal implants. The MAR performance was compared with conventional methods, such as linear intepolaration (LI), and state-of-the-arts learning-based methods, such as DuDoDp.The proposed MaskDM signifcatly suppressed the metal artifacts, while LI introduced severe secondary artifacts due to its modification on the metal corrupted sinogram. Moreover, compared with DuDoDp, the proposed MaskDM is better at preserving details of anatomy during the generation. Our method improves the PSNR by 3.4 dB compared with LI, and 0.4 dB compared with DuDoDp.
Conclusion: Leveraging the prior CT image distribution, the proposed MaskDM significantly reduces the metal artifacts and enhances the CT accuracy.