UPGRADE-Net: Unsupervised Sinogram-Domain Data-Consistent Network for Metal Artifact Reduction

Area of Science:

• Medical Imaging
• Computer Vision
• Artificial Intelligence

Background:

• Computed tomography (CT) is crucial for clinical diagnosis but suffers from metal artifacts caused by implants.
• Existing supervised deep learning methods for metal artifact reduction (MAR) struggle with generalization due to the need for paired artifact-affected and artifact-free data.
• Current MAR methods lack sinogram-domain data consistency for accurate metal trace inpainting.

Purpose of the Study:

• To develop an unsupervised deep learning framework for metal artifact reduction (MAR) in CT imaging.
• To address the limitations of supervised methods by eliminating the need for artifact-free ground truth data.
• To ensure sinogram-domain data consistency for precise metal trace restoration.

Main Methods:

• Proposed UPGRADE-Net, an unsupervised sinogram-domain data-consistent network for MAR.
• Utilized a generative conditional diffusion model guided by prior knowledge for metal trace inpainting.
• Developed a deep unsupervised MAR framework in the reverse process to learn background data distribution.
• Incorporated physics-based conjugate-ray and accumulation-ray consistency loss functions for sinogram-domain data consistency.

Main Results:

• UPGRADE-Net effectively reduces metal artifacts in CT scans.
• The method demonstrates strong performance on both public and clinical datasets.
• Experimental results show superiority over state-of-the-art MAR techniques.
• Achieved accurate metal trace restoration in the sinogram domain.

Conclusions:

• UPGRADE-Net offers a robust unsupervised solution for metal artifact reduction in CT.
• The proposed method overcomes the data acquisition challenges of supervised approaches.
• UPGRADE-Net enhances clinical diagnosis by improving CT image quality in the presence of metallic implants.