Quad-Net: Quad-Domain Network for CT Metal Artifact Reduction

Area of Science:

• Medical Imaging
• Computer Vision
• Artificial Intelligence

Background:

• Metal implants cause severe streaking artifacts in CT images, degrading quality and diagnostic performance.
• Existing metal artifact reduction (MAR) methods, including dual-domain deep networks, still face clinical challenges.

Purpose of the Study:

• To develop an advanced MAR method that synergizes features from multiple domains for optimal artifact elimination.
• To improve CT image quality without compromising subtle structural details.

Main Methods:

• Proposed a quad-domain deep network (Quad-Net) integrating sinogram, image, and their Fourier domains.
• Developed a Sinogram-Fourier Restoration Network (SFR-Net) for inpainting corrupted sinogram data.
• Coupled SFR-Net with an Image-Fourier Refinement Network (IFR-Net) for cross-domain image enhancement.

Main Results:

• Quad-Net demonstrated superior performance over state-of-the-art MAR methods quantitatively, visually, and statistically.
• The method effectively eliminates metal artifacts without requiring precise metal masks.
• Achieved optimal artifact reduction by learning global and local features and their relations across four receptive fields.

Conclusions:

• Quad-Net offers a significant advancement in CT metal artifact reduction, enhancing diagnostic capabilities.
• The proposed quad-domain approach provides a robust and efficient solution for challenging clinical cases.
• The method's ability to work without metal masks is crucial for practical clinical application.