Deep Sinogram Completion With Image Prior for Metal Artifact Reduction in CT Images

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computational Imaging

Background:

• Computed tomography (CT) is vital for medical diagnosis and treatment planning.
• Metallic objects in CT scans cause severe artifacts, compromising image quality and clinical accuracy.
• Existing metal artifact reduction (MAR) methods have limitations.

Purpose of the Study:

• To develop a generalizable framework for metal artifact reduction (MAR) in CT images.
• To leverage both image and sinogram domain techniques for enhanced artifact removal.
• To improve the accuracy of CT-based diagnosis and radiation therapy planning.

Main Methods:

• A novel framework combining image and sinogram domain MAR techniques.
• Formulation as a sinogram completion problem using a neural network (SinoNet).
• Integration of a prior image generation network (PriorNet) and residual sinogram learning.
• End-to-end joint training with differentiable forward projection and filtered backward projection reconstruction.

Main Results:

• The proposed method effectively reduces metal artifacts in CT images.
• Superior artifact reduction compared to existing MAR techniques.
• Preservation of anatomical structures in the reconstructed images.
• Successful validation on both simulated and real-world artifact data.

Conclusions:

• The developed framework offers a significant advancement in MAR for CT imaging.
• The joint deep learning approach enhances artifact removal while maintaining image fidelity.
• This method has the potential to improve clinical diagnosis and radiation therapy outcomes.