Metal artifact reduction in 2D CT images with self-supervised cross-domain learning

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Radiology

Background:

• Metallic implants in CT scans cause severe artifacts, hindering diagnosis and radiation therapy planning.
• Current metal artifact reduction (MAR) methods often require paired images (corrupted and artifact-free), limiting their practical application.

Purpose of the Study:

• To develop a novel deep-learning-based approach for metal artifact reduction (MAR) in CT images.
• To propose a self-supervised cross-domain learning framework that alleviates the need for anatomically identical CT image pairs.

Main Methods:

• A neural network was trained to restore metal trace regions in metal-free sinograms using forward projection of metal masks.
• A filtered backward projection (FBP) reconstruction loss and a residual-learning-based image refinement module were employed.
• Metal-affected projections were replaced with sinograms from CNN output before final FBP reconstruction.

Main Results:

• The proposed method demonstrated superior performance in metal artifact reduction on both simulated and real data.
• The framework effectively reduced secondary artifacts and preserved fine structural details in the reconstructed CT images.
• The approach outperformed existing compelling MAR methods.

Conclusions:

• The self-supervised cross-domain learning framework offers an effective solution for MAR in CT imaging.
• The method shows significant potential for improving image quality and clinical utility across various organ sites.
• This deep learning approach advances the field of artifact reduction in medical imaging.