SemiMAR: Semi-Supervised Learning for CT Metal Artifact Reduction.

Area of Science:

• Medical Imaging
• Computer Vision
• Artificial Intelligence

Background:

• Metal artifacts degrade CT image quality, impacting diagnosis.
• Deep learning (DL) methods for metal artifact reduction (MAR) face domain gap issues with clinical data.
• Existing semi-supervised MAR methods often require large model sizes.

Purpose of the Study:

• To propose a novel, efficient semi-supervised MAR framework.
• To address the challenge of large model sizes in semi-supervised MAR.
• To improve MAR performance on clinical CT data.

Main Methods:

• A semi-supervised framework that learns artifact-free image regions.
• Artifacts are inferred by subtracting learned clean parts from corrupted images.
• Knowledge distillation from a dual-domain MAR network in image and latent spaces using contrastive learning.

Main Results:

• The proposed method utilizes a single generator, reducing model scale.
• It effectively narrows the domain gap between simulated and clinical data.
• Experiments show favorable qualitative and quantitative results compared to state-of-the-art methods.

Conclusions:

• The novel semi-supervised MAR framework offers an efficient solution for metal artifact reduction in CT imaging.
• The approach successfully reduces model complexity while maintaining high performance.
• This method shows promise for improving the clinical utility of CT scans affected by metal artifacts.