X-ray CT metal artifact reduction using neural attenuation field prior

Area of Science:

• Medical Imaging
• Computational Imaging

Background:

• Metal artifacts in CT imaging degrade image quality and hinder diagnosis.
• Existing deep learning methods for metal artifact reduction (MAR) often fail on new data and require extensive training datasets.

Purpose of the Study:

• Propose a novel sinogram inpainting method for MAR using a neural attenuation field (NAF) as a prior.
• Develop NAFMAR, a self-supervised approach that optimizes a neural field, eliminating the need for large training datasets.

Main Methods:

• Optimize NAF to generate prior images for inpainting metal traces in sinograms.
• Utilize 3D forward projection to identify metal traces and employ a metal trace-masked ray sampling strategy for network supervision.
• Implement a metal-aware loss function to prioritize metal-associated regions, improving anatomical feature representation.

Main Results:

• NAFMAR provides an accurate prior for MAR without extensive datasets.
• Corrected images exhibit sharp features and preserved anatomical structures.
• NAFMAR outperformed baseline methods in SSIM and showed comparable PSNR to dual-domain CNNs in dental and pelvic CT imaging.

Conclusions:

• NAFMAR offers an effective, high-fidelity solution for metal artifact reduction in 3D tomographic imaging.
• The method eliminates the requirement for large datasets, making it broadly applicable.