Unsupervised High-Order Implicit Neural Representation With Line Attention for Metal Artifact Reduction

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Image Processing

Background:

• Metallic implants in CT scans cause artifacts, hindering diagnosis.
• Current deep learning methods for MAR require extensive training data.
• Implicit Neural Representation (INR) offers unsupervised image restoration but struggles with spatial correlations.

Purpose of the Study:

• To develop an unsupervised metal artifact reduction (MAR) framework using INR.
• To enhance MAR by capturing local contextual and geometric information from X-rays.
• To reconstruct artifact-free CT images without reference data.

Main Methods:

• Proposed an INR-based unsupervised MAR framework.
• Designed a High-order Line Attention Network to process spatial coordinates and X-ray data.
• Employed a multiple local adjacent ray sampling strategy for richer contextual information.

Main Results:

• The High-order Line Attention Network effectively captures local context and geometric features.
• Second-order feature interaction addresses spectral bias and improves signal detail fitting.
• The framework successfully approximates the implicit continuous function for artifact-free CT reconstruction.

Conclusions:

• The proposed unsupervised MAR framework achieves superior performance compared to state-of-the-art methods.
• The approach effectively reduces metal artifacts in CT images.
• This method holds promise for improving diagnostic accuracy in medical imaging.