SMART: Self-Supervised Learning for Metal Artifact Reduction in Computed Tomography Using Range Null Space...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computational Science

Background:

• Metal artifacts in CT scans degrade image quality and hinder diagnosis.
• Current deep learning MAR methods struggle with anatomical preservation, artifact priors, and spectral variations.

Purpose of the Study:

• To develop a self-supervised MAR method (SMART) addressing limitations of existing techniques.
• To improve diagnostic accuracy in CT imaging by reducing metal artifacts.

Main Methods:

• Proposed SMART, a self-supervised MAR method using range-null space decomposition (RND) and implicit neural representation (INR).
• RND separates metal and tissue linear attenuation coefficients (LACs) for artifact modeling and anatomical preservation.
• INR learns clinical characteristics self-supervisedly and incorporates polychromatic spectra for spectral adaptation.

Main Results:

• SMART demonstrated strong potential in reducing metal artifacts across synthetic and clinical datasets.
• The method effectively balanced artifact reduction with anatomical structure preservation.
• SMART showed superior generalizability to out-of-distribution clinical data due to spectral adaptation.

Conclusions:

• SMART offers a promising self-supervised approach for metal artifact reduction in CT imaging.
• The method's ability to adapt to spectral variations enhances its clinical applicability and generalizability.