DiFiR-CT: Distance field representation to resolve motion artifacts in computed tomography

Area of Science:

• Medical Imaging
• Computational Imaging
• Artificial Intelligence in Healthcare

Background:

• Motion during computed tomography (CT) data acquisition causes artifacts, hindering image quality.
• Existing methods to reduce motion artifacts in CT have limited success due to physical constraints and challenges in estimating complex motion fields.

Purpose of the Study:

• To develop a novel CT reconstruction method for generating time-resolved, artifact-free images.
• To achieve this without explicit motion estimation or modeling.

Main Methods:

• An analysis-by-synthesis approach was developed, focusing on object boundaries represented by neural network-modeled signed distance functions (SDFs).
• Optimization was performed under spatial and temporal smoothness constraints, enabling reconstruction without explicit motion estimation.

Main Results:

• The DiFiR-CT method successfully reconstructed images with complex motions (translation, deformation) in various scenarios.
• Reconstructions were high-quality, robust to noise, and did not require hyperparameter tuning or architectural changes.
• The approach demonstrated utility in multi-intensity scenes and highlighted the importance of realistic initial segmentation.

Conclusions:

• Projection data can be used to accurately estimate temporally-evolving scenes.
• A neural implicit representation and analysis-by-synthesis approach can achieve this without explicit motion estimation.