DiFiR-CT

DiFiR-CT reconstructs time-resolved CT images that suppress motion artifacts by representing object boundaries with neural implicit signed distance functions and an analysis-by-synthesis approach.

Key Features:

• Motion artifact suppression: Reduces motion artifacts in CT reconstructions arising from non-rigid, temporally varying, and patient-specific motion fields.
• Analysis-by-synthesis reconstruction: Generates time-resolved, artifact-reduced images using an analysis-by-synthesis framework without explicit motion estimation.
• Neural implicit SDF representation: Models object boundaries with a signed distance function (SDF) parameterized by neural networks (neural implicit representation).
• Boundary-driven motion representation: Leverages object boundaries as the primary representation to capture temporally-evolving scene geometry and motion.
• Spatial and temporal smoothness constraints: Optimizes the SDF under spatial and temporal smoothness constraints to stabilize reconstructions.
• Sinogram-based operation and noise robustness: Operates directly on sinogram data and demonstrates robustness across a wide range of sinogram noise levels.
• Multi-intensity scene support: Extends to multi-intensity scenes and relies on realistic initial segmentation for effective initialization.
• Validated scenarios and comparative performance: Validated on simulated cases (small circle translation, heart-like ellipse diameter changes, and complex topological deformations) and shown to outperform filtered backprojection reconstructions without hyperparameter tuning or architectural changes.

Scientific Applications:

• Cardiac and dynamic CT imaging: Application to cardiac CT and other dynamic imaging scenarios with non-rigid organ motion.
• Motion artifact mitigation in clinical CT: Improving time-resolved CT reconstructions where motion artifacts compromise diagnostic image quality.
• Computational imaging research: Studying temporally-evolving scenes, multi-intensity phantoms, and reconstruction methods using sinogram data and neural implicit representations.

Methodology:

Analysis-by-synthesis reconstruction using a neural-network-parameterized signed distance function (SDF) to represent object boundaries, optimized directly from sinogram data under spatial and temporal smoothness constraints, without explicit motion estimation.