Ray-Bundle-Based X-Ray Representation and Reconstruction: An Alternative to Classic Tomography on Voxelized Volumes

Area of Science:

• Medical Imaging
• Computational Physics
• Applied Mathematics

Background:

• Classic computed tomography (CT) discretizes volumes into Cartesian grids, causing inconsistencies and resolution discrepancies.
• Existing methods often require interpolation, complex calculations, or matrix inversions.

Purpose of the Study:

• To develop a continuous-volume modeling approach for X-ray tomography that overcomes limitations of traditional grid-based methods.
• To introduce a novel system, T-ReX (Transmission Ray bundles for X-ray geometry), for improved tomographic reconstruction.

Main Methods:

• Modeled forward projection as a continuous-to-discrete mapping using 'ray bundles' representing detector elements.
• Utilized implicit neural representations to map spatial coordinates to attenuation properties.
• Implemented a novel adaptive sampling strategy along ray bundles to focus on high-gradient regions.

Main Results:

• T-ReX effectively reconstructs internal volumes by modeling continuous space and ray bundles.
• The implicit neural representation approach eliminates the need for explicit interpolation or matrix inversions.
• Experiments demonstrated T-ReX's effectiveness in sparse-view X-ray tomography across various scenarios.

Conclusions:

• T-ReX offers a physics-informed approach to X-ray tomography, enhancing accuracy and efficiency.
• The continuous modeling and adaptive sampling strategies are key to T-ReX's performance.
• This method shows significant promise for sparse-view and challenging tomographic reconstruction problems.