EPC-3D-Diff: Physics-Consistent Diffusion for CBCT-to-CT Synthesis
A new framework called EPC-3D-Diff, which utilizes a conditional 3D latent diffusion approach, has been introduced for synthesizing volumetric CBCT-to-CT images. This method incorporates a projection domain equivariance loss based on acquisition physics, recognizing that rotating an in-plane volume results in an angular shift in its projections. During the training phase, the synthesized CT volumes are rotated, projected forward, and aligned with the angle-shifted projections of the corresponding target CT, thus applying a physics-consistent equivariance constraint. By operating within a compact latent space, conditional diffusion effectively captures 3D context. This technique enhances the quantitative reliability of CBCT by mitigating issues related to scatter, noise, and artifacts, leading to improved accuracy in Hounsfield Units for radiotherapy.
Key facts
- EPC-3D-Diff is a conditional 3D latent diffusion framework for CBCT-to-CT synthesis.
- It introduces a projection domain equivariance loss derived from acquisition physics.
- An in-plane rotation of the volume corresponds to an angular shift in its projections.
- Training enforces equivariance by forward projecting rotated synthesized CT volumes and matching to angle-shifted projections of paired target CT.
- The constraint is integrated into the diffusion objective.
- Conditional diffusion is performed in a compact latent space.
- CBCT suffers from scatter, noise, and reconstruction artifacts, limiting HU accuracy.
- The method aims to improve quantitative reliability for radiotherapy.
Entities
—