DVG-Diffusion: Dual-View-Guided Diffusion Model for CT Reconstruction From X-Rays

Area of Science:

• Medical Imaging
• Computer Vision
• Artificial Intelligence

Background:

• Reconstructing 3D CT volumes from few-view 2D X-rays is challenging due to the projection nature of X-rays.
• Existing deep learning methods face difficulties in complex 2D X-ray to 3D CT mapping.

Purpose of the Study:

• To develop an effective deep learning network for high-fidelity 3D CT volume reconstruction from limited 2D X-ray projections.
• To reduce the learning difficulty in CT reconstruction by incorporating new view synthesis and feature alignment.

Main Methods:

• Proposed a dual-view guided diffusion model (DVG-Diffusion) integrating real and synthesized X-ray views.
• Utilized a view parameter-guided encoder for spatially aligned feature extraction.
• Employed a latent diffusion model conditioned on dual-view features for CT latent representation refinement.

Main Results:

• DVG-Diffusion achieved a balance between high fidelity and perceptual quality in CT reconstruction.
• Experimental results demonstrated superior performance compared to state-of-the-art methods.
• Comprehensive analysis of views and reconstruction quality was presented.

Conclusions:

• The proposed DVG-Diffusion model effectively reconstructs 3D CT volumes from few-view X-rays.
• The dual-view guidance and feature alignment strategies significantly improve reconstruction accuracy and quality.
• The method offers a promising solution for low-dose and limited-view CT imaging challenges.