A diffusion-based generative prior approach to sparse-view computed tomography

Area of Science:

• Medical Imaging
• Computational Imaging
• Artificial Intelligence in Radiology

Background:

• X-ray Computed Tomography (CT) image reconstruction from sparse or limited-angle data presents significant challenges.
• Insufficient data leads to artifacts and distortions in reconstructed CT images.
• Deep generative models offer potential solutions for improving CT reconstruction quality.

Purpose of the Study:

• To investigate the application of the Deep Generative Prior (DGP) framework for CT image reconstruction using sparse data.
• To enhance CT image reconstruction by combining diffusion-based generative models with iterative optimization.
• To explore modifications within the DGP framework to improve image generation and iterative algorithms for better reconstruction.

Main Methods:

• Utilized a Deep Generative Prior (DGP) framework integrating diffusion generative models.
• Employed an iterative optimization algorithm for CT image reconstruction from sparse sinogram data.
• Proposed modifications to the generative model and iterative algorithm within the DGP framework.

Main Results:

• Achieved promising results in CT image reconstruction even with highly sparse geometries.
• Demonstrated the potential of the DGP framework for overcoming data limitations in CT imaging.
• Highlighted the effectiveness of combining model-based explainability with neural network generative power.

Conclusions:

• The DGP framework shows significant potential for improving CT image reconstruction from sparse data.
• Modifications to the DGP framework can further enhance reconstruction quality.
• Further research is warranted to fully realize the capabilities of this approach in medical imaging.