Structure-aware diffusion for low-dose CT imaging.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computational Imaging

Background:

• Low-dose X-ray computed tomography (CT) imaging is crucial for reducing radiation exposure but results in noisy images with artifacts.
• Existing deep learning denoising models can cause structural degeneration and blurring.
• Current diffusion models for low-dose CT often start with generic noise, lacking specific data priors, leading to slow reconstruction and suboptimal quality.

Purpose of the Study:

• To develop an advanced framework for high-fidelity CT image reconstruction from low-dose data.
• To address limitations of existing diffusion models by incorporating structural information early in the generative process.
• To improve noise reduction, structure preservation, and generalization capabilities in low-dose CT imaging.

Main Methods:

• Introduced a Structure-Aware Diffusion (SAD) model, an end-to-end self-guided learning framework.
• Developed a nonlinear diffusion bridge to learn physical degradation priors directly from measurements.
• Integrated prompt learning and implicit neural representation using degraded inputs as structural prompts.
• Employed an efficient self-guided diffusion architecture with iterative prompt refinement.

Main Results:

• SAD demonstrated superior performance in noise removal and structure preservation compared to existing methods.
• The model achieved excellent blind-dose generalization capabilities on benchmark datasets.
• Reconstruction quality was significantly improved with very few generative steps, including single-step reconstruction.

Conclusions:

• The Structure-Aware Diffusion (SAD) model offers a significant advancement in low-dose CT image reconstruction.
• Incorporating structural priors through prompt learning effectively guides the diffusion process for high-fidelity results.
• SAD provides a promising solution for clinical applications requiring high-quality CT images with reduced radiation doses.