Noise-inspired diffusion model for generalizable low-dose CT reconstruction.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computational Imaging

Background:

• Deep learning models for low-dose CT (LDCT) struggle to generalize to unseen radiation doses.
• Existing methods often require extensive data or fine-tuning for improved generalization.
• Diffusion models show promise but can generate artifacts due to noise and imprecise guidance.

Purpose of the Study:

• To develop a generalizable deep learning model for LDCT reconstruction that adapts to various unseen radiation doses.
• To address the limitations of current diffusion models in handling CT image noise and prior information.
• To improve the fidelity and robustness of LDCT reconstruction across different dose levels.

Main Methods:

• Proposed a noise-inspired diffusion model (NEED) for generalizable LDCT reconstruction.
• Introduced a shifted Poisson diffusion model to denoise projection data, aligning with LDCT noise characteristics.
• Developed a doubly guided diffusion model for image refinement, using LDCT images and initial reconstructions for enhanced fidelity.

Main Results:

• NEED demonstrated superior reconstruction and generalization performance compared to state-of-the-art methods on two datasets.
• The model effectively reconstructs images at unseen dose levels using a time step matching strategy.
• Qualitative, quantitative, and segmentation-based evaluations confirmed NEED's effectiveness.

Conclusions:

• The proposed NEED model significantly improves generalizable LDCT reconstruction by tailoring diffusion processes to noise characteristics.
• NEED offers a robust solution for reconstructing high-fidelity CT images from low-dose data across various radiation levels.
• The dual-domain approach and noise-inspired design overcome key limitations in current LDCT reconstruction techniques.