Dose-aware denoising diffusion model for low-dose CT.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computational Science

Background:

• Low-dose computed tomography (LDCT) is crucial for reducing patient radiation exposure.
• Deep learning (DL) methods show promise for LDCT denoising but struggle with generalizability and uncertainty.
• Existing methods face challenges in adapting to diverse datasets and dose levels.

Purpose of the Study:

• To introduce a novel dose-aware diffusion model for effective LDCT denoising.
• To address the limitations of current DL methods in generalizability and uncertainty.
• To maintain structural fidelity and improve noise reduction in LDCT images.

Main Methods:

• A physics-based forward process with continuous timesteps for flexible noise representation.
• Incorporation of an efficient noise calibration module to align intermediate results with timesteps.
• A novel method for estimating timesteps for unseen LDCT images to ensure generalization.

Main Results:

• The proposed model outperforms existing methods in preserving noise texture and anatomical structures.
• Consistent performance was observed across different dose levels and on an unseen dataset.
• Qualitative and quantitative evaluations on Mayo Clinic datasets confirm effectiveness.

Conclusions:

• The novel dose-aware diffusion model effectively addresses generalization and uncertainty in LDCT denoising.
• The method demonstrates high structural fidelity and computational efficiency.
• This approach offers a clinically practical solution for LDCT denoising across various dose levels.