Diffusion probabilistic priors for zero-shot low-dose CT image denoising.

Area of Science:

• Medical Image Computing
• Deep Learning
• Image Denoising
• Diffusion Models

Background:

• Denoising low-dose computed tomography (CT) images is crucial in medical imaging.
• Supervised deep learning methods require difficult-to-obtain paired low-dose and normal-dose CT images.
• Existing unsupervised methods often demand large datasets or specialized acquisition protocols.

Purpose of the Study:

• To develop a novel unsupervised method for low-dose CT image denoising.
• To enable zero-shot denoising by training solely on normal-dose CT images.

Main Methods:

• Leveraged a cascaded unconditional diffusion model to generate high-quality normal-dose CT images.
• Integrated low-dose CT images into the diffusion model's reverse process as likelihood.
• Employed iterative maximum a posteriori (MAP) estimations with adaptive coefficient adjustments for noise level adaptation.

Main Results:

• The proposed unsupervised method outperformed state-of-the-art unsupervised and supervised deep learning approaches.
• Achieved high peak signal-to-noise ratio (PSNR) values: 45.02 dB on abdomen CT and 35.35 dB on chest CT.
• Significantly surpassed the Noise2Sim unsupervised algorithm by 0.39 dB (abdomen) and 0.85 dB (chest).

Conclusions:

• A novel unsupervised low-dose CT denoising method based on diffusion models was successfully developed.
• The method effectively addresses data scarcity by training only on normal-dose CT images.
• The approach delivers excellent qualitative and quantitative denoising performance.