Weakly-supervised progressive denoising with unpaired CT images

Area of Science:

• Medical Imaging
• Radiology
• Computational Imaging

Background:

• Low-dose computed tomography (CT) reduces radiation exposure but introduces significant noise, impacting diagnostic accuracy.
• Current fully-supervised CT denoising methods require large paired datasets of normal-dose and low-dose images, which are often impractical in clinical settings.

Purpose of the Study:

• To develop a weakly-supervised CT denoising framework that overcomes the limitations of paired data requirements.
• To enhance the diagnostic quality of low-dose CT images through advanced denoising techniques.
• To improve the flexibility of data acquisition for CT denoising research and application.

Main Methods:

• A physics-based noise model is employed to generate synthetic paired low-dose and normal-dose CT images from unpaired clinical data.
• A progressive denoising module is introduced to gradually reduce noise, avoiding direct mapping from low-dose to normal-dose images.
• Quantitative evaluation metrics including noise power spectrum and signal detection accuracy are utilized to assess image quality.

Main Results:

• The proposed weakly-supervised method achieves remarkable CT denoising performance.
• The framework demonstrates superior signal detectability compared to traditional fully-supervised denoising approaches.
• Experimental results validate the effectiveness of the progressive denoising module in noise compensation.

Conclusions:

• The developed weakly-supervised framework offers a flexible and effective solution for low-dose CT denoising, utilizing readily available unpaired data.
• This approach significantly enhances diagnostic image quality and signal detectability in low-dose CT imaging.
• The method holds promise for broader clinical adoption by alleviating data acquisition constraints.