Training low dose CT denoising network without high quality reference data.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Supervised learning methods for low-dose CT (LDCT) denoising require paired normal-dose CT images, which are difficult to obtain clinically.
• Training deep learning models without clean reference images is a significant challenge in medical imaging.

Purpose of the Study:

• To propose a novel self-supervised denoising method for LDCT imaging that does not require paired clean images.
• To improve the feasibility and clinical applicability of LDCT denoising techniques.

Main Methods:

• A self-supervised learning framework is developed for LDCT denoising.
• Perceptual loss is utilized for data consistency in the feature domain.
• Attention blocks are incorporated into the decoding phase to enhance image quality.

Main Results:

• The proposed self-supervised framework demonstrates effectiveness in LDCT denoising.
• The method achieves competitive qualitative and quantitative performance compared to state-of-the-art supervised and unsupervised approaches.
• Experimental validation confirms the model's capabilities.

Conclusions:

• The developed self-supervised framework offers a flexible and practical solution for LDCT denoising.
• This method eliminates the need for collecting paired training data, making it suitable for real-world clinical scenarios.