Unsupervised learning-based dual-domain method for low-dose CT denoising.

Area of Science:

• Medical Imaging
• Computational Imaging
• Artificial Intelligence in Healthcare

Background:

• Low-dose computed tomography (LDCT) is crucial for reducing radiation exposure in medical diagnostics.
• Deep learning methods have advanced LDCT imaging but face challenges with paired low- and high-dose datasets.
• Supervised learning requires geometrically matched datasets, a significant limitation in LDCT denoising.

Purpose of the Study:

• To develop an unsupervised learning-based method for low-dose CT (LDCT) imaging.
• To address the challenge of acquiring geometrically paired datasets in supervised LDCT denoising.
• To improve image quality in LDCT while minimizing radiation dose.

Main Methods:

• A dual-domain unsupervised learning approach for LDCT denoising.
• Stage 1: Projection domain denoising using the Noise2Self method for statistically independent noise.
• Stage 2: Iterative enhancement combining generative model priors with iterative reconstruction.

Main Results:

• The proposed unsupervised method demonstrates superior denoising performance compared to existing methods.
• The method achieved the highest Structural Similarity Index Measure (SSIM) for denoised images.
• Experimental results validate the effectiveness of the dual-domain approach.

Conclusions:

• The unsupervised learning-based method is a viable alternative to supervised techniques for LDCT imaging.
• This approach is particularly beneficial when labeled datasets for LDCT are scarce.
• The method offers a promising direction for advancing safe and effective CT imaging practices.