BSN With Explicit Noise-Aware Constraint for Self-Supervised Low-Dose CT Denoising.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Image Processing

Background:

• Supervised deep learning for low-dose computed tomography (LDCT) image denoising requires paired low-dose and normal-dose images, which are often unavailable.
• Existing self-supervised deep learning (SSDL) methods fail with spatially correlated noise due to independence assumptions.

Purpose of the Study:

• To develop a novel SSDL approach for high-quality LDCT imaging that does not rely on paired training data.
• To overcome the limitations of existing SSDL methods in handling spatially correlated noise in LDCT images.

Main Methods:

• Introduced the Noise-Aware Blind Spot Network (NA-BSN), a novel SSDL approach for LDCT denoising.
• Developed an explicit noise-aware constraint mechanism based on the observed descending trend of the l1 norm in downsampled CT images with increasing radiation dose.
• Validated the method using various clinical datasets, including different scanning machines, positions, dose levels, and reconstruction kernels.

Main Results:

• NA-BSN achieves high-quality image reconstruction without referencing clean data.
• The method effectively reduces spatially correlated CT noises.
• Crucial image details are retained in complex scenarios, demonstrating robustness across diverse clinical settings.

Conclusions:

• NA-BSN offers a promising solution for self-supervised LDCT image denoising, mitigating the need for paired data.
• The noise-aware constraint mechanism effectively addresses spatially correlated noise challenges in LDCT imaging.
• The proposed method demonstrates significant potential for clinical application in improving LDCT image quality.