Self-supervised dual-domain balanced dropblock-network for low-dose CT denoising.

Area of Science:

• Medical Imaging
• Computer Vision
• Machine Learning

Background:

• Self-supervised learning (SSL) is effective for low-dose computed tomography (LDCT) denoising, but conventional methods neglect sinogram domain information.
• Existing dual-domain SSL methods struggle with blurring artifacts due to noise inhomogeneity in LDCT sinograms.

Purpose of the Study:

• To propose SDBDNet, an end-to-end dual-domain SSL method for LDCT denoising that mitigates blurring artifacts.
• To leverage the properties of inhomogeneous noise in sinograms and moderate sinogram-domain denoising for improved image quality.

Main Methods:

• SDBDNet utilizes a dual-domain approach, processing data in both sinogram and image domains.
• Sinograms are split into subsets to create paired training data with independent noise, then restored using interpolation and learning-based correction.
• Adaptive sinogram denoising is achieved via Dropblock regularization and a weighted average of denoised and noisy sinograms.

Main Results:

• SDBDNet demonstrates effective denoising in both domains without introducing blurring artifacts.
• Numerical experiments confirm that SDBDNet outperforms popular non-learning and existing self-supervised learning methods.

Conclusions:

• SDBDNet offers a novel, high-performance dual-domain SSL solution for LDCT denoising.
• The study highlights the critical role of appropriate sinogram-domain denoising in dual-domain approaches, suggesting avenues for future research.