Dual-Domain Self-Supervised Deep Learning with Graph Convolution for Low-Dose Computed Tomography Reconstruction.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Radiology

Background:

• Low-dose CT (LDCT) reduces radiation exposure but suffers from low signal-to-noise ratio (SNR), impacting diagnostic quality.
• Existing deep learning denoising methods often require paired low-dose and normal-dose images, limiting clinical application.
• Current self-supervised methods make simplistic noise assumptions and focus on single domains (sinogram or image), reducing effectiveness.

Purpose of the Study:

• To develop an effective self-supervised deep learning framework for low-dose CT (LDCT) denoising and reconstruction.
• To address the limitations of existing supervised and self-supervised denoising techniques in CT imaging.
• To improve the diagnostic quality of LDCT images without requiring paired data.

Main Methods:

• Introduced the Dual-Domain Self-supervised (DDoS) framework for LDCT denoising and reconstruction.
• Developed sinogram-denoising and CT image-denoising networks tailored to specific noise characteristics.
• Employed a unified hybrid architecture combining graph convolution and multi-channel attention for feature extraction in both domains.

Main Results:

• The DDoS framework demonstrated superior performance in denoising and reconstruction compared to state-of-the-art methods.
• Experiments on large-scale LDCT datasets validated the effectiveness of the dual-domain approach.
• The method successfully enhanced SNR in LDCT images, meeting diagnostic quality standards.

Conclusions:

• The DDoS framework offers a robust and effective self-supervised solution for LDCT image enhancement.
• This approach overcomes the need for paired data, making it more clinically applicable.
• DDoS significantly improves the diagnostic utility of low-dose CT scans.