Unified-distribution residual diffusion model for cross-scanner generalizable low-dose CT denoising

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Multi-center CT protocols require consistent image quality across scanners.
• Scanner-specific variations challenge cross-scanner low-dose CT (LDCT) imaging.
• Existing deep learning methods struggle with data distribution alignment for cross-scanner LDCT.

Purpose of the Study:

• To propose a single deep learning model for generalizable cross-scanner LDCT image denoising.
• To address the inadequate alignment of data distributions across diverse CT scanners.
• To enhance the performance of LDCT reconstruction in multi-center settings.

Main Methods:

• Developed a unified-distribution residual diffusion model (UDRDiff).
• Introduced a unified-distribution degeneration operator with an input-suppression term (IST) to mitigate inter-scanner discrepancies.
• Implemented a residual diffusion mechanism for a deterministic degradation process and accelerated sampling.

Main Results:

• UDRDiff demonstrated superior performance over existing methods across four scanners and two anatomical regions.
• Achieved improved PSNR and SSIM values on representative abdominal and head image datasets.
• Validated the method's effectiveness and generalization capability across different scanners and anatomical regions.

Conclusions:

• UDRDiff effectively aligns data distributions for cross-scanner LDCT denoising.
• The model enhances generalizability under multi-center, multi-scanning conditions.
• Provides a feasible solution for cross-scanner denoising of clinical LDCT images.