SADiff: A Sinogram-Aware Diffusion Model for Low-Dose CT Image Denoising.

Area of Science:

• Medical Imaging
• Machine Learning
• Image Processing

Background:

• Computed Tomography (CT) image denoising is vital for medical imaging quality.
• Diffusion models show promise for high-quality CT image generation.
• Existing diffusion models lack CT-specific priors and robust generalization.

Purpose of the Study:

• To enhance CT image denoising using diffusion models.
• To address limitations in adaptability and generalization of current methods.
• To improve the diagnostic value of CT scans.

Main Methods:

• Developed a novel conditioning module integrating sinogram-domain image formation priors.
• Implemented a two-phase training strategy for progressive learning of anatomical structures.
• Utilized a diffusion model architecture tailored for CT denoising.

Main Results:

• Achieved significant improvements in CT image quality.
• Demonstrated up to 17% increase in Peak Signal-to-Noise Ratio (PSNR).
• Showed up to 38% increase in Structural Similarity Index Measure (SSIM), outperforming state-of-the-art methods.

Conclusions:

• The proposed conditioning module effectively incorporates CT imaging priors.
• The two-phase training enhances model adaptability and generalization.
• The novel diffusion model significantly advances CT image denoising performance.