Diffusion Meets Sinogram: A Hybrid Learning Framework for Low-Dose CT Image Denoising with Structural and Textural...

CT image denoising is essential for improving the diagnostic quality of low-dose CT (LDCT) scans. While diffusion models have demonstrated remarkable success in image generation, existing diffusion-based denoising methods struggle due to their lack of CT image formation priors and their inability to generalize across varying anatomical structures. To address these issues, we propose SADiff, a novel Sinogram-Aware Diffusion model that integrates sinogram priors with diffusion-based denoising. Our framework consists of a two-stage process: (1) a Degradation Removal (DR) network that effectively suppresses noise and artifacts, and (2) a Stable Diffusion Network enhanced with a CT-conditional (CTC) module, which incorporates sinogram priors for improved feature generation. Additionally, a CT Prompt (CTP) module dynamically generates CT-specific prompts to guide the diffusion process. Extensive experiments on multiple CT datasets demonstrate that SADiff outperforms existing methods, achieving up to 17% PSNR and 38% SSIM improvements, ensuring high-quality, realistic CT image restoration.