Low-dose CT denoising using a diffusion prior via score distillation sampling

Low-dose computed tomography (LDCT) reduces radiation exposure but amplifies noise, which can degrade diagnostic reliability. Recent studies have applied diffusion-based models to LDCT denoising and shown promising performance. However, many of these approaches require paired LDCT and normal-dose CT (NDCT) data or dose-specific retraining, and their reconstruction procedures are often strictly coupled to a predefined diffusion trajectory.

This study proposes an optimization-based LDCT denoising framework that uses a pre-trained NDCT diffusion model as an image prior and can be applied across various dose levels without dose-specific retraining.

We formulate LDCT denoising as an optimization problem guided by a pre-trained score-based generative model (SGM) trained only on NDCT images. The Poisson data fidelity term enforces consistency with the measured projection data, whereas the score distillation sampling (SDS) term acts as a regularizer that guides the reconstruction toward the NDCT distribution learned by the pre-trained SGM. The denoised image is obtained by iteratively optimizing a projection-domain residual variable using the combined Poisson data fidelity gradient and SDS-based regularization gradient.

The proposed method effectively suppressed noise while preserving fine anatomical structures across multiple dose levels using a single pre-trained diffusion model. Quantitative evaluations with respect to full-dose reference images and visual comparisons demonstrated competitive performance compared with supervised and self-supervised LDCT denoising baselines.

The results show that diffusion-based image priors can be leveraged across various low-dose levels within the proposed optimization-based denoising framework. Our work suggests that SDS-guided diffusion priors are a promising direction for LDCT denoising in data-limited settings.