Zero-Shot Low-Dose CT Image Denoising via Patch-Based Content-Guided Diffusion Models

Low-dose computed tomography (CT) techniques reduce the radiation dose while increasing noise and artifacts. Significant advancements have been made in supervised-based methods, with several studies addressing issues such as excessive smoothing and information loss through diffusion models. However, the training process of supervised methods relies heavily on paired normal-dose and low-dose images. Paired CT images are difficult to obtain in practical clinical settings due to ethical constraints. Therefore, we propose a patch-based content-guided diffusion (PCDiff) model for zero-shot low-dose CT denoising to reduce data limitations. The framework is end-to-end; the training process utilizes only normal-dose CT images; original image details are preserved through a patch-based method; and the sampling process effectively denoises low-dose CT images via a content control strategy. Additionally, we introduce two innovative components to enhance detail preservation and the denoising effect: the diffusion refinement block (DRBlock) and the space and frequency interaction downsampling (SFID) block. The DRBlock utilizes the concept of diffusion to iteratively process the feature map channels through a “split-interaction-aggregation” mechanism. The SFID block integrates the spatial and frequency domains to enhance detail expression. Quantitative analysis of phantom and clinical data indicates that PCDiff outperforms popular unsupervised methods and several classical supervised CT denoising methods. Results from slicing demonstrate effective noise suppression and detail preservation.