DNS-Diff:用于低剂量CT图像重建的定向噪声抑制扩散模型OA

Low dose CT images are prone to increased noise and artifacts due to insufficient photons and electronic noise.Re-cently,some studies have attempted to use diffusion models to address the issues of excessive smoothing and unstable training en-countered in previous deep learning-based denoising models.However,diffusion models need to cover complex noise distribu-tions and learn fine-grained reconstruction,thus requiring a large number of training epochs.This article proposes a diffusion model based on feature extraction for directed noise suppression in Low-Dose CT(LDCT)denoising,called DNS-Diff.Firstly,DNS-Diff utilizes LDCT images to extract features such as background,details,and edges,and employs gating method to sup-press noise and content related to the original image features,significantly reducing the number of training rounds.This is due to the rich information of LDCT images used as the starting point of the sampling process.At the same time,in order to reduce the artifacts introduced during the denoising process,obtain good visualization results,and use computing resources as little as pos-sible,this paper designs a post-processing module based on traditional methods,such as bilateral filtering and Gaussian filter-ing,etc.,which can adjust the background,optimize details,and remove artifacts to distinguish core differences faster and better.