RDM2: a two-stage model based on residual learning diffusion model and multi-scale convolution for Low Dose CT denoising

Abstract

Computed Tomography (CT) is widely used in clinical diagnosis, but large amount of radiation accompanied is not expected. Low Dose CT (LDCT) can reduce the radiation effect, however, noise and artifacts will be unavoidably produced. Low dose accompanies large noise intensity, which is difficult to effectively denoise while retaining the details. Aiming at this problem, a two-stage LDCT denoising model, named RDM2, is proposed. In the first stage, a residual learning diffusion model is constructed to eliminate the noise of LDCT. The residuals between LDCT and Normal Dose CT (NDCT) is a kind of complex mixed noise with unknown intensity. In order to fully utilize the residual information, the whole residual is equally divided into small pieces and added iteratively in the diffusion process. Considering even the best trained residual diffusion model may bring unavoidable error when it is used for prediction, a multi-scale convolution encoder decoder convolution neural network (MEDCNN) is proposed in the second stage to further reduce this part of error. The proposed model RDM2 is validated on both the Mayo2020 25% dose LDCT dataset and Mayo2020 10% dose LDCT dataset, the values of PSNR, SSIM, and RMSE on these two datasets are respectively 44.7651, 0.9939, 0.0068 and 35.5302, 0.9601, 0.0172. It is proved that RDM2 outperforms the traditional method, the supervised learning-based method and the GAN-based method, and has the potential to meet clinical needs. Code is available at: https://github.com/zhencunjiang/RDM2.