A Low-Dose CT Image Denoising Method Combining Multistage Network and Edge Protection

Abstract

Low-dose CT is an effective method to reduce the amount of radiation and the potential impact of radiation on patients. However, the reduction of radiation dose will lead to a large amount of noise in the reconstructed image, which will blur the edge details of the internal tissues and the organizational structure, thus causing confusion to the quantitative judgment of doctors in the diagnosis process. For this reason, a multi-stage denoising convolutional neural network emphasizing edge protection (MEP-Net) is proposed in this paper. Firstly, to overcome the problem of fixed input feature scale in the single-stage network, a multi-stage network structure is constructed which can extract multi-scale image noise features. Secondly, in order to protect the edge information of the image, the edge protection module is designed, which is combined with the convolutional neural network to extract the edge features of different stages in the multichannel. Finally, the cross-stage feature fusion mechanism is adopted to restore the edge details of the noisy image. In addition, this paper introduces a compound loss function consisting of Charbonnier loss and edge loss suitable for multi-stage network structure as the loss function to guide the multistage network for training. Experiments are carried out on AAPM2016 public dataset to verify that the edge protection module can effectively protect image edge details. Compared with other advanced algorithms, the proposed multilevel denoising algorithm is optimal in both subjective and objective evaluation indexes.

Keywords

• edge protection
• image processing
• low-dose CT denoising
• multistage