Research On Denoising Algorithms Of Low-Dose CT Images Based On Deep Learning

Since the first CT machine was successfully developed,x-ray technology has become a means of medical clinical diagnosis.With the extensive use of x-ray technology,the inherent radiation hazards of rays have attracted extensive attention.Low-dose CT scanning technology is widely used in clinical medical diagnosis due to reduced radiation dose to the human body,but low-dose CT scanning will cause the image to have strip artifacts and noise,and the signal-to-noise ratio will be reduced,affecting the doctor’s accurate diagnosis of patients.Therefore,how to improve the quality of low-dose CT images has become a hot spot in the field of CT research.This article uses deep learning methods to improve the quality of low-dose CT images,the main contributions are:（1）A recursive residual encoder-decoder network model with a simple structure is constructed,and a recursive residual encoder-decoder network denoising algorithm for low-dose CT image denoising is proposed.This model builds a high-quality network through the recursion of shallow encoder-decoder networks,and has the advantages of residual encoder-decoder networks and recursive networks.The contrast experiments show that the algorithm can effectively remove artifacts and noise in low-dose CT images,and can restore image details and lesion characteristics well.（2）The intensity and scale of noise in different parts of the human body in low-dose CT images are different.To solve the problem of incomplete artifact removal by a single convolution operation,a multi-scale residual network model was constructed,and a low-dose CT image denoising algorithm based on multi-scale residual network architecture was proposed.The algorithm uses atrous convolutions with different dilated factors to extract artifacts at different scales.The comparison of experimental results shows that the algorithm is effective in removing noise and artifacts and maintaining detailed features.（3）In order to avoid the disappearance of gradients during the training of traditional generators,a wasserstein generative adversarial network model based on the total variation regularization term was constructed,and a low-dose CT image denoising algorithm based on the total variation regularization wasserstein generative adversarial network was proposed.The generator of this model adopts the residual encoder-decoder network structure.In addition,the perceptual loss,L₁loss and total variation regularization terms are introduced into the loss function,which cooperates with the wasserstein distance to help the network recover high-quality CT images.Compared with other network algorithms,this algorithm removes artifacts while preserving the image’s structural features and edge details.