Research On Low-dose CT Image Denoising Method

In recent years,computed tomography(CT)technology has been widely used in clinical diagnosis.Because CT generates high radiation and endangers human health,the medical community usually uses low-dose CT scanning technology(LDCT)to obtain CT images.Although LDCT can effectively reduce the amount of radiation,LDCT images are prone to introduce more noise and artifacts,which may eventually affect the diagnosis of radiologists.Since the original CT data is difficult to obtain,a lot of research focuses on directly performing post-processing on CT images to denoise.Existing CT image denoising algorithms are mainly for CT images with specific doses,and there are problems such as loss of image details,image blur,and excessive smoothness after denoising.In order to solve the above difficulties,this paper mainly proposes three algorithms based on deep learning to denoise low-dose CT images.The main tasks are as follows:1.Propose a two-stage CT blind denoising network based on noise estimation.In the first stage,characterization learning is used to train the noise estimation model,and the potential noise features of the image are extracted.In the second stage,the feature fusion network firstly fuses the noise features with the original image,and then uses the residual codec convolutional neural network to achieve image denoising.Experimental results on Piglet's public data set show that the introduction of prior knowledge of noise features can greatly improve the performance of the denoising network.It can achieve better results in the task of denoising CT images with unknown doses.2.Proposed a residual high-resolution codec denoising network.The single-branch serial feature extraction method of the codec network is improved to multi-branch parallel multi-scale feature extraction.Through cross-resolution feature fusion and the use of Concatenate as the fusion method of the residual structure between the codec networks,the network's performance is greatly improved.Denoising performance.The experimental results on Piglet's public data set show that.Compared with some mainstream denoising algorithms,the residual high-resolution codec network proposed in this paper has obtained a great improvement in performance.3.Propose a point-to-point network Multi-Pix based on attention mechanism and pyramid pooling.On the basis of the pix2 pix network,the attention mechanism and the spatial pyramid network are introduced to enhance the relationship between the space and the channel,so as to improve the detailed information of the LDCT image.Experimental results on the Piglet public data set show that combining the attention mechanism and pyramid pooling can improve the performance of the original denoising network.MultiPix has achieved good denoising effects on CT images of different doses.