A Research On Medical Image Denoising Based On Deep Convolutional Neural Network

Since the advent of computed tomography(CT)equipment,it has been diffusely applied in the field of clinical diagnosis,because its advantages of fast scanning and low cost are rather obvious.With the enhancement of people’s health awareness,they strongly recommend to reduce the harm of the radiation from CT equipment.However,as the radiation dose is reduced,the image quality will be deteriorated seriously,which will lead to the serious consequences of misdiagnosis.The traditional image denoising methods for low-dose CT image usually lead to the problems such as the blurring of the image edges and the loss of image details.It is so difficult to maintain a balance between noise removal and detail information retention.Deep convolutional neural network has extremely powerful feature extraction capabilities,and has been successfully applied to computer vision such as object recognition and image classification.Whereupon it has become a hot research topic in academia.Therefore,this thesis will apply the deep convolutional neural network to noise reduction of low-dose CT image,which regards the denoising process as nonlinear fitting process of a complex problem,aiming to improve the quality of low-dose CT image.The main research contents in this thesis are present as follows:(1)A method of medical image denoising based on dual residual convolutional neural network is proposed.The network structure is mainly composed of the structure similar to Inception and dual residual module.With the low-dose CT image as the input and the normal-dose CT image as the label,the complex relationship between them are learned after recurrent training of the proposed network structure.When neural network training is completed,a denoising model can be obtained to predict the denoising image of low-dose CT image.In the network structure design,firstly,the input image’s feature information is extracted by using the structure similar to Inception with three convolutional kernels of different sizes,and the different feature information is fused and output to the dual residual module.Then,the dual residual module is used to further extract the image’s feature information.Last,changing the number of image channels by a group of convolutional kernels of 3 ?3 is to output the prediction image.Compared with traditional methods,this method improves PSNR by at least 1.03 d B and SSIM by at least 0.0182 on the entire image.(2)A method of medical image denoising based on dual attention convolutional neural network is proposed.The previous denoising models based on neural network usually don’t pay attention to the different feature of the image in the two dimensions of channel and space and directly carry out convolution operation on the image.Therefore,a network structure based on dual attention is designed.In the network structure design,firstly,the image feature information is extracted by the structure similar to Inception.Then,the image feature information is further extracted by dual attention modules(spatial attention module and channel attention module),in which the spatial attention module selectively aggregates the feature information of each position,and the channel attention module selectively emphasizes the feature information of each channel.And then the outputs of these two attention modules are added together to improve feature information.Finally,a group of convolutional kernels of 3 ?3 is used to change the number of image channels to output the prediction image.Compared with the traditional method,in terms of the entire image,the method improves at least 1.38 d B on PSNR and 0.0069 on SSIM.In this thesis,experiments are carried out on the simulation data set respectively.It is found that the two methods proposed in this paper not only perform well in noise suppression,but also show great advantages in detail retention through the experimental comparison.What’s more,the methods presented in this thesis have good practicability and can improve the viewing quality of images to a great extent,while reducing the effects of radiation,which are of great significance for the clinical application of medical images and subsequent medical diagnosis.