Research On Low-dose CT Image Denoising Algorithm Based On Deep Convolutional Neural Network

Low-dose computed tomography(CT)technology is widely used in the field of clinical medical diagnosis,because it can reduce the harm of radiation to the human body.However,the reduction of the radiation dose will cause a serious decline in the quality of the reconstructed image,which may lead to misdiagnosis.The existing low-dose CT(LDCT)image processing method is easy to cause problems such as edge blurring and loss of detail of the CT image,and it is difficult to find a balance between removing noise and retaining image detail information.This topic applies the powerful feature extraction ability of deep convolutional neural network(CNN)to low-dose CT image denoising,and transforms the low-dose CT image denoising process into a nonlinear complex problem fitting process.The research is carried out from two angles of image domain and stationary wavelet domain,aiming to improve the quality of lowdose CT images.The main research contents of this paper are as follows:(1)Image domain LDCT image denoising based on deep residual convolutional neural network.Using low-dose CT images as inputs and noise information images as labels,the complex relationship between them is fitted through network training.After the network training is completed,the noise information can be obtained from the low-dose CT image,and the noise-reduced image can be obtained by subtracting the noise information from the lowdose CT image.The network design uses small convolution kernels,joins the bypass connection module and adopts the residual learning mechanism,which can improve the convergence speed of the network and preserve the image details.Through experimental comparison,the model can not only remove noise and artifacts,but also better preserve bone edge information.(2)Stationary wavelet domain LDCT image denoising based on deep residual convolutional neural network.This model is different from the image domain CNN with the images are directly used as the data sets,but the high-frequency coefficients images of the images after stationary wavelet decomposition are used as the data sets.The model uses a deep CNN to end-to-end modeling between high-frequency coefficients of LDCT images and noise information in high-frequency coefficients,and fully combine the advantages of stationary wavelet transform in extracting high-frequency information from images and the advantages of the CNN's powerful feature extraction capabilities.Since the model denoises from high frequency coefficients in multiple directions,it is more conducive to suppressing image noise and preserving image details.Through experimental comparison,the model not only excels in removing noise and retaining bone detail information,but also shows greater advantages in muscle tissue detail retention.