Research On Medical Low-dose DR Image Denoising Based On Convolutional Neural Network

X-Ray digital radiography（DR）has been widely used in the detection of clinical diseases.In order to avoid carcinogenic risks caused by excessive ionizing radiation,reducing radiation dose is an important trend in the field of medical imaging research.However,reducing the radiation dose will generate a lot of noise,which will seriously degrade the image quality and affect the accuracy of lesion diagnosis.In recent years,with the development of hardware technology and deep learning theory,research on image denoising based on convolutional neural network has achieved remarkable results.In this paper,we improve the image quality under the low dose condition by constructing a multi-layer nonlinear transformation network for fitting the noise features of DR images.The content of this thesis is mainly divided into the following three parts:Firstly,the noise model of the low-dose DR image is estimated.In theory,the DR image is affected by Gaussian noise,Poisson noise and multiplicative noise during the acquisition process.Through the analysis of the low dose DR of standard phantoms,it is found that the noise is mainly consist of Gaussian noise and Poisson noise when the intensity is low,and the noise is mainly consisting of multiplicative noise when the intensity is relatively high.In the log domain,the noise level is always limited to a small range,so it can be approximated as additive Gaussian noise.Secondly,based on the idea of depthwise separable convolution and multi-scale residuals,an end-to-end network model MRDNet is designed.This model uses densely connected blocks to estimate the residual information of multiple scales by progressive iteration,and performs residual feature fusion,which enhances the ability of model to tolerate the residual estimation error and to learn the noise features.Experiments show that the model can effectively remove noise and maintain image details on the NIH simulation dataset,which has a certain performance improvement compared to the existing methods.In addition,the denoised results of real radiographic data verify the reasonable assumptions of the noise model.Finally,the data set is divided based on the complexity of the image patches,and the model ED-RDNet is constructed using the ensemble strategy.Due to the uneven distribution of the gray intensity and the complexity in DR images:on the one hand,the difference of complexity between the image blocks is significant which makes the model difficult to fit;on the other hand,the difference in the number of image patches of different complexity is also significant,which affects the denoise preferences of the training model.Therefore,based on the divide-and-conquer strategy,the image patches can be clustered according to their complexity,and the normalized results of each sub-category can be used to train the sub-network model RDNet_1…Krespectively.The sub-model is based on the residual densely connected blocks and depthwise separable convolution.Then,we weight the denoising results of the sub-models to obtain the ensemble network ED-RDNet.Experiments show that,compared to existing advanced methods,ED-RDNet exceeds the performance of a single network through ensemble strategy.A significant performance improvement was achieved on the NIH simulation dataset.The denoising results of real low-dose DR phantom data and CT projection data demonstrated the strong generalization ability and practical value of the ensembled model.