Research On Denoising And Information Reconstruction Of CT Images Under Low-dose Radiation

Computed Tomography(CT)plays a very important role in modern medical diagnosis.However,the security risks caused by X-rays used in this technology have aroused many people's concerns.In recent years,with the rapid development of deep learning technology,researchers have proposed some deep noise reduction algorithms,which achieve better results than traditional methods.However,the existing algorithms still have some problems,such as poor noise reduction quality,over smooth noise reduction image,and so on.More effective low dose CT image denoising solutions still need to be proposed.Giving consideration to both noise suppression and detail preservation,and generate high-quality denoising results,this paper proposes new low-dose CT image denoising model and loss function,which contributes a new technical solution to solve the above difficulties.This paper's research work mainly includes the following two parts:(1)In order to solve the problem of insufficient noise suppression ability of existing methods,an end-to-end noise reduction model with stronger noise fitting ability is proposed in this paper.In this part,this paper designs a new feature extraction module:edge enhancement module based on training Sobel convolution.The module introduces learnable parameters into the traditional Sobel operator to control the intensity of edge feature extraction and enhance the ability of edge feature extraction.In addition,a new low-dose CT image denoising model(the EDCNN model)is proposed.In the model structure,based on the idea of making full use of the input information,dense connections are added between the features extracted by the edge enhancement module and the subsequent modules.At the same time,the model adopts the structure design of learning the gap between the input image and the target image directly.It makes the model more effective for this noise reduction task.(2)Aiming at the problems of image smoothing,loss of details and image blur in image denoising task,a new composite loss function is proposed in this paper.As the objective function of optimization,the loss function integrates multi-scale perceptual loss in addition to the mean square error(MSE)loss commonly used in current mainstream methods.During the experiment,lots of comparative experiments are carried out on the AAPM Mayo dataset,and a variety of configuration schemes of two kinds of loss are discussed.The final design of composite loss function can effectively deal with image blur and over smooth problems.In the visual perception of the image,it significantly improves the image quality of the de-noising results.In this research work,lots of experiments are implemented based on the commonly used AAPM Mayo dataset of this task.Through the ablation experiments of the model's structure and loss,this paper explores the design scheme of the model and the role played by each component in more detail.By comparing with existing methods in this task,the algorithm shows excellent noise reduction performance.