Design And Implementation Of Low-dose CT Denoising Algorithm Based On Deep Learning

Computed Tomography(CT)is an extremely important auxiliary diagnostic method in clinical practice at present,which has a high diagnostic value for central nervous system diseases,head and neck diseases and chest diseases.However,CT scans expose the patient to intense radiation,are likely to destroy some tissue structures of the human body and induce some diseases.In order to reduce the radiation injury of patients in CT scanning,the concept of low-dose CT(LDCT)has been proposed and tried to be used in clinical examination.However,the reduction of radiation dose lead to a large amount of noise and artifacts in reconstructed CT images,and the resolution of low-dose CT images becomes lower,which seriously affect the wide application of LDCT in clinical practice.In order to meet the clinical requirements of LDCT,many LDCT denoising algorithms have been widely proposed,and LDCT denoising algorithm has become a hot spot in the field of CT imaging.With the popularity of deep learning theory,the model based on deep learning has become the most concerned technology in the field of image processing,so some scholars have tried to apply deep learning model to LDCT denoising and achieved many results.Under this background,this paper deeply studies the topic of low-dose CT denoising.At present,there are two main ways to realize low-dose CT in clinic:reducing radiation current and sparse angle scanning.In view of the insufficient performance of the existing denoising models,this paper designs denoising algorithms for these two realization methods,respectively.The details are as follows:(1)Aiming at the quantum noise in low radiation current CT,this paper designs a convolutional neural network including multi-axis gated Multi-Layer Perception(MLP)module to remove the noise in CT images.At present,most of the proposed low-dose CT denoising models are based on convolutional neural networks.However,since the convolution operation can only process one local neighborhood image block at a time,the convolution model cannot obtain long-distance dependence in denoising,that is,it cannot take advantage of the non-local similarity of CT images.However,the non-local similarity of images has been proved to be effective prior information for image denoising,so this paper uses multi-axis gated MLP module to calculate the non-local dependence information of images and add it to convolutional neural network for image denoising.Experiments show that the model designed in this paper has excellent denoising performance.Compared with the simple convolution denoising model,the performance of the model is improved by more than 0.5dB,which can effectively remove the noise of low-dose CT images while retaining more detailed information.(2)Aiming at the problem of missing projection domain in sparse angle scanning CT,this paper designs a recursive repair network for data interpolation in CT projection domain.Previous interpolation network models are all composed of a single network,which is equivalent to an end-to-end mapping.Although this structure is simple,the interpolation effect is limited.In order to improve the performance of the model,the depth of the network must be increased,which will lead to the increase of model parameters and the difficulty of model training.Therefore,this paper designs a recursive repair framework,which can realize continuous optimization of interpolation data through recursive operation,thus improving the quality of interpolation.Through experimental comparison,the network designed in this paper has a good restoration effect on CT projection data obtained by sparse angle scanning at multiple times.Compared with a single interpolation network,the model performance can be improved by more than 1dB,which can significantly improve the reconstructed CT image quality.