Artifact Removal In Sparse-angle CT Based On Perceptual Loss Feature Fusion Residual Attention Network

With the more and more extensive application of CT imaging technology in the process of modern medical diagnosis and treatment,the potential radiation risk in CT examination has also attracted extensive attention.Excessive radiation is easy to induce leukemia,cancer and other diseases.Therefore,it is urgent to reduce the radiation dose in the process of CT scanning.Using the sparse sampling method of scanning at a certain angle interval can obtain the sparse angle CT image with low radiation dose,but there are artifacts in the sparse angle CT image,which reduces the image quality and affects the doctor’s diagnosis.Aiming at removing artifacts from sparse angle CT images and enhancing CT image details,this paper constructs two network models to realize the task of removing artifacts from sparse angle CT images.The main research work of this paper is as follows:Firstly,research on the artifacts removal of sparse angle CT based on convolutional neural network.Several neural networks with different structures are studied to compare the performance of these networks in removing artifacts from sparse angle CT images.The experimental results show that the Feature Fusion Residual Network(FFRN)has a good performance in removing the artifacts of sparse angle CT images,and it has a good effect in parameter quantity and quality of artifact removal,so the improved baseline network is defined as FFRN.Secondly,in order to enhance the effect of artifact removal,a new residual attention module is designed,and the residual attention module is integrated into FFRN network,and a residual attention network based on feature fusion is proposed.The output feature map is used as the input of the residual attention module,and the weight distribution matrix is obtained through a series of calculations.For the channel of artifact feature information,a small weight is assigned to reduce the degree of concern or filter out the artifact information,so as to achieve the goal of artifact removal.The experimental results show that the residual attention network of feature fusion is better than the baseline FFRN in removing artifacts from sparse angle CT images,and the PSNR is improved by about 1.12% compared with FFRN,the SSIM value is about 0.580% higher than the FFRN value.Thirdly,in order to enhance the image detail,the perceptual loss is introduced into the loss function of network training,and a feature fusion residual network based on perceptual loss is proposed.The VGG19 network is used to obtain the eigenvalues of the predicted image and the real image,and the eigenvalues of the predicted image and the real image are taken as the input of the perceptual loss to calculate the loss value.The experimental results show that the effect of residual attention network based on perceptual loss is better than that of baseline FFRN,and the PSNR is improved by 2.43% compared with FFRN,the SSIM value is about 0.856% higher than the FFRN value.