Low-dose CT Image Processing Using Convolution Neural Network

The primary goal of low-dose Computer Tomography(LDCT)research is achieving the imaging quality of normal dose CT under the low-dose scanning conditions.The state-of-art LDCT processing methods can remove noise and artifacts,whereas they may cause blurred edge and reduced image contrast.Deep learning has powerful representation of global and local features through constructing complex nonlinear deep network model,which has made big breakthrough in computer vision field in recent years.Therefore,we conducted LDCT research based on deep learning,aiming at improving the quality of LDCT,meanwhile reducing radiation in medical diagnosis and treatment.This paper contains the following two parts:Filter the projection data to remove the noise using deep covolution networks.Then,reconstruct the filtered projection data via FBP.The main idea of the algorithm is constructing an end-to-end two dimensional residual convolution network reflecting the complex mapping between the low-dose photon number and its contained noise,with the low-dose and high-dose photon number as the training set which are obtained through exponential transforming and Anscombe transforming the corresponding projection data.The network model is then extended to three dimensions to make full use of the information from the different angles projection data.The results show that the 2D model can effectively suppress noise,causing image blurred,whereas 3D model can both improve the denoising effect and preserve the contrast.Deep learning based postprocessing method is proposed to improve the image quality of LDCT containing a large number of noise artifacts.With low-dose and high-dose CT images as trainging set,this paper has construct an end-to-end two dimensional convolution network,an image-to-image regression model,to represent the LDCT mapping to its contained noise and explored the factors may influence the model performance.Experiment results show that the residual network with dropout,a reasonable loss function,increased network depth and width,can improve the network performance.It can be seen that image details,like the vessels,somethimes can be smoothed out due to neglection of three dimensional structure information of the human body.To overcome the issue,a 3D exitension network model is constructed to learn the noise contained in the 3D CT blocks.Compared with the image quality evaluation of results of 2D low-dose post-processing model,3D network can achieve better denoising effect and preseve detail issues like the vessels more effectively by utilizing the 3D structure information to extract the discriminative features of human body structure and noise.