Adaptive Filter Low-dose CT Image Reconstruction Using Deep Convolution Neural Network

X-ray computed tomography has been widely used in modern clinical practice.However,the radiation damage to the patients during CT scanning has raised wide concerns.Excessive radiation exposure might induce leukemia,cancer and other diseases.Therefore,it is necessary to reduce the radiation dose in CT scanning.The reduction in radiation dose would lead to severe noise and artifacts in reconstructed images,which would increase the diagnosis difficulty.In order to improve the CT image quality at low dose,many low dose CT image processing techniques have been presented.These techniques can be divided into five categories: sinogram restorations,post-processing methods,iterative reconstruction techniques,dictionary learning techniques and deep learning methods.Among these techniques,deep learning methods are the most promising low dose CT image processing techniques.These methods utilize a large number of training data to mine the priori knowledge and therefore improve the image quality of low dose CT scanning.These methods have high processing speed and image quality.This thesis mainly focus on low dose CT imaging algorithms based on deep learning.Most existing deep learning methods solve this problem by suppressing noise in projection data or in image data using neural networks.Although there are many impressive research results,these methods have some limitations.Firstly,these methods only focus on projection domain or image domain,without comprehensive use of projection data and image data.Secondly,these methods all require conventional CT reconstruction algorithms.The final effects will be affected by the conventional CT reconstruction algorithms used.Here we handle the low-dose CT reconstruction problem in a totally different approach,which focus on directly reconstructing high quality CT images from low dose projection data by neural network and the neural network can complete image reconstruction and noise suppression simultaneously.Based on the new research idea,this paper firstly introduced a low dose CT image reconstruction network named Dual-Res Unet,which can directly reconstruct CT image from low dose projection data.The network uses two Res Unet networks to perform noise suppression in projection domain and in image domain respectively and a filter backprojection module to perform image reconstruction.Through experiments on AAPM dataset,the effectiveness of Dual-Res Unet has been proved.To further improve the performance,this thesis introduced another low dose CT image reconstruction network named Dual-Res Unet-MR,which can take advantage of multiple reconstruction filter kernels.This network further improves the low dose ct imaging quality.By introducing a reprojection module,this network improve the data consistency between reconstructed results and projection data.Experiments have been conducted on multiple datasets and the experiment results show that this network has excellent performance.