CT Image Reconstruction Based On Deep Learning

Computed Tomography irradiates objects by X-ray,and reflecting the internal structure information by reconstructing through algorithm.However,too much radiation would hurt the patients during the treatment.Therefore,the research about low-dose CT reconstruction is very important and meaningful.Since deep learning has achieved a significant improvement in computer vision,so in this paper,we combine the deep learning techniques with classical CT reconstruction method.We focus on post-process,statistical iteration,analytical reconstruction,and proposed a denosing method and two reconstruction algorithms.All these three methods can improve the quality of low-dose CT reconstruction images.The most direct way to improve the quality of low-dose CT image is to denoise in the image domain.Although there are numerous classical denosing methods that have effectively applied to improve the quality of CT images,easy to appear some problems like artifacts and ladder effects.A series of convolutional neural network based methods can get better results,however there are gradient vanish problem when the network is deeper.In this paper,we introduce the residual blocks to the denoising network,which will make the network deeper and has better results.At present,statistical-based iterative CT reconstruction show the better performance in low-dose CT reconstruction.The key of this method is the design of the penalty term which is difficult to be well designed.In this paper,we start from the optimization objective function of iterative reconstruction,and replace the proximal operator with learning operator in PDHG and ADMM algorithm.What's more,we further extend it to a more general form and expand it as several connected network.This structure can not only avoid the penalty term design,but also improve the quality of low-dose CT images obviously.Analytical reconstruction is widely used in CT reconstruction due to its speed advantage,but the reconstruction quality is worse when use low-dose ray.In this paper,we re-design the filtering function and interpolation method in filtered backprojection algorithm.And add simple convolutional module in projection domain and image domain respectively.Though the new method involves only a small amount of computational cost,it can improve the quality of the low-dose reconstruction image approach to statistical based reconstruction methods.Experiments are carried out on the simulation dataset and the real dataset.And the results in this two datasets both show that our methods get better performance than classical methods.