Low Dose CT Reconstruction Assisted By Image Denoising

Computed Tomography(CT)has become an important non-invasive examing technology.However,recently,researches show that excessive radiation dose in CT is related to the increased risks of cancer,so the worry about health hazard of CT scan is growing.Low dose CT can actually reduce the health risk of CT scan,but will lead to noisy or sparse projection data.Using this low-quality projection data,the images reconstructed by nowadays commercial filtered back projection(FBP)will be corrupted by mass of noise and artifacts.Therefore,how to reconstruct high-quality images from low-quality projection data becomes the key and also the difficulty of low dose CT reconstruction.In the past decades,compressed sensing has attracted a lot of research interest in the field of signal restoration.Compressed sensing uses the sparsity of the signal after being transformed to a certain transformation domain,to precisely reconstruct the original signal.Sparse dictionary learning is an important method for representation learning.Sparse dictionary learning is able to remove noise from signal by searching the sparse representation on an over-complete dictionary.Nowadays,deep convolutional neural network has shown excellent performance in visual recognition.Also,it has become an attractive topic to apply deep convolutional neural network in image denoising.This thesis attempts to introduce the frontier achievements of image processing,such as sparse dictionary learning and deep convolutional neural network,into low dose CT reconstruction.Then these methods are combined with the compressed sensing to achieve better reconstruction quality.The research of this thesis is as follows:1)A low dose CT reconstruction method by combining sparse dictionary learning and compressed sensing is proposed.First,total variation minimization is used to reconstruct intermediate image from low-quality projection data.Adaptive early stopping strategy is applied to schedule the iterative reconstruction by finding an appropriate stopping point.Then,sparse dictionary learning is employed to produce the final image by processing the intermediate image.Experimental results show that the proposed method not only achieves the best imaging quality,but also outperforms the compressed sensing reconstruction in the comparison group while consuming much less time overhead.2)A low dose CT reconstruction algorithm assisted by deep convolutional neural network is proposed.Specifically,a convolutional neural network based on deep residual learning is designed.The network performs two level residual learning.On the one hand,the fact that image residue is more easier to learn is used to help the network to learn the mapping from input image to its residue.On the other hand,residual units promotes the training of deep network with better convergence.Experimental results demonstrates that the proposed method achieve the best reconstruction quality in both visual performance and objective metrics.