Researches On Low Dose CT Image Denoising Based On Multi-scale Residual Encoder Decoder Network

Computed Tomography(CT)can non-invasively acquire the internal structure of human by the difference in X-ray Absorption rate of human tissues.So CT has been widely used in clinical screening and diagnosis of diseases.Given the potential risk of X-ray radiation to human body,low-dose CT Denoising has become a major topic of medical imaging.Currently,reducing X-ray dose and exposure time is the main method to decrease the X-ray cumulative radiation.However,the reduction of X-ray dose will introduce more quantum noise,resulting in a worse CT image quality.How to improve the quality of imaging while reducing the radiation dose is the major challenges of low-dose CT.Now,the mainstream methods of lowdose CT denoising include iterative reconstruction,sinogram domain filtration and postprocessing algorithms.Because of the good portability,post-processing algorithm has gradually become a research hotspot.Low-dose CT denoising is a typical ill-posed problem.Due to the powerful ability to deal with this kind of problem,deep learning has gradually become the mainstream method.The deep learning based method is to train a convolutional neural network(CNN)by mapping the low-dose to normal-dose CT images.However,CNN-based methods heavily rely on the convolutional kernels,and the fixed-size filters lacks the scale transformation of the local neighborhood,resulting in low utilization of image structure information.In addition,CNN-based methods need to processing massive data,and the huge number of parameters brings computing burden and increases the difficulty of network training.Third,CNN-based methods are prone to over-smoothing in the process of denoising,resulting in the loss of image details and blurred edges.To address above problems,this paper performed a series of research in two steps:The first research topic is the research of low-dose CT denoising based on a multi-scale residual encoder decoder network.In order to improve the efficiency of image information utilization and increase the scale of feature extraction,three kinds of multi-scale feature extraction modules were designed.The kind is a feature extraction module with multi-branched dilated convolutions The second kind is a feature extraction module with a serials of dilated convolutions.And the last kind is a feature extraction module with wavelet transform.By combining with the classical residual encoder decoder network REDCNN,the three feature extraction modules are verified experimentally.The experiments show that the three multi-scale feature extraction modules can effectively improve the performance of the network,the multiscale kernels based on dilated convolution has practical significance for improving the performance of the network.The second research topic is the research of low-dose CT denoising network model based on generative adversarial network with attention mechanism.On the basis of the above research,in order to improve the performance and reduce the number of parameters,a REDCNN-based dilated convolution neural network model DREDCNN was proposed,which effectively improved the network performance of REDCNN.To avoid the problem of over-smoothing,a multi-branch attention mechanism module was designed to enhance the ability of the network for maintain the structure of CT images.At last,we design a loss function by adopting a weighted combination of MSE loss,perceptual loss and Earth Mover’s Distance,and then use the framework of generative adversarial network to preserve the image details and structures.The experiment verified that the designed self-attention block and the combined loss function is effective for alleviating the over-smooth phenomenon in low-dose CT denoising.