Research On Medical Image Using Super-resolution Reconstruction

Nowadays,medical images provide an important basis for clinical diagnosis,including the X-ray Computed Tomography(CT)images and Positron Emission Tomography(PET)images.As these medical images are widely used,we have to pay attention to the radiation effects on human body,that caused by the X-ray during CT imaging and radiation agents during PET imaging.In order to reduce this effect,low-dose CT imaging and low-count PET imaging have emerged in the field of medical imaging,which are also hot topics in the field.Image Super-resolution is a popular image restoration technique.With the mature application of compressed sensing theory,there is a better solution to the ill-posed problem in super-resolution reconstruction,which enables reconstruction to be restored beyond the traditional methods.Moreover,with the deep learning of the network framework applied to super-resolution reconstruction technology this year,making it further developed and applied.Therefore,this thesis propose to use the super-resolution reconstruction techniques to solve the problems of low-dose CT imaging and low-count PET imaging.The main research work of the thesis are as follows:(1)This thesis studies the use of sparse representation based super-resolution reconstruction methods to recover poor quality medical images.We treat low-dose CT and low-count PET images as degenerate low resolution(LR)images,and normal CT and PET images as high resolution(HR)images.After super-resolution reconstruction of the LR image,an HR image with improved quality can be obtained.In order to obtain accurate reconstruction results,we propose to use a dilution-based method.This method requires first sparse coding,training a pair of overcomplete dictionary and(is a low resolution dictionary,is a high resolutiondictionary),and then use to get a sparse representation of the input LR imagepatches.The corresponding HR image patches is reconstructed with the high-resolution dictionary and these sparse representation.The complete HRimage can be obtained after all predicted image patches are stitched.(2)This thesis also proposes to use super-resolution method based on LISTAnetwork.This method combines the model of the sparse representation based method with the sparsely encoded LISTA network to obtain a new CSCN network for reconstruction.The focus of this method is on the selection of training data sets.We use a set of CT images with normal dose as the network input and a low-dose CT image with artifacts as the network output to train the parameters in the network.Finally,a network that can effectively suppress artifacts in CT images.This thesis proposes two super-resolution reconstruction methods to process low-dose CT and low-count PET images.Through experiments,the feasibility of the proposed algorithms is proved.Compared with the conventional super-resolution reconstruction method,the proposed methods not only can effectively recover the details of the fuzzy CT and PET images,but also the LISTA-based method can suppress artifacts in low-dose CT images.For low-dose CT imaging and low-count PET imaging,the work of this thesis provides a new opinion to solve them.