Research On Image Restoration Of Unpaired Samples Based On Deep Learning

Image restoration refers to the task of recovering the original high-quality image from the low-quality image.Image degradation occurs during the process of image formation,transmission,and storage,including additional noise,down sampling,blurring,and pixel damage.Therefore,image restoration tasks mainly include:image denoising,image deblurring,image super-resolution reconstruction,and image repairing.The thesis mainly focus on the research of image restoration tasks such as denoising and super-resolution based on CT(Computed Tomography)images.CT is one of the most important imaging methods in modern radiology.Due to its ability to show fine details of internal organs,CT has been widely used for the diagnosis and early screening of diseases.Unfortunately,it also have some disadvantages while CT scans bring convenience to people.For example,BH(Beam hardening)artifacts appearing in reconstructed CT images due to partial energy absorption when X-rays pass through the object during CT scanning.And in order to reduce damage to the human body,the reduction of doses leads to noise appearing in reconstructed low-dose CT images.These problems,including low resolution of the image,will reduce the quality of the reconstructed CT image and may affect the diagnosis of the disease.At present,there are some traditional methods for beam hardening artifact correction,such as hardware filtering,dual energy,statistical polychromatic reconstruction,and linearization.In addition,there are some model-based correction methods and learning-based methods.Although there are so many mature beam-hardening artifact correction methods,there is a lack of systematic software for doctors to correct artifacts.For low-dose CT image denoising,the traditional methods mainly include three types:sine map filtering,iterative reconstruction,and post-reconstruction processing.With the development of machine learning,some low-dose CT image denoising methods based on learning have also appeared.These methods achieve better denoising effects than traditional methods.However,these learning-based methods require paired training data,and In actual situations,paired training data is difficult to obtain.Aiming at the above problems,the thesis conducts research on low-dose CT image denoising and image super-resolution reconstruction based on unpaired training samples.The thesis does the research from the following aspects:(1)BH artifacts correction and software packaging.In this thesis,a model-based BH artifacts correction method is implemented,and a corresponding software is developed.The actual CT images with BH artifacts was obtained using a home-made phantom model,and the function of the software was verified based on this data.The results show that the software developed in this thesis effectively suppresses artifacts on real data.(2)Research on Methods of low-dose CT images denoising.In view of the fact that paired data is difficult to obtain in actual situations,the thesis proposed a network to achieve low-dose CT denoising with unpaired training data.The network includes two generators for generating the denoised image and noise respectively.The synthetic low-dose CT image is obtained by combining the outputs of the two.There are also two discriminators for distinguishing the denoised image from the Conventional dose CT images and distinguish synthetic and true low-dose CT images.After comparative analysis with the existing cycle-GAN network model based on unpaired sample training,it is found that the method proposed in this thesis has better visual results on unpaired image denoising tasks.And the PSNR and SSIM values are 8.01 and 0.05 higher than the cycle-GAN network,respectively.In addition,the results of the proposed method on paired data are basically the same as some existing deep learning-based methods,and even perform better on some image quality evaluation indicators,such as PSNR and SSIM.(3)Research on image super-resolution reconstruction based on deep learning methods.For the natural image dataset,this thesis summarizes and analyzes existing deep image super-resolution reconstruction methods based on deep learning,and combines EDSR and SRGAN networks to implement a new enhanced SRGAN network.Based on the existing methods and improved methods,this thesis implements the super-resolution reconstruction task of natural images under paired and unpaired data.And the effectiveness and pros and cons of various methods are summarized through analysis and comparison.For the CT image super-resolution task,the article first down-samples the collected high-resolution data to obtain low-resolution data,and then applies several models in the natural image super-resolution reconstruction task.And the original cycle-GAN and SRGAN network are combined to implement the SR-cyclegan network for unpaired training.Based on these methods,CT image super-resolution reconstruction tasks under paired and unpaired data are implemented.And a quantitative analysis of the pros and cons of all methods was also performed.