Research On Unsupervised Learning Method For Medical Imaging Reconstruction Problem

With the rapid development of various technologies,medical imaging has become more and more widely used in human body exploration and disease diagnosis,and it plays a very important role in the entire medical research field.However,the medical imaging reconstruction process is to reconstruct the high-quality images from part of the collected observation data,which belong to the category of inverse/inverse problems and is ill-posed.In this regard,researchers have been conducting research for half a century.Researchers proposed to use sparse transformation to obtain image prior information,turning the ill-posed problem into a well-posed problem,so as to achieve the purpose of high accurate reconstruction.With the emergence and development of deep learning,many scholars have proposed many novel methods for constructing of image prior.At the same time,how to use the powerful learning ability of deep learning and use it to extract prior information has become a hot research topic.Based on the theory of unsupervised learning and the principles of image reconstruction algorithms,this theais conducts in-depth research on the construction of image prior information and the application of unsupervised learning in medical image reconstruction.(1)In the application of fast magnetic resonance imaging reconstruction and sparse-view computed tomography reconstruction,the image high-frequency prior information is learned by using a denoising autoencoding network,and then the learned high-frequency prior is used in iterative reconstruction produre,after a certain number of iterations,the high-quality reconstructed image can be obtained.In particular,in prior learning stage,the high-frequency components of image will be modeled,which convey most of the semantic information.To achieve this goal,we first extract a set of multi-profile high-frequency components via a specific transformation and add artificial Gaussian noise to these high-frequency components as training samples.As the high-frequency prior information is learned,we incorporate it into classical iterative reconstruction by proximal gradient descent.The experimental results show that the use of high-frequency priors can effectively reconstruct feature details and present advantages over state-of-the-arts.(2)In low-dose computed tomography reconstruction applications,the score-based generation model is used to learn the image prior,and the data fidelity term is integrated into the iterative reconstruction process as a conditional term.At first,a score-based generative network is used for unsupervised distribution learning and the gradient of generative density prior is learned from normal-dose images.Then,the annealing Langevin dynamics is employed to update the trained priors with the conditional scheme,i.e.,the distance between the reconstructed image and the manifold is minimized along with data fidelity during reconstruction.The experimental results show the superiority of using score-based network to learn prior gradient and combined with simulated annealing method to update iteratively.In summary,this thesis focuses on the application of unsupervised learning to the reconstruction of medical imaging,and explores the construction of image prior and integrates it into the reconstruction algorithm.By summarizing the methods of deep learning to extract image priors for imaging reconstruction,this thesis proposes two unsupervised learning methods based on denoising auto-encoding networks and generative models,and applied to medical imaging applications.Compared with traditional reconstruction algorithms,it has achieved better results with practical feasibility.