A Study On Medical Image Super Resolution Reconstruction

Recently,with the rapid development of medical imaging,ultrasound imaging,computed tomography imaging and magnetic resonance imaging etc.have become the most important techniques for assisted diagnosis.However,due to the limits of radiation dose,acquisition time,imaging setup and the patient's motion,the medical images are usually degraded during the acquisition.Therefore,improving the quality and increasing the spatial resolution of medical images will be very significant for increasing the diagnostic accuracy.To achieve this,in terms of the issues of current image super resolution(SR)reconstruction methods,this thesis intends to investigate the novel medical image super-resolution reconstruction algorithms.The detailed innovations in this work are as follows:(1)Proposing a novel image SR reconstruction method based on regularizations of total variation(TV)and improved non-local total variation(NLTV),which modifies the traditional NLTV regularization by integrating Gaussian,Laplacian,and Cauchy distributions.Such method takes advantage of the merits of TV and improved NLTV,it is able to recover the image details and edges,therefore,it can overcome the defects of traditional regularizationbased SR reconstruction methods and realize the accurate reconstruction of medical gray images.(2)To deal with the issue that the regularization-based SR reconstruction methods are sensitive to the regularization priors,a convolutional neural network(CNN)-based SR reconstruction model is presented,which not only realizes the accurate reconstruction for the medical gray images but also guarantees the accuracy of parameter maps extracted from the reconstructed images.(3)To overcome the problem that the performance of CNN-based SR reconstruction models is highly dependent on the dataset and the model structures,a SR reconstruction algorithm based on the mutual similarity of image patches is finally introduced.In light of the mutual similarity of the image patches at multi-scales,the relationship between the low-resolution and high-resolution image patches is learned using an improved patch matching method,from which the image SR reconstruction can be realized.Without the training with external datasets,such method can get the same or even better SR reconstruction results.