MR Image Super-resolution Algorithm Based On Deep Learning

Magnetic resonance imaging is a widely used medical imaging method in clinical practice,with non-injury,high soft tissue contrast characteristics,commonly used in the diagnosis and treatment of some diseases.Clear magnetic resonance image can provide rich structural information of organs,but the resolution of magnetic resonance image is limited by environmental factors such as noise and scan time.Therefore,the processing of low-resolution and high-noise magnetic resonance images has become an important part of medical image processing.In this paper,the magnetic resonance image super-resolution reconstruction algorithm based on deep learning is studied.Specifically,by building multi-scale residual modules in the depth structure,it is ensured that more high-frequency features of the image are added to subsequent network layer structures.On the other hand,in order to maintain image texture information and reduce image noise,we add the total variation regularization to the loss function for smoothing processing.By comparing the experiments on three clinical MRI brain imaging data sets,it was proved that the proposed algorithm improved significantly in both the peak signal-to-noise ratio and the structural similarity and the visual effect compared with the previous method.The main work of this article is as follows:In the first chapter,we introduce the research background of magnetic resonance image super-resolution reconstruction,summarize the research development of ultraresolution algorithm at home and abroad,and discuss the research motivation and research content of this paper.In the second chapter,we introduce the theory and method of image superresolution reconstruction.In particular,the hypermetric reconstruction algorithms based on convolutional neural networks are discussed from the network structure.In the third chapter,we design a super-resolution algorithm based on multi-scale residual module and full-variation regularization.Specifically,after introducing the network structure from the three parts of feature extraction,nonlinear mapping and image reconstruction,the loss function of multi-scale residual module and the wholevariation regular term is proposed.Further,the operation of the image super-resolution network architecture is realized.In the fourth chapter,the three clinical magnetic resonance data sets and pretreatment processes used in the experiment are described at first.Secondly,the peak signal-to-noise ratio and structural similarity of the two experimental evaluation indicators are introduced.Finally,the numerical results of the proposed algorithm and the comparison algorithm on three clinical data sets are shown.In particular,the effectiveness of multi-scale residual module and full-variation regularity is verified by ablation experiment.The fifth chapter is the conclusion and outlook part.We summarize the main work of this paper,and put forward several follow-up to be further studied direction.