Research On MR Image Super-resolution Based On Samples Space Learning

Magnetic resonance(MR)images can detect small pathological tissue at an early stage.Therefore,it is widely used for lesion location and disease diagnosis,High-resolution MR images can provide clearer structural details and help doctors to correctly analyze and diagnose the disease.At present,the imaging resolution is mainly improved by magnetizing more free water in human tissues and organs,which will cause the main magnetic field of the magnetic resonance imager and the radiation time and radiation intensity of the loaded electromagnetic wave to increase.However,excessive radiation can lead to serious consequences such as body overheating and protein inactivation,so it is not suitable for clinical application.Therefor,the use of super-resolution technology to improve the resolution of MR images is an effective solution.In view of the characteristics of MR imaging,it is difficult to meet the performance requirements using traditional image super-resolution methods.The sample space learning method can effectively utilize the existing image prior knowledge,which has certain advantages compared with the traditional signal processing-based super resolution method,but there is still room for improvement.The main innovations of this thesis include:A super-resolution framework suitable for MR images is constructed,which is divided into small sample space learning and large sample space learning according to the number of samples.A MR image super-resolution method based on sample space learning is proposed,and a dictionary learning method based on sparse representation is used to achieve MR super-resolution in small sample space.A deep convolutional learning network is proposed to achieve super-resolution based on MR images in a large sample space.Combining the MR imaging mechanism to construct an image data set,an image super-resolution quality evaluation method is given,which verifies the performance of the MR super-resolution framework based on sample space learning.Experimental results show that the super-resolution method based on sample space learning is feasible and superior to traditional super-resolution methods.A MR image super-resolution method based on sample space optimization is proposed,which solves the problem that the imbalance of individual sample quality affects the super-resolution performance.First,a MR super-resolution algorithm based on individual sample optimization is proposed,which uses image gray-level consistency and one-two gradient methods to measure the quality of individual samples,eliminates singular samples that affect super-resolution performance,and constructs an optimal sample space;Secondly,a sample space optimization method based on gray-level consistency-gradient joint evaluation is proposed.This method applies complexity and gradient as the characteristic dimensions of the sample distribution,applies the gray-level consistency method to calculate the complexity to determine the global characteristics of the sample,and applies the gradient method to calculate the first-order and second-order degrees of the image to determine the local complexity of the sample,constructing a sample space suitable for MR image super-resolution.Experimental results show that the proposed optimization method can effectively improve the image super-resolution performance in small sample and large sample space.A super-resolution optimization method for MR images based on small-sample spatial dictionary learning is proposed to improve dictionary represen tation ability from the perspective of loss function and reconstruction factor.First,an error loss function based on reconstruction quality constraints is proposed.The high-and low-resolution dictionary reconstruction error is independently calculated,and the individual reconstruction error of the high-and low-resolution dictionary is considered,and the traditional cascade calculation method is abandoned.This method effectively reduces the reconstruction error,and solves the problem that the traditional dictionary joint cascade training does not consider the individual reconstruction error of high and low resolution dictionaries;Secondly,the multi-factor joint learning optimization method is adopted to solve the coupling problems of the balance parameters,overlapping blocks,dictionary block size,block number and other reconstruction factors in the MR super-resolution process,and further improve the dictionary representation ability in the small sample space.The experimental results show that the proposed dictionary learning method of joint optimization of loss function and reconstruction factor effectively improves the super-resolution performance of MR images in a small sample space.A MR image super-resolution optimization method based on deep learning in large sample space is proposed,which improves the high-and low-resolution feature mapping capabilities from the three perspectives of network structure,data characteristics and heterogeneous network fusion.First,a super-resolution network structure based on multi-slice input optimization is proposed,which makes full use of the structural similarity between MR slices.Secondly,in view of the problem that the traditional loss function does not consider the human visual perception,a multi-loss function cascade optimization method is proposed to preserve the image’s color,brightness and high-frequency area contrast characteristics.Third,a method based on heterogeneous network fusion is proposed,which integrates the output of a single network through an additional fusion layer,broadens the width of the network,and can effectively improve the mapping and characterization capabilities of high and low resolution features.The experimental results show that the proposed deep learning network multiple optimization method effectively improves the super-resolution performance of MR images in a large sample space.