Research On Nuclear Magnetic Image Quality Improvement Method Based On Deep Dual Cascaded Network

With the improvement on medical conditions,image-based diagnostic technology has been greatly developed,and related imaging technologies also continue to mature.For the progress of Computed Tomography(CT)and Positron Emission Tomography(PET)to Nuclear Magnetic Resonance imaging(NMR),the transition from planar to stereoscopic multidimensional imaging has been experienced.It is noteworthy that there is no radioactive damage in MRI so that it is more safety,which can directly display multiple axial positions and layers in any direction.The MRI has been widely used in clinical practice.In practical applications,the quality of MRI is limited by some factors such as hardware device performance,space constraints and signal to noise ratio(SNR).In order to improve the image quality of nuclear magnetic images,and assist doctors to determine the location of lesions more accurately,a large number of related works have been conducted.On this basis,this paper deeply studies the method of super-resolution quality improvement of nuclear magnetic image based on deep learning.B y exploring the causes of image degradation,we also combine human visual characteristics and deep learning in our work.The main works are as follows:Firstly,on the basis of existing deep learning based researches,we propose an algorithm for nuclear magnetic resonance image super-resolution reconstruction based on sub-band residual learning.It designed a two-way residual learning network,and by gradually refinement,the feature learning of images and super-resolution reconstruction are effectively combined and optimized.The algorithm has flexible scalability,and greatly leverages the feature representation capabilities of deep learning methods to effectively improve the image reconstruction quality.Secondly,inspired by the concept of receptive field in convolutional neural networks(CNNs),we analyze the role of feedback mechanism in human visual system,and then introduce the feedback mechanism into the image super-resolution task.Specifically,a trainable feedback mechanism is designed and applied in the feature learning process of CNN.Then,a feature learning model based on feedback mechanism is proposed.By simulating the feedback mechanism of image processing in human brain,feature learning comes more effective,which greatly reduces the difficulty of feature learning relying on a single forward network.Finally,we explore the existing problems of super-resolution networks based on CNN.It is found that deep convolutional networks are not conducive to the detailed feature learning of CNN,while image super-resolution tasks require the details of edges,textures,etc.,which are provided by shallow networks.On this basis,a progressive learning strategy is proposed.On one hand,the deep network can be “segmented” to provide sufficient shallow detail features for the network;on the other hand,the network can maintain a certain degree of nonlinear mapping capability.Therefore,it is effectively solved that the problem of how to apply the deep learning method to super-resolution reconstruction when the nuclear magnetic data possess highly similar characteristics.The experimental results show that the proposed method mines the deep relationship between low-resolution images and high-resolution images.Compared with traditional and deep learning-based methods,it both achieves better quality-improvement and provide a more effective idea for MR image processing.