Super-Resolution Of Nuclear Magnetic Imagebased On Hierarchical Fusion Network Undermulti-Attention Mechanism

With the development of modern medical technologies,medical imaging technology has also achieved great progress.Commonly used medical imaging techniques include Cardiovascular Imaging,Electronic Computed Tomography(CT),Positron Emission Tomography(PET),and Magnetic Resonance Imaging(MRI).Among the aforementioned technologies,MRI is widely used in clinical use due to its non-radiation,multiple parameters,and wide examination range properties.However,due to factors such as the high expense of high-precision equipments and large scanning noise interference,the quality of MRI imaging is varied in quality,which can greatly affect doctors' clinical diagnoses.Therefore,improving the quality of MRI images is of great significance in medical research and clinical diagnosis.The super-resolution of MRI has become one of the research hotspots in the field of medical imaging in recent years,but the current super-resolution methods usually only use a single model to process the high-frequency and low-frequency information in the image,such as the texture information and spatial structure information in the image.A certain layer of network can only recover relevant information based on the output of the previous layer.This method generally has the problem of not being able to recover specific frequency information,neither able to fuse adjacent frequency information effectively.To this end,this thesis proposes a super-resolution method for MRI with hierarchical fusion under a multi-attention mechanism,which uses hierarchical learning strategies and a recursive fusion between hierarchies to construct a super-resolution network framework,and introduces multiple attention mechanism to further improve the quality of NMR images.The main work shown as follows.First of all,a hierarchical learning strategy is proposed considering the information of different frequencies existent in the image content and the different degrees of model adaptation.This strategy separates the information of different frequencies.According to the correlation between different frequency information and model complexity,the low-level frequency focus module is used to restore low-frequency information,and the high-level frequency focus module is used to restore high-frequency information,which solves the existing problem on over-recovered and under-recovered information.Secondly,based on the characteristics of high correlation between information and information frequency,a network framework of hierarchical recursive fusion between levels is designed,which restores information of different frequencies according to hidden correlation,and integrates low-level and high-level frequency-focused modules' output information separately,so that the image is first integrated with the associated information of the adjacent frequency during the restoration process,which effectively solves the problem of the existing weak correlation of the adjacent frequency information.Finally,based on the aforementioned hierarchical learning strategy and the inter-level recursive fusion network framework,a multi-attention lower-level fusion NMR image super-resolution network is proposed.The network uses the channel space to integrate the attention module to learn the feature information of each channel and location in the feature map.At the same time,the hierarchical attention module is introduced to adjust the correlation between the features of each layer,which further improves the image quality of MRI super-resolution.The experimental results show that the MRI super-resolution network under the multi-attention mechanism proposed in this thesis makes full use of the information of different frequencies of the image,has better adaptability to medical images,and can effectively improve the super-resolution output quality of the image.Compared with traditional methods and super-resolution methods based on deep learning,this method has better performance in subjective observation and evaluation of various objective indicators,and provides an effective idea for improving the quality of NMR images.