Medical Image Super-resolution Reconstruction Method And System

High-resolution medical images provide doctors with rich pathological information and improve the reliability of diagnosis.Magnetic resonance imaging(MRI)is a commonly used imaging technology in clinic.At present,obtaining higher resolution MRI images usually relies on two ways extending scan time and using more sophisticated instruments.However,extending scan time is not advisable because the patient is required to keep still for a long time during scanning,while the use of more sophisticated instruments obviously leads to the imaging cost increase.Image super-resolution reconstruction aims to obtain a high-resolution image from low-resolution images by using image processing techniques,which is more promising in improving medical image quality.For MRI images,we are focus on development of super-resolution reconstruction algorithms in this paper.Our work includes the following three main parts:(1)Considering the traditional image feature extraction methods cannot well detect the difference in feature structure and the difference of image blocks is not considered when training overall image blocks,a grouped mixed model based on differential curvature is proposed for image super-resolution reconstruction.Firstly,on the basis of gradient feature extraction,the differential curvature algorithm is used to further detect feature structures such as edges and ramp,during which the feature blocks are divided into three groups: smooth regions,texture regions and edge regions.Secondly,the student's t-distribution mixed model is applied to learn the model parameters of the three sets of feature blocks.Finally,multiple distribution models with larger likelihood probability are selected to reconstruct high-resolution image patches,which further improves the accuracy of the reconstructed image patches.Experimental results show that the proposed method obtains not only the higher peak signalto-noise ratio and structural similarity index,but also the more effective visual results,with less time for super-resolution reconstruction of the brain MRI image.(2)Considering the feedforward convolutional neural network cannot combine high-level information with shallow-level information,which is insufficient to obtain contextual information,a feedback network with self-attention mechanism is proposed for MRI image super-resolution reconstruction.Specifically,a feedback network with feedback function is built to correct shallow features and obtain deeper features by combining deep features with shallow features.Then,the self-attention mechanism is integrated into the feedback network,which can adaptively calibrate the characteristic response of the image,reduce redundant information,and improve the ability of the feedback network in processing information.At the same time,residual network with dense skip connections is designed to generate rich feature maps of image patches.After multiple time steps of iterative optimization,a brain image with stronger characterization capabilities is reconstructed.Experimental results show that the proposed network model is superior to some other super-resolution reconstruction algorithms both in quantitative evaluation and qualitative analysis.Moreover,the restored texture details in super-resolution reconstruction results are clearer and closer to the real brain MRI images.(3)A medical image super-resolution reconstruction system is designed.The algorithm proposed in this paper and the classic reconstruction algorithms are integrated into the system,which assists doctors to selecting different algorithms for image super-resolution reconstruction.In addition,the system provides detailed view function,which magnified display of pathological structure and evaluation function,which allows doctors validating the quality of the reconstructed image quantitatively and qualitatively.