Research On Medical MRI Super-resolution Algorithm Based On Deep Network

Magnetic resonance imaging(Magnetic Resonance Imaging,MRI)is mainly used for the diagnosis of pathological sites and image-guided treatment.Without harming the physiological parts of the human body,it is difficult to improve the imaging resolution of hardware.Doctors can use the information provided by high-resolution MRI(Magnetic Resonance,MR)images to effectively diagnose the pathological part after locating it.Deep network refers to the abbreviation of multi-level neural network,which mainly implements a series of functions such as signal input,feature value extraction,and weight value calculation.Compared with the traditional super-resolution method that uses a deep network for super-resolution,the effect of the MR image after super-resolution is not good.Based on the in-depth study of super-resolution deep networks,this paper combines the mechanism of MR images to study the sample generation methods of MR high-and lowresolution images.Aiming at MR images with different contents of water,a deep network is designed to super-resolve MR images.For the three entry points of combining MR characteristics,improving objective indicators,and improving network structure,the optimization method design of super-resolution networks is completed.The superresolution results of various algorithms and optimization methods are mainly evaluated from the objective indicators and visualization results of super-resolution.Finally,the algorithm is integrated into the MR image super-resolution test platform.Aiming at the problem that it is difficult to obtain high-quality,high-and lowresolution samples,this article first analyzes the generation mechanism of MR images,and selects 4 physiological parts that combine different water as the sample set for superresolution.Combining the image degradation mechanism and the MR image imaging mechanism,five high-and low-resolution sample generation methods are used to obtain low-resolution samples,including bicubic downsampling,HR plus noise downsampling,LR plus noise downsampling,and K-space truncated downsampling,The K-space added to the mask is truncated and down-sampling.The results of the study show that after using the traditional bicubic super-resolution of the low-resolution MR images obtained by five kinds of downsampling,the effect of the bicubic downsampling is better than other methods.Therefore,this method is selected as the method of constructing high-and lowresolution samples of prior knowledge.Aiming at the problem of network design for MR image super-resolution,this paper designs seven super-resolution networks for MR image.Contains three types of deep networks: feedforward deep networks,feedback deep networks,and confrontational generation networks.Super-resolution is performed on 4 parts with different binding water.The research results show that there is no obvious jump in the performance of MR image super-resolution effect and conventional image super-resolution effect in terms of objective indicators,and the subtle fluctuations are mainly caused by the abundance of samples.Different organs of the same network have different super-resolution effects,and there are some networks that are not obvious in one of the physiological parts,but have jump changes in the other part,which shows that different physiological organ parts can be customized to design super-resolution networks.From the comparative analysis of different networks in the same part,ESRFBN is superior to other networks in terms of PSNR and SSIM indicators.This mainly relies on the multi-feedback mechanism of the network with weight sharing,so that it can respond to each physiologically in the superresolution process.The parts have strong adaptability.Regarding the optimization of MR image super-resolution network and the construction of MR image super-resolution test platform,this article mainly optimizes from three aspects: combined with the MR image characteristic mechanism,the continuous slice multi-channel input method is adopted,and the purpose is to use the characteristics of adjacent slices to Make up for the feature details lost due to artifacts in the MR image imaging process,making the super-resolution MR image closer to the original high-resolution MR image,and the objective index improvement effect is significant.The optimization method of joint loss function is used to optimize the objective index,which improves the objective index.The three fusion methods of PWF,PNF,and CNF are used to optimize the network.PWF has obvious performance in improving the texture characteristics of MR images after super-resolution,PNF has significant performance in improving objective indicators,and CNF is both in MR image texture richness and objective indicators There are advantages.This article embeds the sample generation method,the traditional super-resolution algorithm,and the deep superresolution network into the MR test platform.This paper tests the functions of the software,and the test results show that the super-resolution network in this paper can effectively improve the super-resolution effect during the super-resolution task.