Research On MR Image Super Resolution Algorithm Based On Depth Residual Network

Magnetic resonance imaging(MRI)is a radiation-free and minimally invasive medical imaging technology.The medical images scanned by this technology have better quality and clearer texture details than those obtained by other methods,so they are widely used in clinical diagnosis.However,in practical application,the resolution of magnetic resonance(MR)image is affected by many interference factors such as long scanning time,unstable patient posture,insufficient hardware conditions,so it is difficult to generate high-quality MR image.Super resolution technology can improve the image resolution without upgrading the hardware cost,which is the most cost-effective method to improve the quality of MR images.However,super resolution technology based on deep learning can reconstruct MR images with clear texture and good quality,which has significant advantages compared with other super resolution technologies.At present,there are two problems in using deep learning method for super-resolution reconstruction of MR images: on the one hand,how to extract more features from MR images and retain more original image information in the extraction process;On the other hand,how to make the model pay more attention to the texture details in MR image during training,so that the reconstructed MR image will have better effect.Therefore,the main research work of this paper as follows:(1)In view of the shortcomings of the existing image feature extraction methods,such as weak anti-interference ability,poor processing effect and low universality,this paper proposes a deep residual network model to improve the resolution of MR images.This model adopts the method of dense connection structure,improving dense convolution block and adding feature fusion layer,which simplifies network parameters and reduces resource consumption to strengthen the feature extraction ability of network model.In addition,the residual module is modified and a group normalization layer is introduced,so that the model can retain more original MR image information while deepening the network layer.At the end of the model,subpixel layer is used to complete the up-sampling work,which improves the image edge smoothness caused by interpolation amplification method.(2)in view of the existing super-resolution algorithm based on the deep learning in training didn’t pay attention to details such as texture,edge area,leading to poor quality of the output image problems,this paper proposes a super-resolution network model based on attention mechanism,channel this model to focus attention module integration module and space attention as a kind of fusion module,And this module is introduced into the improved residual network,so that the network model will allocate more weight to the detail area of MR image during training,so as to extract more valuable texture information and reconstruct the high-resolution MR image with better quality.At the end of the model,transpose convolution layer is used to complete the up-sampling work,which avoids the problem of adding invalid information when using subpixel layer for up-sampling.(3)In order to prove the feasibility of the two models,this paper carries out experiments on the proposed model and compares the results with those of other algorithms.The experimental results show that the network model proposed in this paper has certain advantages in performance.The reconstructed high-resolution MR images by the model have high definition,rich texture details,no obvious artifact or noise information,and the PSNR value and SSIM index have been improved to some extent.