The Research Of Breast Cancer MRI Super-resolution Method Based On Deep Residual Structure

Nuclear magnetic equipment is widely used in clinical diagnosis because of its characteristics of no ionizing radiation and three-dimensional imaging.However,the early breast cancer lesions are relatively small in the whole MR image,which will affect the accuracy of early breast cancer screening.The image super-resolution technology based on deep learning can realize the reconstruction of high-resolution image without improving the hardware equipment,but the existing super-resolution networks are mostly used in the field of natural image super-resolution.Compared with natural images,MR images have no distinct color information,lack of diversity of image information,single image structure and more complex texture features.Therefore,aiming at this problem,we propose a super-resolution neural network based on depth residual structure.Our works can be summarized as:(1)Aiming at the shortcomings of high tissue similarity and less image information in breast cancer MRI images,we propose a chain super-resolution network based on residual structure.The main body of the network is stacked by several residual modules.The residual module can promote the information fusion between layers,enhance the reuse of different features in the network,and help the network to learn the detailed information in MR images.However,due to the complexity and variety of detail information in the MR images,in order to further improve the restoration of the details of the MR image by the network,the way of broadening the network width is adopted to increase the extraction of image features at each layer of the network.However,while increasing the amount of calculation,the too wide network layer may also cause the phenomenon of extracting information redundancy.Therefore,the design of the number of channels in the network layer adopts a gradual change from large to small.The proposed network comprehensively considers the influence of network depth and network width on the reconstruction performance.While ensuring the network depth,the width of the network is widened,so as to extract as much useful feature information from the image as possible.(2)Although the feature extraction ability of the network is improved by broadening the network width,the single chain network structure makes the features between the shallow network and the deep network unable to be transmitted effectively,and the utilization rate of features is not high.Therefore,a super-resolution neural network based on U-net network structure is proposed.The main body of the network uses the improved cross residual module for feature extraction,and then concatenate the output of the shallow network with its corresponding deep network layer through the long connection structure.The features extracted by shallow network may be affected by the noise in the original image,so that there is harmful information in the extracted features that affects the reconstruction results.In order to solve this problem,the channel attention module is introduced into the long connection part of the network.The module is a lightweight network structure.With only a small amount of computation,it can give weight to the information of different channels according to the importance of information,so as to highlight the impact of useful features on the reconstruction effect.The two proposed networks are tested on self built data sets and public data sets,and the reconstruction effect is comprehensively evaluated by using subjective evaluation indexes and objective evaluation indexes.By comparing with different reconstruction methods,it is proved that the proposed method can achieve better reconstruction effect.