Study Of Medical Image Super Resolution Method Based On Convolutional Neural Network

In recent years,with the rapid development of medical imaging technology,medical images have become an important reference for doctors to judge the patients’ condition and find the focus of infection in the clinical diagnosis process.However,influenced by equipment accuracy,rodation dose,imaging time and motion displacement,it’s often difficult to get high-resolution and high-quality medical images in practice.The super-resolution reconstruction algorithm is an effective method to improve the image quality which needs not come at the expense of changing image acquisition conditions.With the gradual maturity of deep learning technologies,many scholars have proposed super-resolution reconstruction algorithms based on deep convolutional neural network,which can automatically learn the mapping relations between high-resolution and low-resolution images through training and apply it to the reconstruction process,thus effectively improve the accuracy of image reconstruction.However,most of the current algorithms improve the performance of the network by increasing the size and complexity of the super-resolution network,without fully considering about their application scenarios.Therefore,combined with the specific medical image reconstruction problem,this thesis studies the super-resolution reconstruction algorithms based on convolutional neural network.The main research work is as follows:(1)Aiming at the super-resolution of 2D medical images,a super-resolution network based on multi-scale residual module,channel attention mechanism and multi-level features fusion module is proposed.Taking the original low-resolution image as the input directly,the image amplification is realized via the sub-pixel convolutional module at the end of the network,which not only reduces the time complexity of the algorithm,but also avoids the error introduced by the interpolation algorithm.The global residual connection method is also used to reduce the difficulty of network training and improve the accuracy of network reconstruction.The main part of the network is constructed by several multi-scale residual modules combined with channel attention mechanism,which not only reduces the number of parameters,but also enhances the ability of feature extraction.What’s more,combined with the characteristics of super-resolution reconstruction problem,a multi-level features fusion module is designed in the network to fuse the information with different levels of abstraction for the final image reconstruction.(2)Aiming at the super-resolution of 3D medical images,a 2D super-resolution network based on multi-channel input,depth separable convolution and global residual connection is proposed.The algorithm decomposes the 3D reconstruction problem into several 2D images’ reconstruction in different directions,and then merges the reconstruction results in different directions to get 3D super-resolution results.First,in order to make full use of the spatial information of the 3D data,we input the adjacent channel images of the image to be reconstructed into the network together to extract the features.Then,Considering that the data amount and calculation of 3D medical images are larger than 2D image,the speed of network is the key of algorithm.Therefore,in the algorithm,the original low-resolution image is taken as the input directly,and the image is enlarged by deconvolution layer at the end of the network.And the depth separable convolution is used to construct the nonlinear mapping element of the network,which can effectively reduce the calculation amount without sacrificing the reconstruction accuracy.In addation,global residual connection is employed in the network to reduce the difficulty of image information restoration.We also design a series of comparative experiments to verify the advanced nature of the above two algorithms.The final results show that,compared with the existing networks,the super-resolution network designed for two-dimensional medical images has a greater improvement in reconstruction accuracy,while the network designed for three-dimensional medical images has better optimization in reconstruction accuracy and operation efficiency.The above results further illustrate that the algorithm designed in this thesis has good practical application value.