Super-Resolution Reconstruction Algorithm Of 3D-MR Image Based On Deep Learning

Magnetic resonance imaging(MRI)is more and more widely used in clinical and science studies due to it has many advantages,but hardware resource limitations and time consumption make it difficult to obtain high-resolution 3D-MR images.The superresolution reconstruction algorithm based on deep learning and orthogonal multiple 3DMR data,it can increase the resolution and effect enormous improvements in image quality.In this work,researcher carried out research on super-resolution of orthogonal multiple 3D-MR data,with the aims of make full use of information of multiple orthogonal images,solve problems of number of parameters is too large in 3D neural network and insufficient high-quality training data supply,obtain high-resolution MR images and improve the quality of 3D-MR images.The main research work and conclusions of thesis are drawn as follows:(1)Thesis proposes a algorithm of calculate common information.The algorithm is implemented by combining zero-order interpolation and deep neural network technology,and it takes advantage of the fact that multiple 3D-MR images contain the same information.Experimental results show that the use of zero-order interpolation preprocessing is better than three-linear interpolation preprocessing in the case of multiimage encoder-decoder(SR8-MIED)network with a symmetric residual structure with the number of encoder-decoder structure is 8.In addition,in order to solve problems of number of parameters is too large supply in 3D neural network and insufficient highquality training data,the 2D neural network is used to realize 3D-MR super-resolution reconstruction and treat 3D-MR images as stacks of 2D slices.(2)Thesis designed the MIED basic network for algorithm of calculate common information,and then designed a deep residual multi-image encoder-decoder(DR-MIED)network with a deep residual structure by directly deepening the network layer,designed the SR-MIED network by deepening the network layer and adding hop-symmetric residual connection,designed a recurrent residual multi-image encoder-decoder(RRMIED)network with recurrent residual structure by adding recurrent connections.All three use symmetrical convolution and de-convolution structures,use convolutional layers to extract abstract features,and then restoring image details through the corresponding de-convolution layers,thereby improving the quality of the obtained superresolution images.Experimental results show that SR8-MIED network proposed in thesis can obtain clearer magnetic resonance super-resolution images.