Super-resolution Reconstruction Of Magnetic Resonance Image Based On Recursive Residual Network

Magnetic resonance imaging(MRI)is a widely used medical imaging technology.It has the characteristics of clear soft tissue imaging,which is conducive to doctors for diagnosis.In order to acquire high-resolution images,long-time scanning is required.But it is difficult for patients to maintain still during the long-time scanning,which will result that the acquired image has motion artifacts.In order to reduce motion artifacts,the layer thickness can be increased to shorten the scanning time.However,it also results in low resolutions of the required images which is not conducive to doctors for diagnosis.Thus,it is very important to improve the resolution of MR images.In order to solve the problem of magnetic resonance images super-resolution,this thesis improves the existing deep recursive residual network(DRRN)successfully which is applied for natural images.Some inherent problems existed in the DRRN will be solved in this thesis,which are serious memory consumption and time-consuming training.The main work of this thesis is described as follows:(1)The DRRN costs a lot of memory and time during training.Due to the limited computing resources,a faster deep recursive residual network(FDRRN)is proposed for MR images super-resolution.Experiments show that the FDRRN greatly reduces the computational burden and the occupation of graphical processing unit(GPU)memory.The required GPU memory for the FDRRN is only 1/3 of that for the DRRN,which accelerates the speed of network training.Its training time is only 1/4 of the DRRN.At the same time,the PSNR and SSIM values achieved by the FDRRN are slightly higher than the those achieved by the DRRN.(2)During the training of the DRRN,some information may be lost when the information is transmitted via the residual units.To solve this problem,a multi-resolution learning convolutional neural network(MRLCNN)is proposed,which contains residual units for feature extraction,multi-resolution up-sampling deconvolution layer and multi-resolution learning layer.The MRLCNN is composed of 17 convolutional layers and one deconvolutional layer.The network realizes image super-resolution in the low-resolution image space.Multiresolution upsampling is used to achieve information fusion of multiple residual units and to accelerate the network.Multi-resolution learning can adaptively determine the contributions of upsampled high-resolution feature maps at various resolutions to super-resolution reconstruction of MR images.Experiments show that the MRLCNN is superior to the FDRRN in terms of training time and GPU memory consumption.Also,it achieves more excellent super-resolution reconstruction performance for MR images compared to the latest deep learning methods.