Research On Fast Nuclear Magnetic Resonance Diffusion Imaging Algorithm Based On Deep Learning

Diffusion magnetic resonance imaging(d MRI)is a widely used in-vivo imaging technique to explore the information of micro-structure of biological tissue by probing the diffusion of water molecules.The long imaging time and susceptibility to motion,however,result in low signal-to-noise ratio and patient discomfort.Therefore,the compressed sensing(CS)technique,which can achieve faster imaging,was introduced to shorten the imaging time by recovering the fully sampled data from undersampled data,and it has been widely used in MRI reconstruction.Plenty of traditional reconstruction methods,based on compressed sensing,have been proposed to accelerate the speed of d MRI reconstruction,but the disadvantages of longer reconstruction time and the poor image quality in these studies do not meet the requirements of clinical applications.Accordingly,this work was designed to improve the efficiency and accuracy of CS reconstruction algorithm for under-sampled d MRI using deep learning models.The main contents of this work are as follows:(1)We propose a CS reconstruction algorithm based on generative adversarial network for reconstructing high-quality diffusion tensor imaging,named CSGAN.The proposed method achieves Nash equilibrium through the optimization process of the minimax game,so that the generator learns the full-sampled data distribution from the under-sampled data distribution for the purpose of reconstruction.To improve the quality of the reconstructed diffusion-weighted images,CSGAN embeds an SE module to add content loss for the spatial domain images and k-space signals in the adversarial loss.The experimental results show that the model is able to recover high-quality diffusion-weighted images of adult and infant ex-vivo heart with high performance in terms of reconstructed diffusion tensor images,FA and MD,and the fiber orientations.(2)This thesis presents a deep convolutional neural network using adaptive multidomain fusion strategy for high angular resolution diffusion imaging(HARDI)reconstruction,called SKQNet.The proposed method constructs SDNet for spatial domain image reconstruction,KDNet for k-space signal reconstruction and KSD fusion network for multi-domain adaptively fusion.The interrelationship between the diffusion-weighted images is also introduced to improve reconstruction quality.The experimental results show that the model can reconstruct the HARDI of the brain with higher quality,better reconstruction performance in diffusion properties GFA,NQA and ODF,as well as better performance of fiber tracking of brain compared with other prevalent image reconstruction algorithms.