Isotropic Super-resolution Reconstruction Of 3D-MRI Images Based On Deep Learning

Magnetic Resonance Imaging（MRI）is an imaging technology used for imaging human tissues and organs.Because of its non-ionizing radiation damage,high contrast in soft tissue imaging and multi-parameter imaging,it has rapidly developed into an important application technology in biomedicine.Spatial resolution is one of the key imaging parameters of magnetic resonance imaging.High-resolution MRI images provide rich structural information for human organ and tissue imaging,which is beneficial for accurate clinical diagnosis and subsequent image processing tasks.However,the resolution of MRI images is limited by various factors such as hardware configuration,scanning time,and signal-to-noise ratio.The improvement of spatial resolution usually comes at the expense of reducing the signal-to-noise ratio or increasing the scanning time.Therefore,under the premise of reasonable image signal-to-noise ratio and MRI scan time,improving the resolution of MRI images is an urgent problem to be solved.An efficient solution is to apply super-resolution reconstruction techniques to MRI images.Super-resolution reconstruction technology aims to extract image feature information from a single or a group of low-resolution images,and reconstruct highresolution images that meet the requirements.Among them,deep learning-based superresolution reconstruction methods are widely used in MRI image super-resolution reconstruction tasks due to their effectiveness in learning high-resolution image details.Aiming at the problems of large layer thickness and low resolution in the selection direction of MRI multi-slice imaging technology,this proposes a research on isotropic super-resolution reconstruction of 3D-MRI images based on deep learning.The purpose is to improve the inter-slice resolution of low-resolution images,reduce the partial volume effect of images,and reconstruct high-resolution isotropic 3D-MRI images.The main work of this as follows:（1）First,research and implement the super-resolution reconstruction of multiple 3DMRI images.Currently,most deep learning algorithms input a single image for reconstruction to improve intra-slice or inter-slice resolution.Compared with a single image,multiple images have more image detail information.Based on the above,this modifies and extends the SRCNN-3D algorithm,and proposes a multi-image superresolution reconstruction network model MISRnet,which uses an end-to-end architecture and uses three orthogonal MRI thick-slice scans in each direction.Anisotropic voxel information to reconstruct isotropic high-resolution images.Experimental results show that the network trained with three orthogonal MRI scans has better performance than the network trained with only one or two MRI scans.（2）On the basis of the research on super-resolution reconstruction of multiple 3DMRI images,an isotropic super-resolution reconstruction of 3D-MRI images based on RMISRnet is proposed.RMISRnet uses residual learning to spread image feature information and gradient information across layers,which is conducive to network training and convergence,and fully extracts three orthogonal low-resolution image features,so that the reconstructed image has more high-frequency information.The results of RMISRnet network reconstruction are compared and analyzed with those of other super-resolution reconstruction algorithms.The results show that RMISRnet has better results in reconstructing normal 3D-MRI images and 3D-MRI images with lesions,and the reconstruction performance is better than other reconstructions.algorithm.In large-scale super-resolution reconstruction,the reconstructed image still has good edge information and high-frequency details,and has good robustness.