Super-resolution Reconstruction Of Magnetic Resonance Images From Anisotropy To Isotropy Based On Deep Learning

High-resolution 3D MR images provide rich anatomical and detail information for clinical diagnosis and medical image research.However,clinical MRI acquisition protocols often need to balance factors such as image signal-to-noise ratio,voxel size(and spatial resolution),and acquisition time.These factors are interrelated and constrained.With an increase in voxel size,the signal-to-noise ratio increases,and the partial volume effect is enhanced.Acquiring higher resolution 3D MR images requires a longer scan time.To balance these factors,anisotropic voxels are usually selected when acquiring MR signals.Anisotropic voxels result in low-resolution 3D MR images in the slice selection direction,lacking high-frequency information,which could negatively impact medical image research and clinical diagnosis.Therefore,improving the resolution of 3D MR images under reasonable image signal-to-noise ratio and scan time is an urgent challenge that needs to be addressed.Currently,the most direct and effective solution is to use image post-processing to apply super-resolution reconstruction technology to MR images.Therefore,this article focuses on using deep-learning-based super-resolution reconstruction algorithms to reconstruct anisotropic 3D MR images with multiple scanning directions orthogonal.The main work of this article is as follows:(1)Research on the super-resolution reconstruction of anisotropic 3D MR images with different thicknesses,using different numbers and scanning directions(sagittal,axial,coronal).Multiple orthogonal anisotropic 3D MR images contain more high-frequency information than single images,and these high-frequency information are complementary.Therefore,this article proposes the ASCnet super-resolution reconstruction network model.ASCnet is an end-to-end network architecture based on residual learning,which reconstructs multiple orthogonal anisotropic 3D MR images as a single isotropic 3D MR image.The experimental results show that ASCnet can effectively extract and fuse structural information and complementary high-frequency information from multiple orthogonal imaging planes of anisotropic 3D MR images,eliminate partial volume effects within the imaging plane,restore the resolution in the slice selection direction,and reconstruct high-resolution isotropic 3D MR images.(2)To further improve the super-resolution reconstruction performance of multiple orthogonal anisotropic 3D MR images,ASCnet is optimized,and the ASCRDN superresolution reconstruction network model capable of multi-scale reconstruction is proposed.ASCRDN introduces residual connections and dense connections,which are beneficial for feature propagation,preventing the degradation of the network model,reducing the extraction of redundant features,and improving the non-linear expression ability of the neural network while reducing the computational cost of the 3D convolutional neural network.This enables the reconstructed isotropic 3D MR images to provide more anatomical and texture details.The experimental results show that the super-resolution reconstruction effect of ASCRDN is better than that of ASCnet and other super-resolution reconstruction algorithms,and can achieve multi-scale and multi-modal super-resolution reconstruction of 3D MR images from multiple orthogonal imaging planes.