Research On Registration And Generation Methods For 3T Magnetic Resonance Images

Magnetic Resonance Imaging(MRI)is an important medical imaging technique,and its quality directly affects its application in clinical diagnosis and treatment.Highfield 3T MRI has advantages over 1.5T MRI,such as higher resolution and better tissue contrast,and is expected to improve the accuracy of early disease diagnosis.However,the 3T magnetic resonance equipment is expensive and has not been widely used in clinical and research.Currently,the 1.5T medical magnetic resonance equipment is still widely used in China,especially in some county hospitals.For this case,a deep learning-based method is proposed to generate approximate 3T MRI from 1.5T MRI,which is proposed to improve the resolution and contrast of MR images.The main research content is as follows:(1)Obtain pairs of 1.5T and 3T MRI of subjects.Due to the differences in scanning equipment and the positional differences of subjects during different scanning times,the spatial positions of each pair of MRIs do not correspond.Therefore,we need to perform medical image registration.In this paper,DA-VM,an MRI cross-modal registration network based on Convolutional Neural Networks,is proposed to correspond 1.5T and3 T MRI in spatial positions.The network is improved based on Voxelmorph.Firstly,some down-sampling layers are replaced by dilated convolutions to expand the receptive field of the network and enhance the expression ability of the feature map.Secondly,by introducing a dual attention mechanism after the feature maps,attention is weighted separately for the spatial and channel dimensions of the image feature maps,adaptively fusing local feature and global dependency,reducing or eliminating areas and channels involving noise information in the feature map,and strengthening the feature representation of the regions of interest,which helps improve the accuracy of registration.The experimental results show that the DA-VM network improves the registration quality and detail compared to other mainstream algorithms.After registration,the 3D MRI is sliced into 2D images and the dataset is expanded by data enhancement,we make a standard database of 1.5T-3T 2D magnetic resonance images.(2)With 1.5T-3T Standard Database,Self-Attention-Cycle GAN(SA-Cycle GAN)method,based on self-attention mechanism fused generative adversarial network,is proposed to generate approximate 3T MR images from 1.5T MR images.The selfattention mechanism is used to calculate the weight parameters of feature maps,and the spectral normalization is used to reduce function oscillation and accelerate model convergence.Introducing prior information into loss function to improve the authenticity of the generated images.The experimental results show that the MR images generated by SA-Cycle GAN have higher PSNR and SSIM values than similar comparison methods.In different MRI sections of the dataset,the PSNR is increased by16.3% ～ 27.4%,and the SSIM is increased by 7.8% ～ 27.4%.The approximate 3T MRI generated by this method has richer details and more accurate texture restoration,and the effect is significant.Based on the above algorithm research,a 3T MRI image generation system based on SA-Cycle GAN is developed.The system is built based on Django and can realize the function of 3T MRI image generation.The methods studied in this paper indicate that the registration algorithm and MRI generation algorithm have excellent performance.Introducing attention mechanism and improving network models can enhance the network performance,and generate approximate 3T magnetic resonance images with higher resolution and richer clinical information,which has significant clinical application value.