Multi Feature Extraction SYMNet Brain MRI Image Registration Based On Improved Skip Connection And Local Regularization Term

The brain is an important organ in the human body.Magnetic resonance imaging(Magnetic Resonance Imaging,MRI)can clearly visualize the soft tissue components of the brain.The research on 3D(Three Dimensions,3D)brain MRI image registration algorithm is of great value for clinical application.Traditional methods have high accuracy,but the algorithm complexity is high,the registration time is long and the universality is poor.Deep learning methods have high precision and strong real-time performance,but most of them ignore the differential homeomorphism of registration.The registration algorithm based on Symmetric Diffeomorphic Neural Network(Symmetric Diffeomorphic Neural Network,SYMNet)ensures the diffeomorphic from two aspects.At the same time,the unsupervised learning method can avoid the difficulty of data label labeling,but the algorithm still has the following three shortcomings: the coding method can not pay attention to the important features of the image and the extracted features are less,resulting in low registration accuracy;The local regular term ignores the zero singularity,which leads to poor differential homeomorphism;Skip connection directly combines the feature maps with large differences to reduce the registration accuracy.Aiming at the problem that the coding method can not pay attention to the important features of the image and the low registration accuracy caused by less extracted features,this paper proposes a SYMNet(Multi-Feature Extraction SYMNet,M-SYMNet)registration algorithm based on multi feature extraction,which introduces maximum pooling into the original coding channel and adds a parallel coding channel composed of maximum pooling and void convolution.Aiming at the problem of poor differential homeomorphism caused by ignoring zero singularities in local regular terms,an improved local regular term for multi feature extraction SYMNet(Improved Local regular term for Multi Feature Extraction SYMNet,LM-SYMNet)registration algorithm based on improved local regular terms is proposed.On the basis of the original local regular terms,the zero points in the deformation field are added to the penalty.Aiming at the problem that skip connection directly combines different feature maps to reduce the registration accuracy,this paper proposes an improved skip connections and local regular term for multi feature extraction SYMNet(Improved Skip connections and Local regular term for Multi Feature Extraction,SLM-SYMNet)registration algorithm based on improved jump connection and local regular term,which uses the interval filling module CCLN to insert skip connection to combine the deep feature map and shallow feature map.Experimental results show that the multi-feature extraction coding channel in M-SYMNET improves the registration accuracy while improving the registration efficiency.The improved local regular term in LM-SYMNET improves the registration accuracy and differential homeomorphism.The improved skip connection in SLM-SYMNET can not only improve the registration accuracy and differential homeomorphism,but also improve the registration efficiency.