Research On End-to-end Multi-dimensional Human Brain Image Registration Technology Based On Convolutional Neural Network

Medical image registration is an essential task in medical image analysis.It is widely used in pathology tracking,tissue structure discrepancy analysis,surgical navigation and so on.Although traditional iterative registration algorithms have welldeveloped,its iterative optimization process makes it time-consuming and easy to fall into local minima.It cannot satisfy the real-time and accuracy requirements of clinical applications.With the great success of deep learning in the field of computer vision,using deep learning to perform medical image registration has become a popular research topic.At present,the existed registration models based on deep learning mainly focus on nonlinear transformations,the end-to-end registration models,including both linear and nonlinear transformations,are rarely reported.In addition,almost all the registration methods are for low dimensional images,for fourdimensional images such as diffusion tensor images have not yet reported.This work intends to investigate the deep learning modes for end-to-end multi-dimensional human brain image registrations,detailed as follows:(1)An end-to-end and unsupervised registration network based on deep learning to register three-dimensional brain magnetic resonance images is proposed.It includes two major parts: linear and nonlinear transformation parameters prediction models,which achieves the coarse-to-precise registration of MRI.The results show that even if the existed methods use linearly aligned images as input,the proposed method takes the original image as input,our method still outperforms the others.It verifies that the superiority of the proposed model.(2)The end-to-end registration model is extended for four-dimensional diffusion tensor image registration.Replacing the output of nonlinear transformation prediction model as velocity fields,and the velocity fields integration module is also included.Moreover,the tensor reorientation module to redirect the transformed diffusion tensor images is also designed.The results show that the proposed method is comparable to that of state-of-the-art method,moreover,it is faster and more stable.