Medical Image Non-rigid Registration Method And System Research

With the development of medicine and computer science,clinical medicine calls for more advanced image processing techniques.As a crucial technology,medical image registration is of great significance with a wide range of clinical applications,such as diagnosis of diseases,monitoring the morphological change of lesions,image-guided radiation therapy and evaluation of treatment.The purpose of image registration is to align the underlying anatomical position represented by each voxel in a given floating image with the corresponding position in the reference image,that is,to find the optimal spatial transformation relationship between the two images so that the points with the same anatomical significance in the two images are in the same spatial position.At present,diseases in lungs have serious impacts on human life and health.Image registration can be used to estimate the movement and deformation of lung organs,assisting doctors in diagnosis and treatment of lung diseases.However,the respiratory-induced large deformation makes registration of lung images rather difficult and challenging.Since many often-used registration methods hardly generate reasonable results,developing more accurate methods is urgently needed for registration of lung images.In this paper,the non-rigid registration method of medical images is studied for the registration of three-dimensional CT images of the lungs.The main work is as follows:（1）A registration algorithm of B-spline free form deformation based on compressible flow is proposed.According to the compressible characteristics of lung organs during respiratory cycle,intensity-variation between CT images and the smoothness constraints of deformation field,the compressible flow theory is used for data fidelity that is combined with squared₂norm of displacement vectors as regular for similarity measure.Then a multi-scale framework of B-spline free form deformation model is designed for the registration from coarse to fine.（2）An end-to-end unsupervised deformable registration method based on convolutional neural network is proposed.After analyzing the advantages and disadvantages of supervised and unsupervised learning methods of registration,a deep convolutional neural network is constructed based on the encoder and decoder models,in which the compressible flow is taken as the lost function,so that the spatial correspondence is learned and lung CT images are matched quickly.（3）A medical image registration and evaluation system is established,in which there are some classical registration methods including the Demons based algorithms,optical flow approach and the proposed B-spline free form deformation algorithm based on compressible flow,as well as the registration method based on convolutional neural network.The system allows for displaying 3D images in different views before and after registration,selecting different registration methods,and estimating registration results both qualitatively and quantitative.