Medical Image Registration Method Research Based On Adversarial Learning

Medical image registration has always been an important processing step in medical image processing system,which can provide aligned image data for subsequent fusion,reconstruction and other tasks.Due to the extremely complex organ structure of medical images,traditional registration methods based on features cannot satisfy the requirements of modern medical image registration for real-time deformable transformation.In the past decade,the rapid development of deep learning has provided a new solution to the problem of medical image registration.Existing medical image registration methods based on deep learning mostly use deep network to predict unidirectional registration results,which cannot maintain the topological structure of the original image.Some methods calculate the inverse transform similarity of the bidirectional deformation field to satisfy the inverse consistency,but the result image is usually obtained from the registration image according to the interpolation and sampling of the deformation field,so this kind of deformation method cannot make the result completely maintain the inverse consistency.Most of the methods only use the similarity measure as the loss function,resulting in the loss of some structural features of the resulting image.Aiming at the above problems existing in the medical image registration method based on deep learning,this paper proposes new methods in the aspects of network structure,cost function,etc.The main work is summarized as follows.Firstly,we propose a new medical image registration method CGMorph(Cycle Generative Adversarial Morph),which use double networks to generate bidirectional registration transformation,then design loss functions based on inverse consistency and smooth regular term.It builds inverse consistent registration network oriented by consequence.Besides,we introduce adversarial mechanism to generate results with similar feature to the training set,which can preserve its original structure feature distribution.It will lead to smooth and reciprocally inverse transformations,which are also diffeomorphic transformations.Diffeomorphic transformations possess bijection property that will guarantee voxels matched one-to-one between moving image and fixed image.This can efficiently reduce the hole and folding in generated images to reach the goal of raising accuracy.Secondly,we propose a new image registration method Non-local CGMorph using non-local adversarial network based on CGMorph.This method designs non-local generator suitable for image registration,which combines the global feature for the response at each position during forward learning procedure by insert non-local block in generators.It makes for a large receptive field,which can get correct registration when there is large deformation and obtain more accurate registration results.Then,in the inverse consistency loss of CGMorph,it is necessary to be guaranteed that two cycle images are the same as their corresponding input images.Thus,the differences between cycle images should be similar with the input images.We propose a difference consistency bound term for loss function to promote the convergence of algorithm and enhance the inverse consistency of network.Besides,CGMorph constructs result-oriented inverse consistent algorithm,and it is also an important aspect that the deformation fields are inverse consistent.We propose deformation inverse bound term of loss function to guarantee the inverse consistency of deformation.In conclusion,in terms of the existing problems in image registration methods,this thesis proposes a new image registration method based on adversarial learning and an image registration method based on non-local adversarial network,and prove the superiority of our method with experiments.This method provides a new idea for image registration,and promote the application of unsupervised adversarial learning in image registration area.