Nonrigid Medical Image Registration Based On Generative Adversarial Networks

Medical image registration is the premise of medical image fusion and analysis.The spatial consistency of anatomical structure is realized by constructing the optimal spatial transformation between two or more medical images.Medical image registration has important clinical significance in accurate medical aspects such as tumor detection,focus location,surgical navigation and so on.With the increasing requirements of modern medicine for the quality and scale of medical images,the development of medical images is becoming more complex and diverse,which makes it difficult for traditional registration methods to meet the needs of high efficiency,high precision and robustness in clinical medicine.With the rapid development of computer technology,medical image processing based on deep learning technology has become the mainstream.Although the registration based on deep learning has been greatly improved in speed,accuracy and robustness,there are still some problems and difficulties in the non-rigid registration of complex deformation.The registration based on deep learning is divided into supervised method and unsupervised method,in which the supervised method requires a large number of label deformation field,and the label deformation field depends too much on manual,which leads to the uneven quality of the label deformation field.Therefore,based on the generated countermeasure network,this paper constructs a kind of unsupervised non-rigid registration method,uses the depth regression network to generate the alignment deformation field directly,and realizes accurate and fast end-to-end registration.The feasibility and generalization ability of the proposed method are verified on multiple data sets.The details are as follows:(1)A pre-registration model is constructed by combining traditional iterative registration and deep learning techniques.The affine transformation and B-spline function are used to preregister the registration image pairs,which reduces the influence of some rigid deformation and non-rigid deformation on the registration to a certain extent,and improves the deformation registration ability of the whole registration model.The addition of pre-registration reduces the burden of network learning in the later stage,and accelerates the speed of network learning without losing the accuracy of registration.Through the non-rigid registration experiment,the feasibility of the pre-registration model is proved.(2)Deformable convolution is embedded in the generation network based on Unet architecture.Deformable convolution is added at the end of each convolution module,that is,before each down sampling,and jump connection to the upper sampling stage to encode the deformation features of different scales,which improves the ability of the network to extract deformation features and realizes the complex nonlinear mapping between the deformation field and the registered image.Single-mode and multi-mode registration are carried out on multiple data sets,and the feasibility of the proposed method is proved.(3)Aiming at global deformation and local deformation,a dual-channel registration network is proposed,which integrates global deformation and local deformation.The registration network based on Unet architecture is adopted in global registration,and the densely connected network is used in local registration,which realizes deformation coding of different sizes and improves the deformation registration ability of the model.A large number of experiments are carried out on the LPBA40 data set to compare the four excellent registration methods so far,to verify the performance of the dual-channel registration method,and to verify the generalization ability of the dual-channel registration method on the IBSR data set.