Unsupervised Single Mode Image Registration Based On Deep Learning

Image Registration refers to finding the spatial mapping relation of the image pair so that the coordinates of the image pair are consistent in the same space.As a key technology of computer-aided diagnosis and treatment,medical image registration can prevent doctors from missing diagnosis or misdiagnosis when they are overtired,and has been developed rapidly.Deep learning has achieved remarkable accomplishments in the medical image registration field,as well as making tremendous advances in image processing in recent years.Based on this,the paper carries out research on unsupervised single-mode registration based on deep learning,mainly including:(1)An unsupervised single-modal image registration method based on improved U-Net is proposed.Most of the existing registration methods use the U-Net network structure for feature extraction and reconstruction.However,the original U-Net uses a long connection at the connection codec,which will result in two connected features due to the sampling depth.The gap causes a large semantic gap,which in turn affects the subsequent registration accuracy.The method in this paper improves the connection method of U-Net,deploys more short connection methods between codecs,and overcomes the problem of large semantic gap.At the same time,a channel attention mechanism is designed on the decoder of the U-shaped network,and the anti-interference ability of the registration model is improved through feature weight recalibration,thereby improving the registration effect.(2)An unsupervised single mode image registration method based on cyclic consistency is proposed.When the existing registration methods use deformation field to distort floating images,there are still problems in maintaining the topological structure,which affects the authenticity of subsequent registration results.In this paper,based on the registration method proposed in(1),a cyclic consistency constraint is designed between the original image and the re-deformed image.The two deformed images are used as the input of the network again,and the original image is obtained after two deformations,which forms a different norm from the input image.The generation of the deformation field is supervised by this function.The topological structure of the deformed image is not changed and the authenticity of the deformed image is guaranteed.In this paper,a large number of experiments are carried out on OASIS and ADNI data sets,and the proposed method is compared with the outstanding registration algorithms so far.The experimental analysis is carried out from the visual and quantitative indexes,which verifies the feasibility of the proposed method,and proves that the registration accuracy and result authenticity of the proposed method are significantly improved compared with the comparison algorithms.