Deformable Medical Image Registration Algorithm Based On Convolutional Neural Network

Medical image registration is a basic topic in medical image analysis,which has important scientific and clinical application values.Through medical image registration,doctors can directly compare cross modal images or deformed single modal images to determine the changes in the lesions to make accurate diagnosis timely.Traditional medical image registration algorithms based on iteration treat registration as an optimization problem,and each registration requires a new iteration,which is time-consuming and depends on the selection of image features and similarity measures.In recent years,registration algorithms based on deep learning have emerged,mostly surpassing traditional algorithms in terms of registration accuracy and speed.Among them,unsupervised learning registration algorithms are more suitable for less labeled or unlabeled medical image data.According to the transformation model,medical image registration can be divided into affine registration and deformable registration.The former uses an affine transformation model,which is commonly used to handle the registration of rigid organs,while the latter uses a nonlinear elastic transformation model,which is commonly used to solve the registration of elastic deformation of tissues or organs,and has a wider range of applications.In addition,the research on affine registration has matured,but there are still many difficulties in deformable registration,such as accurately learning many parameters of nonlinear transformation models and maintaining alignment of topological structures such as boundaries.Most of the existing deformable registration algorithms roughly process the extracted image features,ignoring boundary information,resulting in low overall model registration accuracy and poor boundary registration results.In response to these questions,this thesis has made the following work.Firstly,to enhance the ability of feature extraction and generate deformation fields with high accuracy,proposed a medical image registration algorithm based on spatial skip convolutional network.The spatial skip convolutional network consist of feature extractor,spatial feature enhancement connection,and deformation field generator.The spatial feature enhancement connection is embedded with the central feature enhancement attention module designed in this paper,which is used to weight and enhance the central features of the image.Subsequently,to improve the effect of boundary registration,proposed a medical image registration algorithm based on boundary optimization,which added a boundary registration constraint to constrain the spatial misalignment produced by the registration boundaries of the deformation fields generated by the spatial skip convolutional networks,and designed a boundary loss function as the optimization goal of boundary registration.Finally,conducted some ablation experiments on the open brain MRI dataset for the above two algorithms,and comparative experiments with some classical traditional algorithms and deep learning algorithms.The experimental results show that the medical image registration algorithm based on spatial skip convolutional networks has high registration accuracy for brain MRI images,and the registration algorithm based on boundary optimization has excellent results in boundary registration.