Research Of Non-rigid Medical Image Registration Algorithm Based On Generative Adversarial Networks

Image registration is an important research topic in medical image analysis.The conventional registration methods need to optimize the objective function by reiterating to realize registration,which ignores the sharing registration mode to limits speed of registration,is easy to fall into a local optimum.Deep learning-based registration methods can effectively overcome the limitations of conventional registration methods with their special structure of local weight sharing,good fault tolerance and self-learning capabilities,but can not achieve high registration accuracy for the individual organs with non-linear deformation caused by physiological movements and large deformations,and also depend on a large amount of training data.In fact,medical images of human organs are very scarce and related labels are difficult to obtain.As a result the registration models based on limited training datasets have poor generalization ability on cross-domain datasets.To address the problems,three specific work of this paper are as follow:(1)This paper proposes a method to register multi-modal medical images by combining Residual-in-Residual Dense Block(RRDB)with Generative Adversarial Networks(GANs).Firstly,RRDB are introduced into the standard Generator Network to extract more high-level feature information from unpaired image pairs,thus registration accuracy is improved;Then,a least square loss is used to substitute cross entropy loss constructed by the logistic regression objective.The convergence condition of least-squares loss is more strict,which can alleviate the gradient disappearance and overfitting,therefore robustness of model training is enhanced.In addition,Relative Average GAN(Ra GAN)is embedded into the standard Discriminator Network,namely,adding a gradient penalty in Discriminator Network,which reduce the error of Discriminator.Therefore,the registration error can be decreased and the registration accuracy can be stabilized.Finally,this registration model is trained and validated on DRIVE Dataset,generalization performance tests are performed on Sunybrook Cardiac Dataset and Brain MRI Dataset.Compared with state-of-the-art methods,theoretical analysis and extensive experiments demonstrate the proposed model achieves a good registration results,both registration accuracy and generalization ability have been improved,and it is suitable for medical image registration with large non-rigid deformation.(2)This paper proposes a method to augment the limited training datasets by combining an Enhanced Adversarial Auto-encoder and an Modified Conditional Generative Adversarial Networks.Firstly,a style encoder is added to the original AAE model to generate new reconstructed vessel tree images with richer and more diverse information features,and then combines it with the outer contour mask of retina as an input to the improved CGAN model to obtain diverse fundus retinal images.The registration models based on augmented training datasets can have higher registration quality and better generalization ability.Then,a least square loss is used to substitute cross entropy loss constructed by the logistic regression objective.The convergence condition of least-squares loss is more strict,which can alleviate the gradient disappearance and overfitting,therefore robustness of synthetic model training is enhanced.Both theoretical analysis and experimental results show that the synthesis model can effectively expand the limited training datasets and enrich the structural diversity of the initial datasets.(3)In order to verify the practicability of the synthetic dataset in the field of medical image registration,which further improve the registration accuracy and generalization ability.Firstly,Training registration model of this paper based on this augmented DRIVE datasets;And then generalization performance tests are performed on the multimodal Brain MRI Dataset and Liver Dataset;Finally,compared with stateof-the-art registration methods,this registration model based on this amplified datasets has achieved better registration accuracy and generalization ability by theoretical analysis and experimental results.