Study On Two-cycle Network Model Of Diffeomorphic Medical Image Registration Based On Demons

Medical image registration is used in many clinical applications,such as image guide,motion tracking,segmentation,dose accumulation,image reconstruction and so on.At present,medical image registration methods include traditional iterative registration algorithms and deep learning-based registration methods,both methods have advantages and disadvantages.The traditional methods has high registration accuracy and can keep the characteristic of diffeomorphism,but it has heavy calculation burden and registration speed is slow;The methods based on deep learning is time-consuming and easy to calculate,but it lacks rigorous mathematical theory support,and registration accuracy and diffeomorphic characteristics also need to be further improved.Therefore,combining traditional methods with deep learning methods to give play to their advantages is one of the research focuses of image registration.Demons algorithm is a famous image registration algorithm in traditional methods.This paper is based on Demons algorithm combined with deep learning technology,research is carried out around the establishment of diffeomorphic registration model of medical images.The main innovations are as follows:· A new deep bicyclic variational PDE registration network is proposed.Firstly,the variational PDE module,PDE-Block,is designed based on the iterative steps of Demons algorithm;Secondly,PDE-Block are stacked to construct a unidirectional registration network,PDE-Trans Net;Finally,based on the diffeomorphism definition,the dual cycle registration model is constructed by combining the two networks for forward and reverse registration.Experimental results show that this model is feasible in diffeomorphic registration of medical images,and its performance is better than partial of that partial deep learning-based registration models.· Compared with some registration models and in the the quality of deformation field,the above two-cycle model still has a lot of room for improvement,so the above model is improved.By embedding PDE-Block into UNet structure,and adding the local orientation consistency function into the existing loss function,an improved doublecycle registration model is proposed.The experimental results show that compared with the above model,the improved model solves the previous shortcomings,improves registration accuracy and deformation field quality,and performs better than other models.