Cascaded Deep Convolutional Neural Network Method For Medical Image Registration

With the development of medical imaging equipment,the development of medical image processing technology is also progressing steadily,in which medical image registration plays an important role.For example,registration of brain MR images is helpful for the diagnosis of mental diseases,and registration of lung CT images is helpful for the observation of the pattern of pulmonary breathing movement and the diagnosis of pulmonary diseases.However,the traditional image registration method needs iterative optimization,which requires a large amount of computation and takes a long time.In recent years,following the application of deep learning in various fields,the image registration based on deep learning has also verified its superiority.While achieving the registration effect equivalent to the traditional image registration methods,the deep learning based registration time has been reduced by several orders of magnitude.For registration tasks of MR brain images and 4D CT lung images,this paper proposes registration models based on cascaded deep convolution neural network,which can predict registration image space transformations and registration time is reduced while registration precision is improved by using unsupervised and transfer learning which does not need to provide many real image labels as ground truth.The main contents of this paper are as follows:(1)Multi-stage deep network based registration methods: According to the medical image registration model based on convolutional neural network and the multistage registration idea,a multi-stage network registration model was proposed.The model first performs rough registration of a reference image and an image to be registered to learn the deformation fields by a deep convolutional neural network,and then performs spatial transformation of the image to obtain the registered image.In next stage,the input of the second registration network is the reference image and the resulting image and the output is the deformation fields between the reference image and an image to be registered.(2)Cascaded deep network based registration methods: The cascaded network registration model is proposed.In this model,a series of deep convolutional networks with the same structure and different parameters(which can be learned)are used to learn several small deformation fields between the images end-to-end,while the final large deformation field is obtained by summing all series of small deformation fields step by step,so as to realize the registration of large deformation images.At the same time,the unreal folds of the resulting image is reduced by a Jacobian regularization term of deformation field in the loss function.(3)Unsupervised and transfer learning methods: 3D brain MRI ADNI data set is used for the unsupervised training and testing.The experimental results show that the multi-stage network registration model and the cascaded network registration model can improve the average Dice similarity coefficient of brain tissues and organs,and the cascaded network registration model is of higher registration precision and can effectively reduce the of resulting unreal fold images after registration.Furthermore,the unsupervised and transfer learning,training and testing of the proposed cascaded registration network are carried out for the registration of lung 4D CT data set,DIR Lab.The experimental results demonstrate the feasibility of the cascaded network registration model for the large deformation images.