Research On 3D Medical Image Registration Based On Unsupervised Learning

Medical image registration aims to establish an anatomical correspondence between two medical images by performing a spatial transformation.Medical image registration plays an important role in the areas of lesion location,disease diagnosis,surgical navigation,radiotherapy,etc.Medical registration has a great application significance for clinical diagnosis and treatment.The traditional medical image registration methods cannot meet the high requirements of real-time clinical registration due to its slow speed and low accuracy.With the advancement of medical imaging equipment and the improvement of computing equipment,image registration based on the deep neural networks has been widely studied.In recent years,image registration methods based on deep learning are roughly divided into two categories: image registration based on similarity metric and image registration based on deep regression networks to predict deformation fields.The registration method based on the similarity metric is slow due to the iterative parameter estimation,especially in the case of deformable registration.It is difficult to achieve effective registration.Therefore,the registration method,which directly predicts the deformation field using deep learning network,has become a research hotspot.This kind of method has an excellent performance in accuracy and speed.Following this registration model,this paper improves U-Net and proposes the following image registration methods.This paper presents an unsupervised image registration method based on iterative N-Net dual loss constraints.This paper improves the U-Net network structure and obtains the N-type underlying registration network(N-Net).In order to obtain higher registration accuracy,the registration process is divided into two steps: coarse registration and fine registration,which are achieved by iteratively using the proposed N-Net registration network.The iterative process uses weight sharing to deepen the network level without increasing the number of parameters.At the same time,this paper introduces a loss function with dual loss constraints.The loss function calculates the image similarity loss of coarse registration and fine registration images,respectively,making the regression network better optimized.This paper proposes an unsupervised image registration method based on feature slice reweighting and multi-scale loss constraints.For the feature maps transmitted by the skip connection,we perform "slicing" on the feature maps of each channel.By compressing the information of these feature slices and learning the interdependence between the feature slices,we obtain the importance coefficient of each feature slice.Finally,all feature slices are reweighted according to the corresponding importance coefficient.The reweighting of feature slices enhances useful features and suppresses features that are not very useful for the current task.This allows meaningful slice features to be used more efficiently.In particular,multi-scale loss constraints are introduced in this paper.We output the deformation field obtained by the middle layer of the registration network.Then perform a spatial transformation on the moving image of corresponding resolution to obtain the intermediate layers' registration images.The image similarity losses of these images are included in the loss function.This achieves multiscale loss constraints and promotes better optimization of the regression network.In this paper,we conduct a lot of experiments on four public human brain 3D image datasets.It is proved that the proposed method can obtain excellent results in medical image registration compared with other popular algorithms.By improving and fusing the first two methods,this paper proposes a W-Net image registration method based on iterative learning.We perform network layer optimization for registration networks based on weighted feature slices.By removing the pooling layer and increasing downsampling,a V-shaped basic registration network is proposed.In order to obtain higher registration accuracy,we follow the progressive registration mode based on iterative NNet registration method.The registration image obtained from the first V-shaped network is used as the input of the second V-shaped network.This builds a W-type network(W-Net).The image similarity loss of this network is obtained by two V-shaped networks,which realizes the dual loss constraint.The superior performance of the algorithm is proved by visualization and quantitative analysis of 3D brain MR image datasets.