Segmentation Assistant And Cascade Based Medical Image Registration

Medical image registration technology is the foundation of the image guidance function,and it plays a vital role in the fields of remote surgery,image-guided radiotherapy and precision medicine.Image registration is essentially an optimization problem.The traditional method needs to iteratively optimize each pair of medical images,which is time-consuming,and low accuracy.The existing learning-based image registration method learns a network by optimizing the similarity between images,and the network outputs a spatial transformation to warp the source image to align with the target image.In recent years,the proposed registration method has achieved great success.However,because image registration itself is a complex optimization problem,medical images are more complicated than natural images,and the application scenarios of image registration are diverse.There are still many problems in medical image registration that need to be solved urgently.This paper build registration models from the region of interest and the registration of the edge of the image to improve the registration effect.The main research results of this paper are as follows:Propose a segmentation fusion algorithm based on segmentation network and atlas-based segmentation.It is difficult to obtain annotation of medical images.When there is only a small amount of annotated data,the performance of segmentation networks based on deep learning is usually weak.Atlas-based segmentation can provide better segmentation labels for unlabeled data when there is only one label.If there are multiple labeled data,it is equivalent to multiple atlases,and segmentation based on multiple atlases can generate multiple segmentation labels for an unlabeled image.Therefore,in this paper,when there is only a small amount of labeled images,the fusion module designed by the convolutional network combines the segmentation generated by two methods to obtain more accurate segmentation.During the training process of segmentation,semi-supervised training‘tricks’are also used,and the segmentation of unlabeled images obtained by the segmentation model proposed in this paper are used to optimize the segmentation model parameters.A segmentation assisted registration network is proposed.Most of the current registration networks are designed for the registration of whole image.In practical applications,many lesions are only connected to specific regions.Therefore,compared with the whole registration effect,the registration effect of specific regions is more important.With the guide of semantic information,the algorithm designed in this paper can improve the registration performance of the region of interesting(ROI).Specifically,with few labeled images,the region of interest is extracted by the segmentation generated by the segmentation model and the original image,and the region of interest is encoded and feed into the existing model framework.In addition,this paper also proposes a region similarity loss,which is used to measure the similarity of the region of interest and guide the registration model learning better parameters.The final experiment shows that the registration accuracy of the algorithm proposed in this paper is better in different regions of interest.A cascaded registration network is proposed.Due to the low voxel value of the edge of each tissue in the image and the large contrast of the image,the effectiveness of a single registration network is not strong.Therefore,this paper cascades two registration networks,and uses the designed fusion module to fuse the deformation fields output by the two registration networks to improve the image registration effect.The input of the first network in the cascade network is the deformed image and the target image,the deformed image is obtained by the output deformation field,and the gradient map of the deformed image and the gradient map of the target image are input into the second network.Finally,experiments show that compared to a single registration network,the cascaded registration algorithm proposed in this paper is more effective.This paper searches and explores the problems in the medical image registration:Combining the atlas-based segmentation method with the learning-based segmentation approach to improve the accuracy of image segmentation.Use semantic information to assist registration,based on the existing registration network.The region of interest extracted via the segmentation information is encoded with an encoder,and feed to the registration network to improve the registration effect of the region of interest;Cascade a registration network with the input of normal images and a registration network with the input of gradient images.The performance of the edge of images is improved by fusing the deformation fields output by the two networks.The experimental results on the LPBA40 and MindBoggle101 data sets show that the performance of registration algorithm proposed in this paper improves in different evaluation indicators.