Research On Deep-Learning Based 3D Medical Brain Image Registration

Medical image registration is a basic step in the task of medical image analysis.With the rapid development of deep learning algorithms,convolutional neural network frameworks have been very popular in the field of medical image registration in recent decades.Since medical image datasets usually lack real deformation displacement fields as training labels,unsupervised registration networks have become the mainstream choice for medical image registration.At present,deep-learning medical image registration algorithms are facing several severe challenges.One is that the number of images in the datasets is scarce,usually only dozens or hundreds of images are available in the medical datasets.The other is that the information between the images to be registered is not fully utilized,which is a defect caused by the traditional two-input registration structure.Thirdly,the existing loss function does not impose sufficient constraints on the registration network,which contributes to the poor effectiveness of registration experiments.In this paper,researches on the above difficulties are carried out,mainly focusing on the 3D medical brain image registration algorithm based on deep learning.The main research content can be divided into the following points:(1)For the first problem,a novel triple-input structure for registration is proposed in this paper.The registration algorithm based on convolutional neural networks usually input an image pair into the network for training,but we randomly select three images as the input group while not fixing the target image.This will produce two advantages: One is that the triple-input structure can enhance the datasets and greatly alleviate the limitation caused by the lack of training data.The second is that the network can extract the internal correlation features of three pictures instead of two pictures during each training epoch.So the network can better learn the relevant information between these images.Compared with the best-performing double-input network Voxelmorph＿unet,triple-input network TIU-Net＿nodsloss,proposed in this paper has improved the Dice value of cerebrospinal fluid,gray matter,and white matter on the LPBA40 dataset by 7.8180%、1.0462%、2.0190%.(2)For the second problem,a deformation enhancement loss function is proposed in this paper.An additional deformation enhancement loss is computed between two registered source images,thus the deformed source image is not only registered to the reference target image but also the anatomical structure of each other is more matched.It not only improves the accuracy of registration and the rationality of deformation,but also strengthens the constraints on the network.Compared with the basic triple-input network TIU-Net＿nodsloss,the experimental results of the triple-input network with enhanced loss function TIU-Net has improved the Dice value of cerebrospinal fluid,gray matter,and white matter on the LPBA40 dataset by 11.7715%、2.2349%、3.1134%.(3)For the third problem,a new concept of segmentation-guided registration is proposed in this paper.In the training stage,the segmentation masks of training images are additionally used to calculate the dice loss,so the weight parameters of the network are iteratively updated through backpropagation to guide the registration network predicting a more reasonable deformation field.Introducing this idea into different network structures,we can construct double-input-based-segmentation neural network for registration and the triple-input-based-segmentation neural network for registration.The experiment also verifies that the idea of segmentation-guided registration can effectively improve registration accuracy and has good universality for different network structures.