Research On Medical Image Registration Based On Unsupervised Deep Learning

Image registration is the basis of medical image analysis,which can establish the spatial correspondences of anatomical structure between different images,and is the key technology to realize precision medical treatment.Image registration plays a significant role in clinical applications such as accurate disease diagnosis,atlas analysis,image-guided radiotherapy and surgical navigation.With the development of medical imaging and equipment,the types,dimensions and scale of medical image data are expanding,and the complexity and diversity of medical image deformation are increasing.As a result,conventional registration algorithms show some limitations in speed,accuracy and robustness.In recent years,deep learning technology has been applied to medical image registration tasks,and does well in speed,accuracy and robustness.This thesis will focus on the remaining registration problems and challenges in different clinical fields,such as complex and diverse anatomical deformation in brain MR image registration,gray difference and artifacts in CBCT and CT image registration,and sliding motion in lung 4D-CT image registration.This thesis is based on unsupervised deep learning technology and aims to propose specific and effective unsupervised registration models through thorough studies on the registration network,learning strategy,similarity metric and regularization term.The main research works are as follows:1.An unsupervised registration model based on adversarial learningAiming at the brain MR image registration with complex and diverse anatomical deformation,this thesis proposes an unsupervised registration model based on adversarial learning.On one hand,in order to solve the problem of complex and diverse anatomical deformation,this thesis designs a multi-scale registration network to comprehensively analyzes the multi-scale image features,so as to accurately evaluate the multi-scale brain deformation.On the other hand,in order to solve the problem that single gray similarity function is not enough to ensure accurate learning of the spatial correspondence of complex brain structures,this thesis adds an additional discriminator network into the unsupervised registration model to distinguish the two registered images,and thus adds an adversarial objective term into the training function to urge the registration network to generate well-registered image pair that is indistinguishable to the discriminator network.The experimental results demonstrate that the adversarial learning strategy can effectively enhance the learning effect of the unsupervised registration model,and improve the accuracy of brain image registration with complex and diverse anatomical deformation.2.An unsupervised registration model based on similarity metric networkAiming at the CBCT and CT image registration with gray difference and artifacts,this thesis proposes an unsupervised registration model based on similarity metric network.On one hand,in order to solve the problem of considerable gray difference between CBCT and CT images,this thesis designs a registration network with two input channels,which analyzes the CBCT and CT image respectively to alleviate the gray difference.Moreover,this thesis proposes a CBCT and CT image similarity metric network,which analyzes multi-scale deep features of CT and CBCT images,to replace the gray similarity metric functions.On the other hand,in order to solve the problem of the artifacts in CBCT images,this thesis proposes the spatial weighting module in the similarity metric network,so as to pay more attention to significant spatial voxels and ignore artifact voxels.The experimental results demonstrate that the similarity metric network can evaluate the similarity between CBCT and CT images and guide the unsupervised learning of registration network more accurately.3.An unsupervised registration model based on spatially adaptive regularization termAiming at the lung 4D-CT image registration with sliding motion,this thesis proposes an unsupervised registration model based on a spatially adaptive regularization term.In order to solve the problem that the traditional regularization term cannot guarantee both the sliding motion and the smooth motion,this thesis incorporates a lung segmentation network into the unsupervised registration model.Through the lung boundary information from the segmentation network,this thesis unites the smooth regularization term and the non-smooth regularization term to construct a spatially adaptive regularization term.The experimental results demonstrate that the proposed adaptive regularization term can flexibly regulate the learning of the lung registration model,and achieve accurate registration of the sliding motion at lung boundary and the smooth motion inside the lung at the same time.