Research On The Algorithms Of Medical Image Registration Based On Deep-learning

Medical image registration is a fundamental task in medical image analysis.It aims to establish the spatial alignment of corresponding anatomical structures in different subjects and different stages of medical images to realize the key technology of precision medical treatment.It is crucial in practical clinical applications such as medical atlas analysis,image-guided radiotherapy and navigation surgery,and accurate disease diagnosis on images.Traditional medical image registration methods search and calculate the deformation field between image pairs iteratively.The iterative optimization methods consume a lot of computing resources.They are time-consuming to predict the deformation field between a pair of images,and their accuracy and robustness also have certain limitations.In recent years,with the rapid development of deep learning technology,medical image registration based on deep learning is more competitive than traditional methods in terms of accuracy and robustness.However,the registration technology based on deep learning still faces the challenges of improving the registration accuracy,smoothing the deformation field,and establishing the correspondence between remote voxels.This conducts in-depth research on key technologies,including registration models and deformation smoothing based on the unsupervised training strategy for these challenges in medical image registration tasks.Three high-performance unsupervised registration models based on deep learning are proposed.The main contributions and innovations of this thesis are as follows:1.Research on the unsupervised heart-brain image registration model based on separate encoding and folding correction.Existing deep learning-based deformable registration methods use concatenated image pairs as input to their models,which ignores the independence of anatomical information within images.Furthermore,global regularization leads to the registration model over-or under-constrains the smoothness of the predicted deformation fields smooth,affecting their model registration accuracy.To solve the above two issues,this thesis proposes a twinning model,SEN-FCB,which consists of two sub-networks:(1)the proposed separate encoding convolutional neural network model,SEN,for predicting high-accurate deformation fields;(2)the folding correction block,FCB,for correcting the deformation field to reduce foldings that occur during deformation.The experimental results show that the proposed model in this study improves the registration accuracy and effectively reduces the folding in the deformation field,and the proposed FCB outperforms the global regularization on the restriction of deformation smoothing.2.Research on the improved Transformer-based unsupervised symmetric brain image registration model.To address that the performance of the registration model based on the convolutional neural network is limited by the size of the receptive field,this thesis uses Transformer to model the relationship of remote voxels.However,the standard Transformer has a vast number of parameters and high computational complexity.It can only be applied to the bottom of the model,where only coarse-grained feature information can be obtained at the lowest resolution,which limits the contribution of the Transformer in the registration model.To address this problem,this thesis proposes a convolution-based efficient multi-head self-attention module CEMSA,which reduces the number of parameters of the standard Transformer and captures local spatial context information to reduce semantic ambiguity in the self-attention mechanism.Based on the proposed CEMSA,a symmetric Transformer-based registration model SymTrans is designed in this thesis.SymTrans uses Transformers in the encoder and decoder to model long-range correlations in image pairs.The experimental results of displacement and diffeomorphism registration show that SymTrans has good registration performance,and the ablation experiments demonstrate that the proposed symmetric structure is effective.3.Research on the unsupervised upper abdominal image registration model based on the Swin Transformer.Establishing voxel correspondences in the abdominal images is more complex than in other organs,and establishing the perception of long-range semantic relatedness is critical for abdominal image registration.Therefore,we use Swin Transformer as the basic module for the registration model in this thesis.However,the Transformer-based models split an image into "words," which limits the ability of the Transformer to model and output coarse-grained spatial features,thus restricting the contribution of the Swin Transformer to the registration model.To address this issue,we propose a recovery feature resolution model,RFRNet,which enables the Swin Transformer to provide finegrained spatial feature representations and rich semantic correspondences of image pairs.Furthermore,the inflexible shifting window partitioning strategy of the Swin Transformer,makes it unable to perceive semantic information of uncertain distances and automatically establish information interaction across the global range of windows.Therefore,this thesis proposes a weighted window attention mechanism,WWA,to automatically build global interactions between windows after the regular and cyclic shift window operations of the Swin Transformer.Based on RFRNet and WWA,the proposed unsupervised deformable image registration model,RFR-WWANet,enables to perceive the correlation of distant voxels in abdominal images,thereby facilitating the establishment of meaningful semantic correlations of anatomical structures.The results show that RFR-WWANet achieves significant registration performance improvements.Furthermore,ablation experimental results demonstrate the effectiveness of the RFRNet and WWA designs.