Research On Deformable Medical Image Registration Based On Unsupervised Learning

Deformable medical image registration,a method aiming to establish dense nonlinear correspondences between a pair of medical images,holds significance in the field of medical imaging.It plays an irreplaceable role in the development and advancement of clinical medicine,drug research,and medical image analysis.Traditional image registration methods formulate image registration as an optimization problem,requiring iterative optimization for each new image pair,which is time-consuming and demands substantial computational resources.Therefore,it is crucial to develop more accurate and efficient deformable medical image registration methods.Supervised learning methods require a large amount of labels,which are mostly manually annotated,and are susceptible to human errors.In contrast,unsupervised learning for image registration does not require manually labeled data and can achieve desirable performance by utilizing large-scale unlabeled medical images,thus receiving more and more attention from researchers nowadays.However,unsupervised learning methods for image registration require more powerful backbone networks to model the spatial correspondence between moving and fixed images,and must also address the correlation between global and local information in 3D medical images,as well as the ability to handle anatomical structures of different scales.These challenges limit the extensive application of relevant methods in clinical practice.Therefore,this thesis conducted research on deformable medical image registration based on unsupervised learning,and the specific content is as follow:(1)To address the issue of spatial correspondence modeling between moving and fixed medical images,this thesis proposes a framework based on large-kernel modern hierarchical convolutional neural network.The framework uses depth-wise convolution with large receptive fields to extract features of 3D medical images,and employs a hierarchical structure to extract information at different scales with fewer parameters.The large-kernel modern hierarchical convolution is utilized to capture the spatial correspondence between the input moving and fixed images in a simpler and more efficient way.(2)To address the challenge of the correlation and interaction between the global and local information of 3D medical images,this thesis proposes a novel deformable medical image registration method based on both global and local 2.5D attention.By jointly embedding the global and local features in the network framework and constructing a pseudo-3D feature weighting map,the method can effectively model the context information between the image pairs and achieve a desirable deformation field through reweighting the extracted features.(3)To address the issue of multi-scale feature integration,this thesis proposes a deformable image registration framework based on multi-scale cross-attention.The framework consists of a dual-parallel feature extraction encoder that uses multi-scale cross-attention modules to process the input moving and fixed images separately.This allows for effective communication between the features of the moving and fixed images,and the discovery of multi-level voxel correspondences.A multi-scale feature integration module is designed to better integrate multiscales feature information and generate a more refined deformation field,ultimately achieving effective registration.