Research On Large Deformation Medical Image Registration Algorithm Based On Deep Learning

Medical image registration is a fundamental and core technology in the field of medical image processing and analysis.Its purpose is to align or similar anatomical structures in images acquired from different sources or different times and perspectives to a unified spatial coordinate system through spatial transformation,thus realizing the fusion of multi-modal or multi-temporal image information.Medical image registration is of great clinical value in precision medical applications such as lesion detection,tumor growth monitoring,and image-guided surgical navigation and radiotherapy.In recent years,deep learning has made significant progress in the field of medical image registration.Deep learning-based registration methods use neural networks to directly predict the deformation fields between images,overcoming the limitations of the complex optimization process and a priori assumptions in traditional methods,significantly improving the effectiveness and efficiency of registration.However,existing deep-learning medical image registration methods still face challenges when faced with large-scale deformation problems such as large structural differences or topological changes.Therefore,this thesis proposes an unsupervised 3D medical image registration method based on deep learning to address the problem of large deformation in medical image registration.The main innovation points and work of this method are as follows:(1)A rigid body pre-registration network based on contour features is proposed and integrated into the overall registration framework.The network predicts the global transformation parameters of the floating image from the contour information in the image,allowing it to be matched with the fixed image as a whole,effectively reducing the complexity of large deformation registration.Additionally,a novel regularization term is designed for the global transformation parameters generated by the rigid body registration network to ensure training stability with joint optimization.(2)A deformation registration network based on feature multi-scale and accumulation optimization is proposed.The network constructs a novel multi-scale feature fusion module that can extract image features at different levels and directions and fuse them into a high-dimensional feature representation.The network employs an attention mechanism to enhance registration accuracy in local anatomical regions.The thesis also designs a continuously optimized deformation field prediction mechanism that combines long-and short-range information in the network to generate final spatial displacement vectors,enhancing the model’s ability to cope with large-scale deformations.The registration accuracy is further enhanced while controlling the number of model parameters and avoiding image interpolation distortion.(3)In this thesis,the designed rigid body pre-registration network deformation registration network is integrated into a unified framework,and a large number of qualitative and quantitative experimental analyses are carried out on publicly available medical image datasets such as lung CT,liver CT,and brain MRI to verify the effectiveness of the proposed method in handling large deformation medical image registration tasks.The experimental results are comprehensively evaluated and compared using various evaluation metrics and visualization methods,demonstrating the advantages of this thesis’ s method in maintaining image structural integrity,improving registration accuracy,and adapting to different types and modalities of images.Compared with the current state-of-the-art unsupervised medical image registration methods,the technique proposed in this thesis shows significant improvements in registration accuracy and robustness.