Research On Medical Image Registration Based On Unsupervised Deep Learning

Medical image registration is a complex technique in medical image analysis that aims to establish the spatial mapping relationship between related images from different sources.It plays a crucial role in clinical fields such as disease diagnosis,surgical navigation,atlas analysis,and modeling of tissue and organ motion.With the progress of computer and artificial intelligence technology,unsupervised deep learning-based techniques for medical image registration have been extensively researched and developed,gradually replacing traditional iterative methods and supervised learning approaches.Nevertheless,there are still numerous key and challenging issues to be addressed in clinical applications,such as feature distribution differences in multimodal medical image registration,large-scale deformation,and complex motion modeling in soft tissue organ image registration,and utilization strategies for time series features in image sequence registration.In this article,this thesis has conducted innovative research and proposed effective registration algorithms based on unsupervised deep learning.The main contributions and innovations of this work are summarized as follows:1.This thesis proposed a multimodal medical image registration algorithm that utilizes multi-strategy fusion to address the challenge of distribution differences and unclear correspondences between multimodal images.The proposed medical image registration model combines two registration strategies,namely unified modality metric and cross-modality metric,to accurately capture the spatial correspondence between multimodal images.Additionally,the proposed method optimized the image disentanglement network for registration task.The proposed method utilized a fusion network to perform secondary inference on the registration results obtained previously,resulting in improved registration accuracy.Our research findings demonstrate that employing a multi-strategy fusion approach can effectively enhance the performance of unsupervised registration models.2.This thesis proposed novel models for full-cycle 4D CT registration of the lungs based on image sequences.Proposed approach extends traditional image registration from pairs to sequences,and effectively utilizes temporal information contained within the sequence to achieve superior registration results.This thesis proposes a sequence registration model based on time-line modeling,which incorporates Conv-LSTM modules into the U-Net bottleneck layer to capture temporal clues of image sequences.The model also includes motion constraints and registration strategies that adhere to the motion laws of the image sequence.To further enhance its explicit and efficient modeling ability for time-series features,this thesis introduces a sequence registration model with spatio-temporal feature fusion.The proposed model employs Conv-LSTM as the backbone module to construct a U-shaped network and formulates the registration task as a spatio-temporal sequence prediction problem,thereby extending the network’s processing capability from spatial dimensions to temporal dimensions.Experimental results demonstrate that both registration algorithms exhibit superior performance in time series registration tasks.3.This thesis investigates on the task of registration strategy for image sequence registration.In comparison to image pair registration algorithms,image sequence registration algorithms offer superior accuracy and smoothness,as well as more flexible registration task definition strategies.This thesis presents a thorough overview and examination of strategies for sequence registration,which are categorized into three types:inter-frame registration,crossframe registration,and inter-frame/cross-frame compose registration.With the proposed spatiotemporal feature fusion registration network as the basis,this thesis provides a detailed discussion on the implementation and performance of three registration strategies.Additionally,through experiments,this thesis analyzes the performance of spatial and temporal-spatial networks under these registration strategies and offers recommendations for implementing registration tasks.