Unsupervised 3D registration for multi-tasks with local-global self-attention training

Motivation: Image registration is crucial in medical analysis, but traditional iterative methods are time-consuming. While deep-learning techniques are popular, they often underperform in different tasks. Goal(s): The study aims to develop a method that ensures stable and reliable registration accuracy across various scenarios. Approach: We propose LGANet, a deformable image registration method that integrates a dual-stream pyramidal network and local-global self-attention. It initializes the deformation field with a local-global attention module and refines it in a coarse-to-fine manner through feature interaction and fusion. Results: The results tested on three public MRI datasets show that our method achieves superior performance over several start-of-the-art registration methods. Impact: We presents an unsupervised 3D image registration method based on deep learning principles. By incorporating local-global self-attention training, this method is robust across different registration tasks. It ensures consistently accurate registration results and is applicable to multi clinical registration scenarios.