ITR-Net: A Hybrid Deep Learning Architecture for Precise and Efficient Medical Image Registration

This research proposes a weakly supervised-based learning registration network called Improved Transformer Registration Net (ITR-Net) to improve medical image registration accuracy. Firstly, we improved the transformer module by incorporating patch embedding and a feed-forward layer. These enhancements enable the transformer module to focus on local features in close proximity while establishing connections between distant voxels. Secondly, we embedded this module within U-Net, utilizing a CNN structure to extract image features more precisely and enhance matching accuracy. Then we evaluated the performance of our approach using three-phase CT imaging data of kidneys and lungs. The results demonstrate that our method surpasses traditional and pure CNN-based registration algorithms in terms of both registration accuracy and efficiency.