The Inverse-Consistent Deformable MRI Registration Method Based on the Improved UNet Model and Similarity Attention

Introduction: Deformable image registration is an essential task in medical image analysis. The UNet model, or the model with the U-shaped structure, has been popularly proposed in deep learning-based registration methods. However, they easily lose the important similarity information in the up-sampling stage, and these methods usually ignore the inherent inverse consistency of the transformation between a pair of images. Furthermore, the traditional smoothing constraints used in the existing methods can only partially ensure the folding of the deformation field. Methods: An inverse consistent deformable medical image registration network (ICSANet) based on the inverse consistency constraint and the similarity-based local attention is developed. A new UNet network is constructed by introducing similarity-based local attention to focus on the spatial correspondence in the high-similarity space. A novel inverse consistency constraint is proposed, and the objective function of the new form is presented with the combination of the traditional constraint conditions. Experiment: The performance of the proposed method is compared with the typical registration models, such as the VoxelMorph, PVT, nnFormer, and TransMorph-diff model, on the brain IXI and OASIS datasets. Results: Experimental results on the brain MRI datasets show that the images can be deformed symmetrically until two distorted images are well matched. The quantitative comparison and visual analysis indicate that the proposed method performs better, and the Dice index can be improved by at least 12% with only 10% parameters. Conclusion: This paper presents a new medical image registration network, ICSANet. By introducing a similarity attention gate, it accurately captures high-similarity spatial correspondences between source and target images, resulting in better registration performance.