Multi-scale Lung CT Image Registration Based on Mamba-ConvNet

To address the issue of large deformations in medical image registration, we proposed a novel multi-scale lung CT image registration method based on Mamba-ConvNet. By effectively learning multi-scale features and incorporating long-range modeling, the capability of the registration network in dealing with complex deformations is improved. Furthermore, to preserve the topology of transformation, we adopt diffeomorphic deformation with a stationary velocity field, resulting in a more accurate deformation vector field (DVF). Experiment results on the DIR-Lab dataset demonstrate the superiority of the proposed method over current deep learning-based methods in terms of target registration error (TRE), especially in cases with large initial deformations, the proposed method improves by 37%.