RegMamba: An Improved Mamba for Medical Image Registration

Deformable medical image registration aims to minimize the differences between fixed and moving images to provide comprehensive physiological or structural information for further medical analysis. Traditional learning-based convolutional network approaches usually suffer from the problem of perceptual limitations, and in recent years, the Transformer architecture has gained popularity for its superior long-range relational modeling capabilities, but still faces severe computational challenges in handling high-resolution medical images. Recently, selective state-space models have shown great potential in the vision domain due to their fast inference and efficient modeling. Inspired by this, in this paper, we propose RegMamba, a novel medical image registration architecture that combines convolutional and state-space models (SSMs), designed to efficiently capture complex correspondence in registration while maintaining efficient computational effort. Firstly our model introduces Mamba to efficiently remotely model and process potential dependencies of the data to capture large deformations. At the same time, we use a scaled convolutional layer in Mamba to alleviate the problem of spatial information loss in 3D data flattening processing in Mamba. Then, a deformable convolutional residual module (DCRM) is proposed to adaptively adjust the sampling position and process deformations to capture more flexible spatial features while learning fine-grained features of different anatomical structures to construct local correspondences and improve model perception. We demonstrate the advanced registration performance of our method on the LPBA40 and IXI public datasets.