Enhancing vision Mamba with two-dimensional position embedding and multiscale fusion for medical image segmentation

Medical image segmentation plays a critical role in diagnostic and scientific research, enabling the delineation of regions of interest and the identification of fine-grained lesion at the pixel level. Recent advances in deep learning, particularly in Vision Transformers (ViTs), have significantly accelerated this field. However, existing ViT models often struggle to capture long-range dependencies while maintaining computational efficiency. Therefore, we aimed to design a model that can not only accurately identify target lesions but also ensure computational efficiency.

To achieve this goal, we established vision Mamba, an efficient visual state space model (SSM), and proposed enhancements through the integration of two-dimensional (2D) position embedding and a multiscale feature fusion block (MB). Our 2D position embedding method effectively incorporates spatial information into patch embedding, while the MB supplements detailed information lost during Mamba's sequential processing.

Experiments on three public datasets showed the superiority of our approach. Compared to a suboptimal model, our proposed method achieved average performance improvements of 1.29%, 1.18%, 5.52%, 1.18%, and 1.03% across the three datasets in terms of the dice similarity coefficient, volumetric overlap error, average surface distance between objects, Jaccard coefficient, and recall, respectively.

Our method not only enhances segmentation accuracy but also retains computational efficiency, making it a promising candidate for practical medical image segmentation applications.