MBGNet: Mamba-Based Boundary-Guided Multimodal Medical Image Segmentation Network

Ke Xu, Min Li, Guangjian Liu, Chen Chen, Cheng Chen, Enguang Zuo, Xiaoyi Lv

Abstract: Multimodal medical image segmentation plays an important role in fields such as medical image diagnosis and biomedical research. Although Mamba performs well in medical image feature extraction, it still faces challenges in capturing fine boundaries in lesions. Therefore, in this paper, a Mamba-based boundary-guided multimodal medical image segmentation network (MBGNet) is proposed. To address Mamba’s deficiency in capturing boundary information, we designed a Boundary Information Encoding Module (BIEM). This module employs multiple boundary extraction strategies to capture boundary information across different modalities and uses an external attention mechanism to enhance the interaction and understanding of boundary relationships. Additionally, we designed an Information Guidance Module (IGM) to address information loss during boundary and content fusion. This module uses the boundary segmentation map as a guideline, integrating local features and global context for content segmentation, effectively overcoming information loss. Finally, experimental results on the BraTS2019 and BraTS2020 glioma tumor datasets show that MBGNet achieves DICE coefficients of 88.17% and 88.12%, and Hausdorff 95 distances of 5.21 and 5.08, respectively. These results confirm the superior performance of MBGNet in the segmentation of complex lesion regions, providing a more accurate and reliable method for multimodal medical image analysis.

External IDs:dblp:conf/cvm/XuLLCCZL25