Edge-interaction Mamba Network for MRI Brain Tumor Segmentation

Abstract

Recently, state space models, which show competitive performance to Transformers in capturing long-range dependencies at a lower computational cost, have attracted increasing attention in various medical image tasks. Following the merits of Mamba, we propose a novel edge-interaction Mamba network called EiMamba-UNet for MRI brain tumor segmentation. The core of EiMamba-UNet includes a Mamba branch and an edge branch to extract richer brain tumor features, which are further fused through the edge-semantic fusion module (ESF) to obtain more discriminant encoder results. Besides, we introduce an improved cross region attention Mamba module, i.e., CRAMamba, to obtain scale-aware and enhanced feature representation from the Mamba branch. EiMamba-UNet is extensively evaluated on the public BraTS2020-2021 brain tumor segmentation datasets, and experimental results shows its effectiveness and competitiveness compared with the state-of-the-art methods.