Dual-Branch Brain Tumor Segmentation Algorithm by Fusing CNNs and Mambas

Due to the varying shapes and sizes, blurred boundaries, and unstable positions of brain tumors in magnetic resonance imaging, extracting both local and global features from the images is crucial for accurate segmentation of tumor regions. Existing deep learning methods have insufficient ability to model long-range dependencies in 3D medical images and face challenges in fusing global and local features. To effectively extract and utilize both global and local features, a dual-branch parallel brain tumor segmentation model combining CNNs and Mambas named MamFuse is proposed. MamFuse consists of two branches: one branch employs Mamba module, built on a selective state-space model (SSM), to model long-range dependencies, while the other branch utilizes the convolutional module to extract local information. Additionally, we introduced an innovative fusion module, GAFusion, to integrate multi-level features from both branches. Experimental results on the BraTS21 dataset show that the proposed network outperforms other state-of-the-art segmentation networks, achieving more accurate brain tumor segmentation.