MambaFormer3D: A Hybrid Transformer-Mamba Network for 3D Brain Tumor Segmentation

In recent years, Vision Transformers (ViTs) have demonstrated strong performance in 3D medical image segmentation, particularly in capturing global features and modeling long-range dependencies. However, the self-attention mechanism in ViTs faces significant computational challenges due to its quadratic complexity, especially when applied to three-dimensional medical images, resulting in a high number of parameters and increased computational overhead. This limitation hampers their deployment on resource-constrained devices. To address this challenge, this paper proposes a lightweight MambaFormer architecture, which leverages the efficient representation learning capability of the Mamba SSM and the global feature extraction strength of Transformers, achieving linear computational complexity. The synergy between the MALR and LMA modules ensures computational efficiency while significantly enhancing segmentation performance, particularly in brain tumor segmentation tasks. Experimental results demonstrate that MambaFormer3D exhibits superior performance in processing multi-modal BraTS data, showing great potential in computationally constrained environments.