<p>Brain tumor segmentation from multimodal magnetic resonance imaging is essential for computer-aided diagnosis and therapy formulation. There are still a number of obstacles in the accurate segmentation for the computer-assisted application. CNN-based encoder faces challenges in capturing long-range dependencies and lacks robustness when using skip connections, which leads to loss of details and affects segmentation accuracy. Transformer-based encoder can obtain global context, but it is limited by computational constraints when processing high-resolution medical images. To address these issues, we propose MHC-SegNet, a novel Mamba–Hadamard collaboration segmentation network that aims to more efficiently capture global patterns across modalities. We design a collaboration encoder–decoder with scale-aware Mamba encoder and triple-axis group Hadamard product to capture long-range dependencies and extract multi-perspective 3D spatial features with computationally efficient. In addition, scale-aware feature refinement layers are also proposed to represent and aggregate multi-scale information in the collaboration encoder, and a hierarchical fusion skip module is designed to address the loss of detail caused by vanilla skip connections. The proposed modules enable the network to achieve a comprehensive understanding of the input data, and accurately identify tumor features. Experimental results demonstrate that our model achieves the best performance compared to recent state-of-the-art models in segmentation accuracy.</p>