<p>Brain tumor segmentation from multi-modal magnetic resonance imaging (MRI) is a fundamental yet challenging task in computer-assisted diagnosis because of substantial tumor heterogeneity, ambiguous boundaries, and the difficulty of jointly modeling fine local structures and global volumetric context in 3D data. We propose a Multi-Branch Hierarchical Fusion Network (MBNet) for accurate, robust, and computationally efficient brain tumor segmentation. MBNet adopts a symmetric multi-branch backbone that preserves high-resolution detail while progressively aggregating low-resolution semantic context through hierarchical cross-scale fusion. In addition, a self-calibrating attention mechanism is introduced to adaptively enhance tumor-relevant responses in high-resolution features using semantically stronger low-resolution guidance. To improve contextual modeling under a compact model budget, we further design a lightweight 3D convolution block with large-kernel depthwise convolution, which enlarges the effective receptive field while controlling parameter growth and computational overhead. Experiments on the BraTS2021 dataset show that MBNet consistently outperforms representative convolution-based and Transformer-based baselines. Specifically, MBNet achieves an average Dice score of 0.914 and an average Jaccard index of 0.855, while maintaining favorable parameter efficiency (8.56M parameters) and a competitive computational cost for large-scale 3D volumetric segmentation. These results demonstrate that the proposed framework effectively balances detail preservation, multi-scale contextual integration, and computational efficiency, highlighting its effectiveness for high-resolution 3D brain tumor segmentation.</p>