Multimodal magnetic resonance imaging (MRI) provides complementary tissue contrast, which is crucial for the accurate assessment of brain tumors. Therefore, the effectiveness of segmentation models largely depends on the degree of integration of information from different MRI modalities. Traditional fixed fusion strategies, such as early, mid, and late fusion, while offering simple modality combination methods, are limited by their static nature in their ability to adapt to modality-specific differences within the tumor region. Meanwhile, existing attention-based fusion methods, such as MSFF and MAF-Net, while improving cross-modal interaction, typically rely on multi-branch architectures or stacked attention layers, which increases the number of parameters and computational cost.