Brain tumor segmentation in MRI is critical for early diagnosis and individualized treatment. However, it remains challenging due to complex tumor morphology, blurred boundaries, and the loss of fine details. To tackle these challenges, we propose the Frequency-Spatial U-Net with Multi-Scale Feature Fusion for Brain Tumor Segmentation (FS-UNet), a dual-domain segmentation framework that integrates spatial boundary enhancement and frequency-domain modeling, aiming to improve both accuracy and robustness. The encoder incorporates a Bitemporal Feature Aggregation Module (BFAM) with multi-scale dilated convolutions and SimAM attention to fuse shallow textures and deep semantics, enhancing boundary sensitivity and contextual consistency. Meanwhile, a Frequency-Spatial Attention Module (FSAM), whose spatial branch is enhanced with a larger receptive field for this task, is integrated in the skip connections to apply 2D Fourier Transform and refine frequency components for enhanced boundary awareness. Experimental results on the BraTS2019 and BraTS2020 datasets demonstrate that FS-UNet outperforms existing methods, particularly in regions with blurred boundaries, showing its robustness and generalization.