In recent years, deep learning has made significant strides in MRI segmentation, while its effectiveness is often curtailed in domain shift scenarios. Unsupervised domain adaptation methods have been proposed to address this issue, with mainstream methods aligning feature distributions across different domains via adversarial learning. However, most methods lack an effective transfer strategy to preserve crucial textural and structural information, leading to suboptimal segmentation outcomes. To address this limitation, we introduce the Frequency-Constrained Vector Quantized Variational Autoencoder (FC-VQVAE), a novel domain adaptation method applicable to various MRI segmentation models. A VQ-VAE is employed for dimensionality reduction of medical images, aligning data from different domains through adversarial learning within the low-dimensional latent space. Meanwhile, additional frequency constraints are introduced to prioritize the retention of vital information for MRI segmentation tasks. Our experimental results indicate that the integration of the FC-VQVAE module significantly enhances the cross-domain segmentation capabilities, underscoring its potential as a versatile solution for domain shift in MRI segmentation.