The Student-Teacher model is a prominent approach in semi-supervised medical image segmentation. However, the process of knowledge distillation from the Teacher model to the Student model is still constrained by certain modelable approaches, which often rely on specific features in the original domain such as time, space, and other domain-specific characteristics. To overcome this limitation, we propose integrating Fourier domain information into the learning process. By leveraging frequency-based features, the student model gains the ability to capture underlying structures more effectively, thereby improving segmentation performance. This study examines several leading methods in medical image segmentation, including Adversarial Entropy Minimization, Interpolation Consistency Training, Mean Teacher, Uncertainty-aware Self-ensembling Models, Deep Adversarial Networks, and Uncertainty Rectified Pyramid Consistency. Experimental results on the ACDC and BraTS19 datasets reveal significant performance improvements of 6.9% and 2.8% in 2D and 3D segmentation tasks, respectively, by adding parameters of 0.56% and 0.13%.