Due to variations in imaging devices and pathological differences, traditional deep learning segmentation models often encounter domain shifts between training and clinical test data, leading to significant performance degradation. Given the high cost of acquiring multi-source target domain data in medical scenarios, Single-Source Domain Generalization (SDG) has emerged as a promising approach, enabling models trained on a single-source dataset to generalize to unseen target domains. However, existing SDG methods based on augmentation techniques suffer from the augmentation gap, where augmented data fail to fully encompass the variations present in clinical test scenarios, potentially resulting in model failure. To address this challenge, we propose a novel SDG algorithm that leverages frequency spectrum analysis and band-specific perturbations to bridge the augmentation gap and enhance domain generalization. Specifically, we identify augmentation regions in the frequency spectrum where conventional augmentations introduce minimal perturbations and apply perturbations to enhance data diversity. Considering model stability, we incorporate frequency band attention mechanisms and an adversarial training framework, ensuring semantic consistency during the augmentation and feature extraction processes. Extensive experiments on medical image datasets demonstrate that our method significantly improves segmentation performance across unseen domains, outperforming state-of-the-art SDG approaches. These findings highlight the potential of our method for real-world clinical deployment, where model robustness to domain shifts is critical.