Abstract: Domain shifts caused by variations in imaging principles, acquisition protocols, and equipment manufacturers pose significant challenges for deploying deep learning models in medical image segmentation. Single-source domain generalization (SDG) , which relies on data from only a single source domain without requiring access to target domain data, offers a practical solution to this issue. However, existing SDG methods often rely on complex feature operations or auxiliary generative networks, which hinder their efficiency and generalizability. In this research, a Guided Domain Expansion (GDE) method tailored for medical images is proposed. The proposed method employs a monotonic nonlinear mapping function based on Bézier curves to expand the data distribution while preserving the inherent gray-scale characteristics and anatomical spatial structures of medical images. By constraining the transformation intensity, the method can precisely control the direction of domain expansion, thereby avoiding chaotic data distributions and reducing the training burden of the model. To further enhance generalization,a Semantic Information Refining Network is introduced that integrates an Anatomic-Consistent Constraint Framework and an Adversarial Domain Classifier with a gradient reversal layer. Experimental evaluations on cross-modality and cross-sequence datasets demonstrate that this method consistently achieves superior performance, confirming its effectiveness in addressing domain shift challenges in medical imaging.