<p>Ultrasound imaging is widely utilized for the clinical diagnosis of breast cancer owing to its convenience and non-invasive nature. However, automatic segmentation of breast cancer in ultrasound images remains challenging due to uneven intensity distributions, indistinct edges, and ambiguous tumor shapes and sizes. To address these issues, we propose an edge guided Transformer (EGTFormer) for breast cancer ultrasound image segmentation. Specifically, we first design a semantic-aware encoder to learn a deep semantic representation of breast cancer region and establish long-range dependencies from ultrasound image via a sparse self-attention module. Then, our tumor segmentation branch adopts a tumor saliency enhancement module to enhance both local and global features, thus capturing more accurate tumor regions and sizes. Additionally, we devise an edge perception branch to generate detailed edges of tumors as prior information. Finally, we utilize multi-branch learning to effectively aggregate tumor segmentation features and edge features, establishing a new paradigm where explicit edge guidance significantly boosts segmentation precision while preserving clinically interpretable tumor morphology. The comprehensive experimental results demonstrate that EGTFormer outperforms the various comparative methods on the BUSI and BUSI-WHU datasets, achieving F1 of 83.55%, and 90.50%, respectively.</p>