Semantic-aware breast tumors segmentation network.

Breast cancer is a major health threat to women, and early detection via ultrasound imaging is vital for improving survival rates. However, ultrasound images often suffer from low contrast, blurry boundaries, and significant shape variations of tumors. To address these challenges, we propose the Semantic-aware Breast Tumors Segmentation Network (SBTSN). The core contribution is a novel Semantic-aware Block (SAB) designed to refine tumor boundaries by leveraging deep-level semantic guidance to calibrate low-level spatial features. Unlike traditional static attention mechanisms, the SAB dynamically generates attention maps from decoder feedback to suppress background noise and enhance target-related features. Integrated into a lightweight U-Net-based backbone, SBTSN achieves a superior balance between segmentation accuracy and computational efficiency. Extensive experiments on the BUSI and STU datasets demonstrate that SBTSN outperforms state-of-the-art models, achieving superior scores in key segmentation metrics such as Dice and IoU, effectively reducing missed detections in complex clinical scenarios.