Breast cancer remains a leading cause of morbidity and mortality worldwide, making accurate diagnosis in ultrasound imaging a critical challenge due to inherent speckle noise, low contrast, and variable lesion morphology. In this work, we propose a deep learning framework for breast ultrasound segmentation that integrates a Hierarchical Mix Transformer (MiT-b2) encoder with a U-Net decoder. Unlike traditional convolutional networks, this architecture leverages efficient self-attention mechanisms to capture global contextual dependencies while preserving fine-grained spatial details through multi-scale feature fusion. This approach can help women by enabling earlier and more accurate detection, potentially reducing the number of deaths from breast cancer. To ensure reliability and reproducibility, experiments were conducted on the Breast Ultrasound Images Dataset (BUSI) using a rigorous protocol involving multiple random seeds, Test-Time Augmentation (TTA), and adaptive thresholding. The proposed method demonstrated high stability and performance, achieving a mean Dice Coefficient of 0.7956, an Intersection over Union (IoU) of 0.6728, and a remarkably high Specificity of 0.9883, indicating effective suppression of false positives. Furthermore, geometric evaluation yielded an average Hausdorff Distance (95%) of 28.14 pixels, validating the model’s boundary delineation capabilities. These findings suggest that hierarchical Transformer-based models provide a robust and clinically consistent solution for computer-aided diagnosis in breast ultrasound.