Breast cancer detection through ultrasound imaging plays a crucial role, particularly for patients with dense breast tissue where mammography may be less effective. The purpose of this research is to propose an improved version of the U-Net model which includes a Self-Attention Mechanism. Additionally, it includes a ResNet50-V2 backbone that combines residual learning and a Pyramid Pooling Module (PPM) in order to efficiently capture complex tumor characteristics. Additionally, an attention-based decoder improves the segmentation process by concentrating on essential areas, lowering the number of false positives, and increasing the level of accuracy. The technique that has been presented makes considerable improvements to the identification of tumors, which in turn facilitates early diagnosis and superior clinical decision-making. The model obtains a segmentation accuracy of 98.73%, indicating greater performance in comparison to approaches that are traditionally performed.