Breast ultrasound (US) is a first line tool for screening of patients for breast cancer disease, but automated lesion segmentation is difficult due to speckle noise, poor contrast and very heterogeneous lesion boundaries. We present a compact U-Net based pipeline developed for robust breast lesion segmentation trained on the Breast Ultrasound Images (BUSI) dataset and released with a deployable demo. The end-to-end system pre-processes the input to 256-by-256 pixels, applies intensity normalization for pixel-wise masks with a lightweight U-Net model enhanced with batch normalization and dropout for stability. It adopts an 80/20 train-test split, keeps learning in check through early stopping and model snapshots, and exports the best model as.keras artifacts. Binary masks for clinical review can be created from probability maps through a simple post-processing threshold. This model is hosted in an interactive Gradio app, which predicts masks and overlays them onto raw scans for transparent visualization, friendly to clinicians. This paper covers the data pipeline, model architecture, training protocol, deployment conditions, and practical lessons learned in terms of reproducible segmentation under resource constraints