Ultrasound Breast Tumor Segmentation Using Deep Learning Models

In ultrasound images breast tumor segmentation is vital for accurate diagnosis and treatment, but presents challenges due to tumor variability in shape, boundary, and intensity. This work presents a new technique incorporating visual saliency information into a deep learning model to improve segmentation accuracy. The model successfully captures complicated tumor morphologies and borders by emphasizing spatial locations of high radiological relevance using a U-Net design with attention mechanisms. Evaluated on a dataset of 780 ultrasound images divided into benign, malignant, and normal cases, the suggested model’s highest accuracy was $98 \%$, with improvements validated through metrics such as the Dice coefficient, sensitivity, and specificity. By presenting a precise, reliable method for breast tumor segmentation in ultrasound pictures, this study enhances medical image analysis and provides a path to increased diagnosis accuracy. It may also apply to other medical imaging problems.