US-Net: A Breast Ultrasound Image Segmentation using Deep Learning

Segmentation of medical images is a crucial step in many clinical applications, including the precise diagnosis and treatment of diseases like breast cancer. Therefore, automated segmentation of breast tumors from breast ultrasound images remains a challenging task. In this paper, we developed a new model, called Ultrasound Network (US-Net), which uses the U-Net architecture with attention gates embedded in the skip connections to assign weights to feature maps based on their importance for the segmentation task. Our method underwent evaluation on three public datasets: BUSI, UDIAT, and STUHospital, using the Dice coefficient as the primary metric for segmentation performance. Notably, US-Net achieved impressive Dice coefficients of 86.99%, 94.38%, and 94% on BUSI, UDIAT, and STUHospital, respectively. Experimental results showed that our network outperformed the latest image segmentation methods for lesion segmentation in breast ultrasound.