A Novel Deep Learning Approach for Breast Cancer Ultrasound Image Segmentation

Ultrasound imaging is a widely accessible, cost-effective, and radiation-free modality for early breast cancer screening, particularly valuable in resource-limited settings. In this study, we propose a modified attention U-Net architecture for the automated segmentation of breast lesions in ultrasound images. Leveraging attention gate mechanisms within the U-Net framework, the model adaptively suppresses background noise and enhances lesion-specific feature representation, yielding improved performance over conventional segmentation methods. Evaluated on a dataset comprising 780 annotated breast ultrasound images, the proposed model achieved a test accuracy of 98.24% and demonstrated superior Intersection over Union (IoU) metrics compared to the baseline U-Net. Interpretability analyses using Grad-CAM confirmed that the network effectively localizes diagnostically relevant regions while avoiding false positives. These findings underscore the clinical potential of attention-enhanced deep learning approaches for precise and interpretable breast cancer segmentation in ultrasound imaging.