Deep Learning for Breast Lesion Classification on Ultrasound Images: A Novel Multi-Dataset Approach
Year: 2026
Type: conference-paper
Abstract: Breast cancer represents the most frequently diagnosed cancer among women globally, with Vietnam experiencing a dramatic doubling of incidence over two decades while lacking comprehensive screening programs. This study proposes a fully automated deep learning pipeline integrating YOLOv8m detection, U-Net segmentation with ResNet152 backbone, and ResNet101 classification for breast ultrasound analysis. Training and validation utilize four heterogeneous public datasets (BUSI, BUS-BRA, BUS-UCLM, QAMEBI) through rigorous five-fold cross-validation. The detection module achieves 99.09% precision and 94.74% recall, while classification demonstrates 91.05% accuracy with ground-truth masks and 82.30% accuracy in fully automated mode. A Streamlit-based clinical application enables real-world deployment in resource-limited Vietnamese healthcare settings. The multi-dataset approach enhances generalizability across diverse populations and imaging conditions, positioning this framework as a practical AI-assisted tool for early breast cancer screening.
Language: English
Topic: AI in cancer detection
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: diamond
OpenAlex ID: W7133802441
