USR PUBLISHER

Breast cancer is the most frequently diagnosed cancer among women worldwide, and early, accurate detection remains central to improving survival outcomes. Conventional screening modalities such as mammography and ultrasound are limited by false-positive results, reduced sensitivity in dense breast tissue, and marked inter-reader variability. Deep learning offers a potential route to more consistent, automated interpretation of breast imaging. Here we compare a baseline six-layer convolution neural network (CNN) against a ResNet50 model adapted via transfer learning for three-class classification (normal, benign, malignant) of breast ultrasound images. Using a public dataset of 1,564 ultrasound images from 600 female patients, the fine-tuned ResNet50 model reached a validation accuracy of approximately 90%, compared with approximately 65% for the baseline CNN. Loss curves for the ResNet50 model diverged between training and validation in later epochs, indicating over fitting that we attribute to the limited size of the dataset relative to model capacity. These results support transfer learning as a substantial improvement over a shallow CNN baseline for breast ultrasound classification, while underscoring that larger, multi-centre datasets and explicit regularization are needed before such models could be considered for clinical decision support.