Transfer Learning with ResNet50 for Breast Cancer Classification in Ultrasound Images

Abstract

Breast cancer remains a predominant cause of mortality among women, highlighting the importance of accurate early detection. Mammography can be hampered by breast density and radiation exposure, while ultrasound is safer and more accessible. However, ultrasound images have speckle noise, low contrast, and often blurred lesion boundaries, making readings subjective. Transfer learning-based Convolutional Neural Network (CNN) approaches are widely used, but structured comparisons of architectures for three-class breast cancer classification (benign, malignant, and normal) on the combined BUSI and Mendeley datasets, particularly with respect to computational efficiency, remain limited. This study evaluates five transfer learning models (VGG16, VGG19, InceptionV3, Xception, ResNet50) on 1,030 images from BUSI and Mendeley. The images were standardised to PNG, converted to RGB, resized to 224×224, noise-reduced with a 3×3 median filter, normalised according to the architecture preprocessing, and augmented conservatively. The dataset was split into an 80% training set, a 10% validation set, and a 10% testing set, utilizing ImageNet weights and implementing partial fine-tuning. ResNet50 exhibited superior performance, attaining an accuracy of 90.29%, a precision of 90.48%, a recall of 90.29%, and an F1-score of 88.74%, outperforming VGG19 (F1-score 87.58%) and Xception (F1-score 86.51%). These findings suggest that computationally efficient models can deliver reliable results and may support computer-aided ultrasound diagnosis in resource-limited environments.