Computer-assisted classification for early tumor detection in breast ultrasound (BUS) images remains a major challenge in biomedical engineering. The use of deep learning networks (DLs) achieves high performance to solve this problem, but large datasets are required. As a solution, we evaluate transfer learning as a method of training DLs previously trained to solve another problem. Transfer learning differs from other methods by fine-tuning a network trained to a specific domain to another, which can be related or not. We fine-tuned five state-of-the-art DLs for binary classification of benign and malignant tumors: VGG16, ResNet, EfficientNet, MixNet, and RegNetY. The DLs were trained with a publicly available BUS dataset, which is expanded by a factor of 29 using data augmentation. The best network, MixNet XL, achieved a balanced accuracy of 99.3% and an area under the ROC curve of 99.8%. The proposed method, compared to other studies, has an equal or better performance using a fraction of training epochs.