The task of medical image classification is increasingly supported by algorithms. Deep learning methods like convolutional neural networks (CNNs) show superior performance in medical image analysis but need a high-quality training dataset with a large number of annotated samples. Particularly in the medical domain, the availability of such datasets is rare due to data privacy or the lack of data sharing practices among institutes. Generative adversarial networks (GANs) are able to generate high quality synthetic images. This work investigates the capabilities of different state-of-the-art GAN architectures in generating realistic breast ultrasound images if only a small amount of training data is available. In a second step, these synthetic images are used to augment the real ultrasound image dataset utilized for training CNNs. The training of both GANs and CNNs is conducted with systematically reduced dataset sizes. The GAN architectures are capable of generating realistic ultrasound images. GANs using data augmentation techniques outperform the baseline Style- GAN2 with respect to the Frechet Inception distance by up to 64.2%. CNN models trained with additional synthetic data outperform the baseline CNN model using only real data for training by up to 15.3% with respect to the F1 score, especially for datasets containing less than 100 images. As a conclusion, GANs can successfully be used to generate synthetic ultrasound images of high quality and diversity, improve classification performance of CNNs and thus provide a benefit to computer-aided diagnostics.
