[www] [BibTex]

Note:

Abstract: The task of medical image classification is increas-ingly supported by algorithms. Deep learning methods likeconvolutional neural networks (CNNs) show superior perfor-mance in medical image analysis but need a high-quality train-ing dataset with a large number of annotated samples. Partic-ularly in the medical domain, the availability of such datasetsis rare due to data privacy or the lack of data sharing practicesamong institutes. Generative adversarial networks (GANs) areable to generate high quality synthetic images. This work in-vestigates the capabilities of different state-of-the-art GAN ar-chitectures in generating realistic breast ultrasound images ifonly a small amount of training data is available. In a secondstep, these synthetic images are used to augment the real ul-trasound image dataset utilized for training CNNs. The train-ing of both GANs and CNNs is conducted with systemati-cally reduced dataset sizes. The GAN architectures are ca-pable of generating realistic ultrasound images. GANs usingdata augmentation techniques outperform the baseline Style-GAN2 with respect to the Fréchet Inception distance by upto64.2%. CNN models trained with additional synthetic dataoutperform the baseline CNN model using only real data fortraining by up to15.3%with respect to the F1 score, espe-cially for datasets containing less than 100 images. As a con-clusion, GANs can successfully be used to generate syntheticultrasound images of high quality and diversity, improve clas-sification performance of CNNs and thus provide a benefit tocomputer-aided diagnostics