Generative adversarial networks for generating synthetic breast ultrasound images from small datasets

While a large and diverse set of medical data is essential for the development of computer-aided diagnostic systems, accessing these datasets in the medical field is particularly challenging due to the need for expert annotators and concerns regarding patient privacy. Artificial Intelligence (AI), specifically Generative Adversarial Networks (GANs), have emerged as a promising solution for augmenting medical image datasets. However, GANs trained on small datasets often encounter issues such as collapse, overfitting, and instability, which result in the generation of unrealistic images. Therefore, we introduce enhanced Deep Convolutional GANs (DCGANs) to generate realistic synthetic breast ultrasound images (BUS) under data-limited conditions. The proposed approach consists of key modifications to the conventional DCGANs: Spectral Normalization (SN), Squeezeand-Excitation (SE) block, Scaled Exponential Linear Unit (SELU), label smoothing, and the use of a specific learning strategy. The evaluation results state that our proposed DCGAN outperforms selected state-of-the-art GANs in the Inception Score (IS), the Structural Similarity Index (SSIM), and the Mean Squared Error (MSE). Additionally, visual analysis results supported by five radiologists with over 15 years of clinical experience demonstrated that on average 64% of the synthetic images generated by our model were considered real images from an ultrasound (US) machine.