Abstract
Deep learning from medical images is typically hindered by limited access to images and severe imbalance of classes that reduces the effectiveness of typical machine learning algorithms. Generative adversarial networks can be employed to address such issues by creating natural-appearing synthetic images to complement training sets. In this study, we compare two advanced GAN architectures, StyleGAN2 and StyleGAN3, using two publicly available breast imaging datasets: BUSI (ultrasound, 210 malignant cases) and CBIS-DDSM (mammography, 509 malignant cases). Evaluation was based on Fréchet Inception Distance and Kernel Inception Distance. On BUSI, StyleGAN3 achieved FID = 140.7 and KID = 0.06 at 1000 epochs, whereas StyleGAN2 achieved FID = 259.7 and KID = 0.25. On CBIS-DDSM, StyleGAN3 achieved FID = 90.6 and KID = 0.06, and StyleGAN2 achieved FID = 124.8 and KID = 0.10. These results demonstrate that StyleGAN3 has a tendency to synthesize images that are more natural and diversified under limited dataset conditions, at the cost of increased training times, whereas StyleGAN2 provides similar quality at less expensive computational costs. The results indicate the potential of generating medical images and the tradeoff between image quality and efficiency in data augmentation for breast cancer image improvement.