Generative Adversarial Networks for Retinal Fundus Image Synthesis

To effectively develop a machine learning algorithm, one must first gather a sufficiently big and representative collection of samples. Collecting such a well-curated dataset is expensive, especially as the labeling and annotation sometimes involve the work of more than one expert from the field of interest, as in the case of medical applications. A possible solution to the lack of diverse and large annotated databases is to use Generative adversarial networks (GANs) for medical image synthesis. This thesis aims to present various GAN architectures able to synthesize very realistic-looking fundus images. The data used to train the GANs is a collection of high-resolution retinal images offered by EyePACS and is publicly available. Also, different evaluation methods will be presented and employed to assess the GAN’s output, including quantitative and qualitative validation methods. Lastly, it will be shown how a trained GAN can be efficiently employed for transfer learning experiments, especially in the small data regime setting.