Human Blastocyst Image Generation Using Generative Adversarial Networks

Several medical fields such as In Vitro Fertilization (IVF) suffer from limited databases d require augmentations, to be suitable for machine learning (ML) tasks such as classification. We test the use of various Generative Adversarial Network (GAN) architectures to tackle the problem of limited data by generating high-quality samples. We compare two GAN-based data augmentation techniques for a limited medical image database, utilizing synthetic images. A GAN can create synthetic images by learning to model a probability density function that can represent large data like images. Synthetic samples are then evaluated using metrics and visual evaluation with respect to the original "real" samples.