Conditional Generative Adversarial Networks for Biological Image Synthesis

Generative Adversarial Networks formulate the task of learning probability distributions as a game theoretic problem. They have been especially popular for generating images due to their ability to generate sharp, plausibly looking samples. GANs already can successfully generate handwritten digits, human faces and indoor scenes, but this list is growing steadily. Osokin et.al. [6] have shown that GANs can be fruitfully applied to biological images, in order to model mutual spatial arrangement of proteins in a cell. This is a task of immediate practical interest, since current biological imaging techniques have certain limitations, and machine learning can, in principle, help to overcome these constraints. In this work, we modify the conditioning technique of Mirza et.al. [20] and construct a GAN-based architecture that is able to jointly generate images of 6 proteins in connection with one another. We also show, that projection discriminator technique, which was recently proposed by Miyato et.al. [21], enables generation of 41 proteins. In addition to that, we explore properties of several GAN architectures and conditioning methods using a number of evaluation metrics on a synthetic data and some conventional datasets, like MNIST and CIFAR-10. Keywords: Generative Adversarial Networks, Conditional Generative Adversar- ial Networks, synthesis of biological images.