[This is a presentation of Jonas Nienhauses diploma project, done under the
supervision of Jan Kybic. He successfully implemented several GANs to
artificially generate microscopy images. Jonas is a double-degree master
student
from RWTH Aachen.]


Artificial image generation based on artificial neural networks is a current
research topic with potential applications on data augmentation in machine
learning tasks, particularly in the fields of medical and biological research.
In this work, a conditional generative
adversarial neural network (cGAN) was implemented and successfully applied to
generate synthetic histology images. For this purpose, several input masks such
as cell and region segmentations, and cell center annotations have been
generated. Based on such masks, images from two datasets, red blood cell images
and lymph node histology, have been used to train the model in order to
generate
realistic images similar to the training images from these datasets. The
quality
of the generated images is subjectively very good in general, although it
strongly depends on the used datasets and level of information provided the
masks.