Abstract
In this paper, we describe how to apply image-to-image translation techniquesto medical blood smear data to generate new data samples and meaningfullyincrease small datasets. Specifically, given the segmentation mask of themicroscopy image, we are able to generate photorealistic images of blood cellswhich are further used alongside real data during the network training forsegmentation and object detection tasks. This image data generation approach isbased on conditional generative adversarial networks which have provencapabilities to high-quality image synthesis. In addition to synthesizing bloodimages, we synthesize segmentation mask as well which leads to a diversevariety of generated samples. The effectiveness of the technique is thoroughlyanalyzed and quantified through a number of experiments on a manually collectedand annotated dataset of blood smear taken under a microscope.