Variational Autoencoder as a Data Augmentation tool for Confocal Microscopy Images

Abstract

Retinoblastoma is an ocular tumor characterized by malignant cells in the retina of the eye. For its treatment, doctors apply different methods, one of them is the localized injection of a chemotherapy drug called Topotecan. Scientists in the area of medicine are interested in study the speed of Topotecan penetration in tumor clusters. In order to do this, several sequences of microscopy images of Retinoblastoma cell cultures were taken. However, this process is very complex because Retinoblastoma's tumor clusters require stringent growth and maintenance conditions. That is why there is only one reduced set of images, unique in Argentina. Nonetheless, for this same reason statistical studies applied to uncover this penetration dynamics are unreliable. The objective of this work is to generate synthetic images that have the same characteristics as the original ones and imitate the penetration process of Topotecan. For this purpose, a Variational Autoencoder is applied as a data augmentation technique to maximize the extraction of information from this dataset.