Application Of Generative Adversarial Network for Data Augmentation and Multiplication to Automated Cell Segmentation of the Corneal Endothelium

Considering the automatic segmentation of the endothelial layer, the available data of the corneal endothelium is still limited to a few datasets, typically containing an average of only about 30 images. To fill this gap, this paper introduces the use of Generative Adversarial Networks (GANs) to augment and multiply data. By using the ``Alizarine'' dataset, we train a model to generate a new synthetic dataset with over 513k images. A portion of this artificial dataset is then used to train a semantic segmentation model for endothelial layer segmentation and its performance is evaluated showing that in average the mean intersection over union for all datasets is equal to 81\%. In our opinion, the images of the endothelial layer, together with the corresponding masks generated by the GAN, effectively represent the desired data. The obtained results seem optimistic after visual inspection, since the segmentation is very precise.