Generative Adversarial Network and Retinal Image Segmentation

Abstract

With the advancements in medical equipment technologies, our capabilities to perform diagnosis from retinal images are increasing. However, the need for having more retinal image data is also rising. To investigate and detect possible abnormalities, the accessibility of the retinal images database is vital. While there exist a number of retinal images datasets mainly available from research institutions, the number is scarce and additional data is desired. As a matter of fact, it is not always feasible to collect more data from patients. This issue can be addressed by producing synthetic retinal images from already existing limited data. Hence, it is desirable to develop and validate machine learning algorithms which may generate additional synthetic retinal image data. Generative models provide for generating retinal images. Generative models can learn essential concealed features of data which can be utilized for generating new retinal images having close resemblance to the original images. This chapter provides the reader a detailed discussion on the theoretical and mathematical model of generative adversarial networks (GANs), their use cases and their capability to generate synthetic retinal images.