Generative Adversarial Networks in Retinal Image Classification

The recent introduction of generative adversarial networks has demonstrated remarkable capabilities in generating images that are nearly indistinguishable from real ones. Consequently, both the academic and industrial communities have raised concerns about the challenge of differentiating between fake and real images. This issue holds significant importance, as images play a vital role in various domains, including image recognition and bioimaging classification in the biomedical field. In this paper, we present a method to assess the distinguishability of bioimages generated by a generative adversarial network, specifically using a dataset of retina images. Once the images are generated, we train several supervised machine learning models to determine whether these classifiers can effectively discriminate between real and fake retina images. Our experiments utilize a deep convolutional generative adversarial network, a type of generative adversarial network, and demonstrate that the generated images, although visually imperceptible as fakes, are correctly identified by a classifier with an F-Measure greater than 0.95. While the majority of the generated images are accurately recognized as fake, a few of them are not classified as such and are consequently considered real retina images.