Résumé: Inaccessibility to medical image datasets in today’s technology limits deep-learning studies in the healthcare field. Generative adversarial networks (GANs) can fill this gap by synthesizing data comparable to actual images. GAN is a generative-modeling approach that emulates dataset content using deep learning techniques. Vanilla GAN is not compatible enough to synthesize images, so variants have been developed. In this study, the performances of the deep convolutional GAN (DCGAN) using the sigmoid-based cross-entropy loss function and the least squares GAN (LSGAN) using the mean square error function on diabetic retinopathy images were analyzed. Inception score, which measures visual acuity, and Frechet inception distance, which calculates structural similarity, were used to validate the qualitative results of the generated images. In detailed analyzes, the DCGAN model performed better than the LSGAN model. The evaluations made depend on the loss of generator and discriminator, classification accuracy, quality of generated images and training epoch of the models. As a result, were reported the effect of changing hyperparameters in DCGAN and LSGAN models and the compatibility of the produced images with the quantitative results.

Langue: Anglais