Umělá inteligence pro generování syntetických retinálních obrazů: Laboratorní úloha

Abstract

This bachelor’s thesis focuses on the use of Generative Adversarial Networks (GANs) for the generation of retinal image data and the design of a laboratory task related to this topic. The thesis summarizes the basics of eye anatomy, principles of fundus imaging, and the theory of GANs, including their properties and training challenges. The practical part deals with the design and implementation of a GAN model in the MATLAB environment. The model was first tested on synthetic datasets and subsequently applied to real retinal data. The experiments showed that the quality of generated images strongly depends on the characteristics of the input data, with the best results achieved using a filtered dataset with reduced variability. The model was able to capture the basic structure of the retina, particularly color and intensity distribution; however, the generated images lacked fine details, especially vascular structures. The thesis also includes the design of a laboratory task in which students implement selected parts of a GAN model based on a provided guideline and gain practical experience with generative models.