Simulating Glaucoma Progression in the Retinal Ganglion Cell Layer with Generative Adversarial Networks

Abstract

Objective: The main objective of this study is to develop a tool capable of synthesizing images of the ganglion cell layer (GCL) that simulate glaucoma progression using generative antagonistic networks (GANs).

Material and methods: The dataset includes 406 GCL images of 76 eyes with glaucoma and progression, recorded by a spectral domain optical coherence tomograph (OCT). The Pix2Pix model, a conditional antagonistic generative network, was used to transform the current GCL images into images representing glaucoma progression. A total of 70% of the samples were used for training and 30% for model testing. The structural similarity coefficient was used to analyze the similarity between the real and generated images, and finally, an expert's opinion was used to assess the originality of the generated images.

Results: The synthesized images successfully replicate glaucoma lesion patterns, with good generalizability and reproducibility. The results show a mean structural similarity between 0.76 and 0.78 in the different tests. The test with the expert obtained an accuracy of 57% in distinguishing between real and generated images.

Conclusions: The system developed can generate synthetic images of the GCL with a high similarity to the real ones, demonstrating the effectiveness of the model in synthesizing images that represent the evolution of glaucoma.

Keywords: antagonistas; antagonists; aprendizaje profundo; células ganglionares; deep learning; ganglion cells; glaucoma; tomografía; tomography.