Generative Artificial Intelligence for Retinal Image Translation to Improve Glaucoma Screening With Deep Learning

The purpose of this study was to improve automated glaucoma detection by utilizing generative adversarial networks (GANs) to translate underutilized scanning laser ophthalmoscopy (SLO) fundus images into synthetic color fundus (CF) photographs.

A Cycle-Consistent GAN model (CycleGAN) framework was used to translate 16,936 SLO fundus photographs into corresponding synthetic CF images. Five deep learning models were trained using real CF, synthetic CF, SLO fundus, and combined datasets to classify glaucoma from a holdout test set of real CF photographs. Model performance was evaluated using the area under the operating characteristic curve (AUC) and sensitivity at 90% and 95% specificities.

The "GAN+CFP" model, trained on real and synthetic CF images, achieved the highest AUC (0.94, 95% confidence interval [CI] = 0.93-0.96, P < 0.05) and sensitivity at 90% and 95% specificities (0.83 and 0.77, respectively), outperforming the "CFP" (AUC = 0.89, sensitivities = 0.77 and 0.66), "SLO+CFP" (AUC = 0.88, sensitivities = 0.71 and 0.56), and "GAN" models (AUC = 0.82, sensitivities = 0.51 and 0.33). The "GAN+CFP" and "SLO+CFP" models demonstrated consistent sensitivity across racial and ethnic groups, with "GAN+CFP" yielding superior results across demographics.

GANs effectively translate SLO images into synthetic CF photographs, addressing domain shifts and increasing dataset sizes to enhance glaucoma detection.

GANs may improve glaucoma classification models by improving dataset consistency and mitigating domain shifts. By generating synthetic CF images from SLO data, GANs expand available training data in a clinically relevant imaging modality.