Assessment of Deep Generative Models for High-Resolution Synthetic Retinal Image Generation of Age-Related Macular Degeneration

Question  Can deep learning be used to synthesize fundus images of age-related macular degeneration (AMD) that appear realistic to retinal specialists?

Findings  In this study of fundus images from 4613 study participants from the Age-Related Eye Disease Study and 133 821 fundus images, the ability of 2 retinal specialists to distinguish real from synthetic fundus images of varying stages of AMD was close to chance, and their diagnostic accuracy similar for real and synthetic images. Machines trained with only synthetic images showed performance nearing that resulting from training on real images.

Meaning  These findings suggest that deep-learning–synthesized fundus images of AMD are realistic and could be used for education of humans across various levels of expertise and for machine training.

Importance  Deep learning (DL) used for discriminative tasks in ophthalmology, such as diagnosing diabetic retinopathy or age-related macular degeneration (AMD), requires large image data sets graded by human experts to train deep convolutional neural networks (DCNNs). In contrast, generative DL techniques could synthesize large new data sets of artificial retina images with different stages of AMD. Such images could enhance existing data sets of common and rare ophthalmic diseases without concern for personally identifying information to assist medical education of students, residents, and retinal specialists, as well as for training new DL diagnostic models for which extensive data sets from large clinical trials of expertly graded images may not exist.

Objective  To develop DL techniques for synthesizing high-resolution realistic fundus images serving as proxy data sets for use by retinal specialists and DL machines.

Design, Setting, and Participants  Generative adversarial networks were trained on 133 821 color fundus images from 4613 study participants from the Age-Related Eye Disease Study (AREDS), generating synthetic fundus images with and without AMD. We compared retinal specialists’ ability to diagnose AMD on both real and synthetic images, asking them to assess image gradability and testing their ability to discern real from synthetic images. The performance of AMD diagnostic DCNNs (referable vs not referable AMD) trained on either all-real vs all-synthetic data sets was compared.

Main Outcomes and Measures  Accuracy of 2 retinal specialists (T.Y.A.L. and K.D.P.) for diagnosing and distinguishing AMD on real vs synthetic images and diagnostic performance (area under the curve) of DL algorithms trained on synthetic vs real images.

Results  The diagnostic accuracy of 2 retinal specialists on real vs synthetic images was similar. The accuracy of diagnosis as referable vs nonreferable AMD compared with certified human graders for retinal specialist 1 was 84.54% (error margin, 4.06%) on real images vs 84.12% (error margin, 4.16%) on synthetic images and for retinal specialist 2 was 89.47% (error margin, 3.45%) on real images vs 89.19% (error margin, 3.54%) on synthetic images. Retinal specialists could not distinguish real from synthetic images, with an accuracy of 59.50% (error margin, 3.93%) for retinal specialist 1 and 53.67% (error margin, 3.99%) for retinal specialist 2. The DCNNs trained on real data showed an area under the curve of 0.9706 (error margin, 0.0029), and those trained on synthetic data showed an area under the curve of 0.9235 (error margin, 0.0045).

Conclusions and Relevance  Deep learning–synthesized images appeared to be realistic to retinal specialists, and DCNNs achieved diagnostic performance on synthetic data close to that for real images, suggesting that DL generative techniques hold promise for training humans and machines.