Discovery of imaging biomarkers for healthy aging and age-related macular degeneration using counterfactual generative adversarial networks

Abstract

Purpose : There is a lack of clinically usable biomarkers for the diagnosis and prognosis of age-related macular degeneration (AMD). Visualizing the effect of healthy and pathological aging on the retina may aid the discovery of novel imaging biomarkers for AMD. To this end, we explore the use of deep learning to synthesize counterfactual OCT images that reflect hypothetical scenarios in which age, sex or disease stage of the scanned subject are changed while the subject’s identity remains fixed.

Methods : We developed a counterfactual generative adversarial network (GAN) that alters existing OCT images to visualize the retina at a different, operator-selectable age, sex or AMD disease stage. Two datasets, which have been curated for the PINNACLE study, were used for GAN training and validation: 175,869 OCT images of predominantly healthy participants in the UK Biobank population study and 57,875 images of AMD patients undergoing treatment at Southampton Eye Unit. The visual quality of the counterfactuals was quantified by conducting a Turing Test, in which five expert ophthalmologists were asked to distinguish between real and artificially generated OCT images. Additionally, we measured whether the generated images faithfully depict the counterfactual age and sex and preserve the subject identity by predicting these demographics using independently trained ResNet-50 neural networks.

Results : The generated counterfactuals were indistinguishable from real OCT images in most cases (Turing Test accuracy: 76.6%±18.4%). The GAN realistically modified image features associated with subject age and sex (correlation between counterfactual and predicted age: Pearson’s R of 0.86±0.02; agreement between counterfactual and predicted sex: 79.7%±5.8%), while preserving subject identity in 88.8%±6.4% of counterfactuals. Several observed retinal changes were linked to plausible biomarkers, such as thinning of retinal layers with aging or increase in drusen size with progressing AMD.

Conclusions : We have demonstrated the ability of GANs to generate realistic counterfactual OCT images, which can be used to visualize the individual course of retinal changes caused by healthy or pathological aging.

This abstract was presented at the 2022 ARVO Annual Meeting, held in Denver, CO, May 1-4, 2022, and virtually.

 

Two representative examples of counterfactuals visualizing the effect of age, sex and disease progression.

 

Quantitative analysis of the GAN’s ability to model age and sex.