Innovative Artificial Intelligence (AI) approaches solve mysteries of retinal diseases


Denver, Colo.—Three studies presented this week at the Association for Research in Vision and Ophthalmology’s (ARVO) 2022 Annual Meeting in Denver, Colo. shows how the use of AI technology and machine learning can help solve a variety of issues that scientists and healthcare professionals are challenged to figure out related to retinal diseases. These studies highlight how their research can help fellow researchers and clinicians, and ultimately, optimize care for patients.

Algorithm generates images of eyes at various stages of development

With age-related macular degeneration (AMD), it is often difficult to treat patients or provide them with a timeline-based prognosis, because scientists do not fully comprehend what causes AMD to progress or even occur. As part of the PINNACLE consortium, lead researcher Martin Menten, PhD from the Technical University Munich in Germany, along with researchers from other institutions in Germany, Austria, the United Kingdom, Switzerland, and the U.S., developed a machine learning algorithm that has the capability to reproduce realistic medical images of the eye. The study focused on using this algorithm to compose counterfactual optical coherence tomography (OCT) images that show hypothetical cases where sex, age, or disease stage of the scanned individual are changed while the individual’s identity remains fixed.

Menten and his team created 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.” 175,869 OCT images of primarily healthy patients and 57,875 images of AMD patients undergoing treatment were used. They got five expert ophthalmologists involved in distinguishing between the artificially generated OCT and real images in a visual Turing Test.

Menten and team found that in most cases, the generated counterfactuals were indiscernible from real OCT images. The GAN was able to modify image features related with the patient’s age and sex while preserving their identity in 88.8% ± 6.4% of counterfactuals. Many retinal changes that were observed could be linked to plausible biomarkers like thinning of retinal layers with aging. Menten and his team’s research established GAN’s ability to generate realistic counterfactual OCT images. He said, "We hope that our research will advance the understanding of age-related macular degeneration, so that ultimately doctors will be able to make more accurate prognoses and improved treatment decisions."

  • Abstract title: Discovery of imaging biomarkers for healthy aging and age-related macular degeneration using counterfactual generative adversarial networks
  • Presentation start/end time: Wednesday, May 4, 12:30 – 12:47pm MT
  • Location: Four Seasons Blrm 1 (Denver Convention Center)
  • Also available on the virtual meeting site at beginning May 11.
  • Abstract number: 3855


Another machine proves to be just as accurate as its’ maker

Approximately 85% of people that become afflicted with AMD have the non-exudative (dry) type. There are no proven treatments for dry AMD, hence, recent efforts have focused on the identification of intermediate stages of the disease as potential intervention points. Recently, the Classification of Atrophy Meetings (CAM), a program where global retina experts provide new consensus definitions for atrophy by utilizing multimodal imaging modalities, introduced the term iRORA – incomplete retinal pigment epithelial and outer retinal atrophy – defined as direct precursor of complete retinal pigment epithelium and outer retinal atrophy (cRORA) on OCT B-scans.

Giulia Corradetti, MD, and her team from the Doheny Eye Institute and collaborators from the Computational Medicine Department at the University of California Los Angeles investigated a deep learning algorithm’s ability to automatically identify iRORA and cRORA in AMD eyes using OCT B-scans. They also wanted to validate its’ capabilities. Corradetti and team trained a Resnet18 model to detect presence of iRORA and cRORA at both volume and B-scan level. The training dataset encompassed OCT B-scans from nonneovascular AMD patients and were interpreted by a qualified grader for the presence of iRORA and/or cRORA: 101 OCT volumes (6,138 OCT B-scans) from 37 patients with intermediate and late AMD with no evidence of macular neovascularization (MNV) and 87 OCT volumes (4,128 OCT B-scans) from 34 patients with early and intermediate AMD without evidence of iRORA. The algorithm was tested on two independent external datasets. For the first test set involving 1,117 OCT volumes sampled from the general population, they found that the deep learning algorithm model predicted cRORA and iRORA with an area under the receiver operating curve (AUROC) of 0.99 and 1.0 and area under the precision recall curve (AUPRC) of 0.61 and 0.83, for iRORA and cRORA, respectively.

For the other separately collected test set of 60 OCT B-scans that were enhanced for iRORA and cRORA lesions, the model was able to accurately identify both lesions proving that the algorithm was indeed valid and performed significantly above chance for both iRORA and cRORA. The model can accomplish equivalent or better outcomes in comparison to human graders, eliminating the tedious, time-consuming effort and the inter-grader variability associated with manual annotations.

Corradetti said “iRORA is the direct precursor of cRORA and its reliable detection is critical for facilitating early intervention trials. In this study, we evaluated the performance of a deep learning algorithm for automated detection of iRORA and cRORA on OCT B-scans, which makes the assessment of these lesions practical in the context of clinical trials and clinical practice.”

  • Abstract title: Automated Identification of Incomplete and Complete Retinal Epithelial Pigment and Outer Retinal Atrophy Using Machine Learning
  • Presentation start/end time: Wednesday, May 4, 1:55 – 2:12pm MT
  • Location: Four Seasons Blrm 1 (Denver Convention Center)
  • Also available on the virtual meeting site at beginning May 11.
  • Abstract number: 3860


An AI algorithm efficacious enough in predicting the most common inherited retinal disease genes

Inherited retinal diseases (IRDs) affect approximately 1 in 3,000 people globally. Depending on the type of IRD, some patients might be blind at birth and others might have their vision decline progressively over time. Genetic mutations are linked with IRDs so finding out the affected gene in a patient is crucial and the primary step that leads to diagnosis, prognosis, and treatment. This is usually managed by specialist eye hospitals with IRD experts that can both carry out and interpret imaging and genetics tests. Unfortunately, the availability of this IRD specialist service is not widely available, causing poor diagnostic rates worldwide, resulting in delayed treatment options and vision loss assistance.

Lead researcher Nikolas Pontikos, PhD, and his team from the University College London Institute of Ophthalmology in the United Kingdom along with researchers from other establishments in the UK and Germany are addressing this gap in their research. They created Eye2Gene, an AI algorithm that can deduce the possible IRD causative gene from suspected IRD patients’ retinal scans. They trained and tested Eye2Gene on IRD patients’ retinal scans with a known diagnosis from an established eye hospital. The dataset contained 44,817 images from 1,907 IRD patients “covering 3 modalities: Fundus Auto-Flourescence (FAF), Infrared (IR), and Spectral-Domain Optical Coherence Tomography (SD-OCT).” General applicability was evaluated on a held-out dataset comprised of 264 patients from the original eye hospital and an external group of 37 patients from a different hospital. To measure Eye2Gene against human performance, 8 ophthalmologists assessed a subset of 50 FAF scans.

Eye2Gene produced a top five accuracy of 88% in the held-out dataset, 83% in the external group dataset, and 72% in the human performance dataset compared to 78% for the ophthalmologists. Pontikos and his colleagues’ research shows that the algorithm is adept in deducing the 36 topmost common IRD genes to a top five accuracy of >80% and its performance is comparable to human experts. Pontikos said "as more gene-targeted treatments are being developed for inherited retinal diseases and genetic testing is becoming more widely available, Eye2Gene, a decision support system to democratise and accelerate the genetic diagnosis of inherited retinal disease, will help to identify patients that are eligible for existing treatments or can participate in clinical trials." And hopefully become available in more places worldwide. More information about the project can be found at

  • Abstract title: Eye2Gene: prediction of causal inherited retinal disease gene from multimodal imaging using AI
  • Presentation start/end time: Monday, May 2, 11:42 – 11:59am MT
  • Location: Four Seasons Blrm 1 (Denver Convention Center)
  • Also available on the virtual meeting site at beginning May 11
  • Abstract number: 1161


The Association for Research in Vision and Ophthalmology (ARVO) is the largest eye and vision research organization in the world. Members include approximately 10,000 eye and vision researchers from over 75 countries. ARVO advances research worldwide into understanding the visual system and preventing, treating and curing its disorders. Learn more at

The 2022 ARVO Annual Meeting will take place in Denver, Colo. from May 1 – 4 and virtually May 11 - 12. The Meeting is the premiere gathering of nearly 10,000 eye and vision researchers from around the world. During the Meeting, 4,800 abstracts will be presented on the latest basic and translational research in eye and vision science.

All abstracts accepted for presentation at the Annual Meeting represent previously unpublished data and conclusions. This research may be proprietary or may have been submitted for journal publication. Embargo policy: Journalists must seek approval from the presenter(s) before reporting data from paper or poster presentations. Press releases or stories on information presented at the ARVO Annual Meeting may not be released or published until the following embargo dates:

  • May 1: Official launch of presentations of all posters (both presented in-person and virtually)
  • Rolling basis: Paper session, Symposia, Minisymposia, Cross-sectional Groups, and invited speaker sessions that have specific presentation times will be embargoed until the end of those individual time slots.


Media contact:
Jenniffer Scherhaufer