Published in Research

Eye2Gene study supports AI analysis of SPECTRALIS platform for IRDs

This is editorially independent content
6 min read

New research published in Nature Machine Intelligence evaluated the use of Heidelberg Engineering’s artificial intelligence (AI)-based, multimodal SPECTRALIS imaging system in accelerating the genetic diagnosis of inherited retinal diseases (IRDs).

Important to note: The data is based on the Eye2Gene research project as part of a collaboration with the Germany-based high-tech imaging solutions company.

Let’s begin with this imaging system.

With an upgradable, modular design, the SPECTRALIS optical coherence tomography (OCT) imaging platform combines scanning laser fundus imaging with high-resolution OCT.

Check out the array of modules currently available; as for its multimodal imaging options, these include:

  • OCT, multiple scanning laser fundus imaging modalities; widefield and ultra-widefield; scanning laser angiography; and OCT-angiography (OCTA; FDA cleared in July 2024)

And in recent news: Heidelberg received FDA clearance for the SPECTRALIS Flex Module in October 2024.

Now to Eye2Gene—what is it, exactly?

This is a collaborative research project based out of the University College London Institute of Ophthalmology and funded by the National Institute for Health and Care Research and Moorfields Eye Charity.

  • See here for the institutions and organizations involved in this collaboration.

Its purpose: The project has engineered a form of AI (a deep learning algorithm with the Eye2Gene name) that can:

  • Recognize and comprehend retinal scans
  • Use that information to predict the type of IRD a patient has
  • Specify which gene is likely to be affected as a result of that patient’s IRD
  • Once validated, the AI then helps to make decisions about proper patient care

And which imaging modalities are used to capture those scans?

Eye2Gene utilizes fundus autofluorescence (FAF); infrared reflectance (IR), and spectral-domain OCT (SD-OCT) to non-invasively predict the likely at-fault gene for an IRD.

Also keep in mind: The deep learning algorithm is built on 15 “convolutional neural networks (CNN)"—with five per imaging modality—that, collectively, generate patient-level predictions based on 63 distinct IRD genes.

  • What this enables: Ideally, improved accuracy in its identification as well as ensuring adaptability to “variations in imaging conditions across different sites.”

Gotcha. So how is Heidelberg involved in the project?

The company partnered with Eye2Gene in May 2024 to integrate two of its imaging technologies (SPECTRALIS BluePeak Autofluorescence and HEYEX 1’s Heidelberg Appway) with the project’s AI algorithm.

The intent: To determine how images acquired with the SPECTRALIS can be interrogated by Eye2Gene to predict which gene is causing the inherited disease.

Now to this new research.

Eye2Gene was trained on over 58k scans from 2,451 patients (4,801 eyes; obtained from 9,291 appointments at the Moorsfield Eye Hospital (MEH) and split into those three imaging modalities, with five CNNs trained on each).

Using a single input scan from one of those three modalities, Eye2Gene then obtained a single scan-level gene prediction of IRD—this was then repeated on multiple scans to produce a single production for the patient (based on the average scan-level predictions per modality).

And how was the AI algorithm evaluated?

Researchers simulated Eye2Gene’s retinal scan process using two different patient datasets:

  • 28,174 retinal scans obtained from 524 patients from within the MEH dataset
  • 39,596 retinal scans obtained from 836 patients from five external IRD clinics

For each patient: Eye2Gene was run on all scans to obtain an overall prediction per patient; investigators then compared the prediction of Eye2Gene to the underlying gene diagnosis.

So what did they find?

Eye2Gene actually attained an 83.9% (based on a range of 81.7% to 86%) overall top-five accuracy compared to human experts.

  • Note: This was defined as the number of cases the correct gene appeared in the top-five ranked choices of the model.

Which of those three imaging modalities was better at interpreting?

Between FAF, IR, and OCT, Eye2Gene demonstrated noteworthy superiority in interpreting only FAF images. Its accuracy percentage: 76% (compare this to just 36% achieved by clinicians).

  • This finding was also “consistently reproduced” across the five external IRD clinical sites partaking in the study.

What else to know?

Investigators also evaluated Eye2Gene for its potential as a “next-generation” phenotype tool to identify new gene-phenotype groups.

  • The data: In +75% of tested cases, the AI algorithm outperformed popular phenotyping-only tools in prioritizing disease-causing genetic variants—“increasing the likelihood of achieving a definitive diagnosis.”

As the study authors noted: “Eye2Gene gene predictions aid the interpretation of the results of a multigene genetic test by scoring and thereby prioritize genetic variants where the gene matches the phenotype.”

Plus: See here for details on how it performed in distinguishing between genetic and non-genetic diseases, and click here to learn how the AI algorithm outperformed other AI approaches.

Any notable limitations to be aware of with this?

But of course. See them here.

Alrighty, so what are the key takeaways for the future?

As Heidelberg pointed out, the use of Eye2Gene supports:

  • Earlier IRD referrals to genetic testing and clinical trials
  • Aid in complex differential diagnoses
  • Access to clinician-level interpretation in settings of limited specialist expertise

“By integrating phenotype data into variant prioritization, Eye2Gene increases the diagnostic yield, improving the likelihood of reaching a genetic diagnosis,” the company stated.


How would you rate the quality of this content?