A study published in Nature Machine Intelligence highlights the potential of Eye2Gene – Heidelberg Engineering’s AI-powered analysis tool that enables real-time gene prediction directly from multimodal Spectralis imaging – to accelerate genetic diagnosis in patients with inherited retinal diseases (IRDs).1
Eye2Gene integrates within the Heyex 2 platform via Heidelberg AppWay, leveraging fundus autofluorescence (FAF), infrared reflectance (IR), and spectral-domain optical coherence tomography (SD-OCT), providing a non-invasive decision-support tool for clinicians.
In more than 75% of tested cases, it outperformed popular phenotyping-only tools in prioritising disease-causing genetic variants
A Deep Learning Model with Expert-Level Performance
The AI system was trained on 58,030 multimodal retinal scans from 2,451 patients with confirmed genetic diagnoses and further externally validated on 775 patients from five sites. Covering 63 disease-associated genes, Eye2Gene captures more than 90% of IRD cases in Europe, demonstrating a broad clinical relevance.
Lead author of the study, Associate Professor Nikolas Pontikos from University College of London said the research demonstrated “a top five prediction accuracy of 83% compared to world-leading experts”.
Particularly noteworthy was Eye2Gene’s superiority in interpreting only FAF images, where it reached an accuracy of 76%, compared to 36% or less by experienced clinicians who took part in the study. These results were consistently reproduced across five independent clinical centres – including institutions in Tokyo, Bonn, São Paulo, Oxford, and Liverpool – demonstrating the model’s robustness and generalisability across populations and imaging standards.
In more than 75% of tested cases, it outperformed popular phenotyping-only tools in prioritising disease-causing genetic variants, thereby increasing the likelihood of achieving a definitive diagnosis.
Clinical Value and Global Impact
At its core, Eye2Gene is powered by an ensemble of 15 convolutional neural networks – five per imaging modality – which together generate patient-level predictions by averaging across scans and modalities. This architecture not only improves accuracy but also ensures that the system can adapt to variations in imaging conditions across different sites.
The clinical implications of Eye2Gene are wide-ranging. The tool supports earlier referrals to genetic testing and clinical trials, assists in complex differential diagnoses, and makes expert-level interpretation accessible in settings where specialist expertise may be limited.
Reference
- Pontikos N, Woof WA, Lin S, et al. Next-generation phenotyping of inherited retinal diseases from multimodal imaging with Eye2Gene. Nat Mach Intell . 2025;7:967-978. doi: 10.1038/s42256-025-01040-8.
