Multimodal AI Reduces Barriers to Diagnosis of Retinal Genetic Diseases… Demonstrates 88.5% Accuracy in Clinical Trials
Background
Inherited Retinal Disease (IRD) is a representative rare disease that leads to blindness. With over 270 causative genes and varying clinical manifestations among patients, accurate diagnosis is challenging. This is why, on average, it takes more than five years from the onset of symptoms for patients to have their genetic cause identified.
Ophthalmologists perform various imaging tests, such as Color Fundus Photography (CFP) and Optical Coherence Tomography (OCT), to preserve patients' vision. However, it is difficult for humans alone to identify clues to genetic mutations from vast amounts of image data. In healthcare settings with a shortage of skilled retinal specialists, there has been a growing need for an assistive tool to aid in decision-making before genetic testing. Efforts continue to predict genetic mutations in advance to reduce testing costs and time.
Key Findings
The research team led by Professor Xiaodong Sun of Shanghai Jiao Tong University School of Medicine, in collaboration with a global research team including researchers from Korea and Poland, introduced a new clinical decision support system (CDSS) called 'Retina4IRD'. This model is based on 'RETFound', an ophthalmology-specific large visual model that has been pre-trained on approximately 900,000 CFP and 700,000 OCT images. The research team refined Retina4IRD using data from 1,843 genetically confirmed patients collected from the Department of Ophthalmology at Yonsei University Severance Hospital and Gangnam Severance Hospital, as well as the Department of Ophthalmology at the Medical University of Silesia in Poland. In particular, the multimodal design, which analyzes CFP and OCT images as well as the patient's age of onset, gender, and family history, is noteworthy.
The researchers designed a multi-center randomized controlled trial (RCT) to precisely measure the actual diagnostic support performance of Retina4IRD. In a trial conducted on 295 patients suspected of having IRD, the ophthalmologist group assisted by AI achieved a Top-5 Accuracy of 88.5% in identifying the actual causative gene within the top five candidate genes. In contrast, the accuracy of the control group of specialists who diagnosed based only on images and charts was 67.3%, a difference of more than 20 percentage points (P < 0.001). In addition, in the post-management index, which evaluated the patient's ability to conduct additional detailed examinations or establish a treatment plan, the AI-assisted group scored 37.7 points, outperforming the control group (28.5 points) and leading to better treatment decisions. To clarify the model's diagnostic basis, the artificial intelligence highlighted the areas of the image that it focused on in the form of a heatmap, thereby increasing its explainability.
Significance and Prospects
This study is expected to be an important milestone in the approval and introduction process of medical AI solutions. This is because the performance of medical AI has been directly demonstrated through a randomized controlled clinical trial, beyond simple retrospective analysis. The results, which have proven that it statistically significantly improves the diagnostic capabilities of physicians in a real clinical setting, are expected to play a major role in resolving the controversy over the clinical effectiveness of artificial intelligence. In particular, it is of great clinical value in that it helps to quickly identify eligible patients for gene therapy and prevent them from missing the optimal treatment window.
However, there are still several challenges to be solved before it can be widely adopted in actual clinical practice. Currently, the range of genes that Retina4IRD can diagnose is limited to 17 types, which is a clear limitation in covering all 270 or more causative genes of retinal diseases. Further research is needed to address issues such as variations in image quality depending on device specifications and a lack of data on patients with extremely rare mutations. In order to reduce healthcare disparities around the world, it will be necessary to optimize the system so that it can run smoothly even in primary healthcare institutions with poor equipment infrastructure.
Nature Medicine, Published online: 24 July 2026; doi:10.1038/s41591-026-04545-wRetina4IRD, an AI-based clinician decision support system, enhances inherited retinal disease diagnosis through multimodal imaging and clinical data integration, achieving 88.5% accuracy in a randomized clinical trial.
Retina4IRD is practical in that it narrows down the patient's causative gene using inexpensive and accessible fundus examinations and OCT scans before performing costly genetic decoding tests. Previously, Whole Exome Sequencing (WES) or panel tests, which cost millions of won, had to be performed randomly. Now, by targeting the genes suggested by AI and conducting precision tests, the patient's financial burden can be reduced.
In addition, it is a useful tool for researchers at pharmaceutical companies who are recruiting clinical trial subjects for gene therapy. In the case of Luxturna, a high-priced therapy that targets specific gene mutations, early identification and administration of the target patient is critical to the success of the treatment. By using this model, it will be possible to quickly screen potential treatment subjects without expensive whole-genome tests, further accelerating the development and clinical introduction of new drugs.