Abstract
The accurate and timely diagnosis of inherited retinal diseases (IRDs) represents an unmet clinical need in ophthalmology, as the current pathways rely on resource-intensive phenotyping, multidisciplinary expertise and genetic testing. Here we developed Retina4IRD, an artificial intelligence (AI)-based clinician decision support system (CDSS) that predicts 17 genotype categories from retina images. Retina4IRD uses a Vision Transformer model pretrained with RETFound. We then trained and validated Retina4IRD using multimodal data with color fundus photographs and optical coherence tomography scans from 1,843 genetically confirmed patients (3,376 eyes) across China, South Korea and Poland. The top-5 prediction accuracy was 0.904 (95% confidence interval (CI): 0.896–0.912) and 0.856 (95% CI: 0.850–0.863) for internal and external validation, respectively. We conducted a randomized controlled trial with 300 participants with suspected IRD randomized 1:1 to either Retina4IRD-assisted specialist arm or specialist-only arm. Of these, 295 participants (median age 33 years, 114 (38.6%) females) with available next-generation sequencing reports were included in the final analysis. The primary outcome was met: top-5 genetic accuracy was significantly higher in the Retina4IRD-assisted specialist arm versus the specialist-only arm (88.5% versus 67.3%, P < 0.001). For secondary endpoints, top-1 to top-4 accuracies all favored the Retina4IRD-assisted specialist arm, with top-1 accuracy of 37.8% versus 22.4% and top-4 accuracy of 81.8% versus 53.1%, respectively. Post hoc analyses demonstrated that, with Retina4IRD assistance, clinicians made better management decisions, and the composite downstream management score indicated significantly higher scores relative to the control group (37.7 versus 28.5, P < 0.001). Our study shows that Retina4IRD is a CDSS tool prior to genetic testing and aligns with clinical workflow for patients with suspected IRDs. ClinicalTrials.gov identifier: NCT06839170.
This is a preview of subscription content, access via your institution
Access options
Access Nature and 54 other Nature Portfolio journals
Get Nature+, our best-value online-access subscription
$32.99 / 30 days
cancel any time
Subscribe to this journal
Receive 12 print issues and online access
$259.00 per year
only $21.58 per issue
Buy this article
- Purchase on SpringerLink
- Instant access to the full article PDF.
USD 39.95
Prices may be subject to local taxes which are calculated during checkout
Data availability
Data can be shared only for noncommercial academic purposes and will require a formal material transfer agreement. Interested investigators can obtain and certify the data transfer agreement and submit requests to X.S. (xdsun@sjtu.edu.cn). Investigators who consent to the terms of the data transfer agreement, including, but not limited to, the use of these data only for academic purposes, and who agree to protect the confidentiality of the data and limit the possibility of identification of patients will be granted access. Requests will be evaluated on a case-by-case basis within 1 month before receipt of a response. All data shared will be deidentified. For the reproduction of our algorithm code, we have also deposited a minimum dataset at Zenodo (https://zenodo.org/records/20489924 (ref. 40), which is publicly available for scientific research and noncommercial use. Source data are provided with this paper.
Code availability
The code being used in the current study for developing the algorithm is provided at https://github.com/zycl2001/Retina4IRD
References
Georgiou, M. et al. Phenotyping and genotyping inherited retinal diseases: molecular genetics, clinical and imaging features, and therapeutics of macular dystrophies, cone and cone-rod dystrophies, rod-cone dystrophies, Leber congenital amaurosis, and cone dysfunction syndromes. Prog. Retin. Eye Res. 100, 101244 (2024).
Ben-Yosef, T. Inherited retinal diseases. Int. J. Mol. Sci. 23, 13467 (2022).
Schneider, N. et al. Inherited retinal diseases: linking genes, disease-causing variants, and relevant therapeutic modalities. Prog. Retin. Eye Res. 89, 101029 (2022).
Sullivan, L. S. & Daiger, S. P. RetNet: Retinal Information Network. https://retnet.org/
Britten-Jones, A. C. et al. The diagnostic yield of next generation sequencing in inherited retinal diseases: a systematic review and meta-analysis. Am. J. Ophthalmol. 249, 57–73 (2023).
Sheck, L. H. N. et al. Panel-based genetic testing for inherited retinal disease screening 176 genes. Mol. Genet. Genomic Med. 9, e1663 (2021).
Nguyen, Q. et al. Can artificial intelligence accelerate the diagnosis of inherited retinal diseases? Protocol for a data-only retrospective cohort study (Eye2Gene). BMJ Open 13, e071043 (2023).
Wong, W. et al. Inherited retinal disease pathway in the UK: a patient perspective and the potential of AI. Br. J. Ophthalmol. 109, 1266–1271 (2025).
Pierce, E. A. et al. Gene editing for CEP290-associated retinal degeneration. N. Engl. J. Med. 390, 1972–1984 (2024).
Botto, C. et al. Early and late stage gene therapy interventions for inherited retinal degenerations. Prog. Retin. Eye Res. 86, 100975 (2022).
Suh, S., Choi, E. H., Raguram, A., Liu, D. R. & Palczewski, K. Precision genome editing in the eye. Proc. Natl Acad. Sci. USA 119, e2210104119 (2022).
Gulshan, V. et al. Development and validation of a deep learning algorithm for detection of diabetic retinopathy in retinal fundus photographs. JAMA 316, 2402–2410 (2016).
Ting, D. S. W. et al. Development and validation of a deep learning system for diabetic retinopathy and related eye diseases using retinal images from multiethnic populations with diabetes. JAMA 318, 2211–2223 (2017).
Dai, L. et al. A deep learning system for predicting time to progression of diabetic retinopathy. Nat. Med. 30, 584–594 (2024).
Meng, Z. et al. Non-invasive biopsy diagnosis of diabetic kidney disease via deep learning applied to retinal images: a population-based study. Lancet Digit. Health 7, 100868 (2025).
Fujinami-Yokokawa, Y. et al. Prediction of causative genes in inherited retinal disorder from fundus photography and autofluorescence imaging using deep learning techniques. Br. J. Ophthalmol. 105, 1272–1279 (2021).
Fujinami-Yokokawa, Y. et al. Prediction of causative genes in inherited retinal disorders from spectral-domain optical coherence tomography utilizing deep learning techniques. J. Ophthalmol. 2019, 1691064 (2019).
Shah, M., Roomans Ledo, A. & Rittscher, J. Automated classification of normal and Stargardt disease optical coherence tomography images using deep learning. Acta Ophthalmol. 98, e715–e721 (2020).
Miere, A. et al. Deep learning to distinguish ABCA4-related Stargardt disease from PRPH2-related pseudo-Stargardt pattern dystrophy. J. Clin. Med. 10, 5742 (2021).
Pontikos, N. et al. Next-generation phenotyping of inherited retinal diseases from multimodal imaging with Eye2Gene. Nat. Mach. Intell. 7, 967–978 (2025).
100,000 Genomes Project Pilot Investigators et al. 100,000 Genomes pilot on rare-disease diagnosis in health care—preliminary report. N. Engl. J. Med. 385, 1868–1880 (2021).
Conway, M. P. et al. The role of the ophthalmic genetics multidisciplinary team in the management of inherited retinal degenerations—a case-based review. Life (Basel) 14, 107 (2024).
Parker, M. A. et al. Test–retest variability of functional and structural parameters in patients with Stargardt disease participating in the SAR422459 gene therapy trial. Transl. Vis. Sci. Technol. 5, 10 (2016).
Guziewicz, K. E. et al. BEST1 gene therapy corrects a diffuse retina-wide microdetachment modulated by light exposure. Proc. Natl Acad. Sci. USA 115, E2839–E2848 (2018).
Rahman, N., Georgiou, M., Khan, K. N. & Michaelides, M. Macular dystrophies: clinical and imaging features, molecular genetics and therapeutic options. Br. J. Ophthalmol. 104, 451–460 (2020).
A preliminary assessment of the potential impact of rare diseases on the NHS. https://imperialcollegehealthpartners.com/a-preliminary-assessment-of-the-potential-impact-of-rare-diseases-on-the-nhs/ (Imperial College Health Partners, 2018).
Reddy, S. Explainability and artificial intelligence in medicine. Lancet Digit. Health 4, e214–e215 (2022).
Kundu, S. AI in medicine must be explainable. Nat. Med. 27, 1328 (2021).
Desai, A. N. Artificial intelligence: promise, pitfalls, and perspective. JAMA 323, 2448–2449 (2020).
Miere, A. et al. Deep learning-based classification of inherited retinal diseases using fundus autofluorescence. J. Clin. Med. 9, 3303 (2020).
Zhou, Y. et al. A foundation model for generalizable disease detection from retinal images. Nature 622, 156–163 (2023).
Kim, C. et al. Transparent medical image AI via an image-text foundation model grounded in medical literature. Nat. Med. 30, 1154–1165 (2024).
Guo, L. L. et al. A multi-center study on the adaptability of a shared foundation model for electronic health records. NPJ Digit. Med. 7, 171 (2024).
Wang, M. et al. Enhancing diagnostic accuracy in rare and common fundus diseases with a knowledge-rich vision-language model. Nat. Commun. 16, 5528 (2025).
Han, R. et al. Randomised controlled trials evaluating artificial intelligence in clinical practice: a scoping review. Lancet Digit. Health 6, e367–e373 (2024).
Angus, D. C. Randomized clinical trials of artificial intelligence. JAMA 323, 1043–1045 (2020).
Lawrence, N. D. A unifying probabilistic perspective for spectral dimensionality reduction: insights and new models. J. Mach. Learn. Res. 13, 1609–1638 (2012).
Consultation on development of standards for characterization of vision loss and visual functioning: Geneva, 4–5 September 2003. https://iris.who.int/handle/10665/68601 (World Health Organization, 2003).
Vasey, B. et al. Reporting guideline for the early-stage clinical evaluation of decision support systems driven by artificial intelligence: DECIDE-AI. Nat. Med. 28, 924–933 (2022).
Sun, X. An Al-based clinician-decision support system for inherited retinal diseases: a multicenter clinical validation trial. Zenodo. https://zenodo.org/records/20489924 (2026).
Acknowledgements
We sincerely thank all of the hospitals for providing IRD data and the participating retinal specialists who generously donated their time and efforts for the reader studies.
Funding
X.S. discloses support for the research of this work from the National Natural Science Foundation of China (grant 82388101); the National Key R&D Program of China (grant 2022YFC2502800); and the Noncommunicable Chronic Diseases—National Science and Technology Major Project (grants 2023ZD0509202 and 2023ZD0509201). H.J. discloses support for the research of this work from the National Natural Science Foundation of China (grant 82471130); the Shanghai Municipal Talent Work Bureau (grant QNWS2025029); and the Shanghai Jiao Tong University School of Medicine (grant 20250412). B.Q. discloses support for the research of this work from the National Natural Science Foundation of China (grant 62502206). B.S. discloses support for the research of this work from the National Natural Science Foundation of China (grant T2525004) and the Noncommunicable Chronic Diseases—National Science and Technology Major Project (grants 2023ZD0509202 and 2023ZD0509201). D.Z. discloses support for the research of this work from the National Natural Science Foundation of China (grant 62136004) and the National Key R&D Program of China (grant 2023YFF1204803). The funders had no role in study design, data collection and analysis, decision to publish or preparation of the manuscript.
Author information
Authors and Affiliations
Contributions
X.S. and T.Y.W. initiated the project, organized a collaborative team and provided study supervision. H.J., B.S., X.S. and T.Y.W. conceived of the study and its design. Y.Q., J. Wang, J.C., T.L., G.Z., Y. Jin, H.C., J.X., K.X., Y.R., Y. Jiang, Y.Y., W.L., Q.G., Ya Li, Yang Li, B.L., J. Wu, Q.Z., F.L. and W.Z. collected the data for the model’s training and testing. B.S., B.Q., C.Z., S.L., X.C. and D.Z. developed the architectures, training and validation setup. J.H., S.H.B., J.L., A.S, M.S., T.H.R. and A.G. collected the data for the international external validation. X.S., H.J., Yang Li, J. Wu, Q.Z., F.L., W.Z. and Q.G. lead the multicenter RCT. H.J., B.Q. and Y.Q. wrote the paper. X.S., T.Y.W. and B.S. critically revised the paper, and all authors discussed the results and provided feedback regarding the paper. J.H., A.G., Yang Li, J. Wu, Q.Z., F.L., W.Z. and B.L. are the co-leaders of this study and senior authors of this paper. All authors had final responsibility for the decision to submit for publication.
Corresponding authors
Ethics declarations
Competing interests
T.Y.W. is a consultant for 4MDT, Aldropika Therapeutics, Bayer, Boehringer Ingelheim, Carl Zeiss, Plano, Quaerite Biopharm Research Ltd., Roche and Shanghai Henlius. He is an inventor, holds patents and is a co-founder of the start-up company EyRiS, which has interests in, and develops digital solutions for, eye diseases. X.S. is a consultant for Alcon, Allergan, Bayer Healthcare, Carl Zeiss, Innovent Biologics and Roche. The other authors declare no competing interests.
Peer review
Peer review information
Nature Medicine thanks Nikolas Pontikos, Gui-shuang Ying and the other, anonymous, reviewer(s) for their contribution to the peer review of this work. Primary Handling Editor: Mattia Andreoletti, in collaboration with the Nature Medicine team.
Additional information
Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
Extended data
Extended Data Fig. 1 Performance of Retina4IRD based on CFP and metadata.
a, Evaluation of the model on the internal test set (n = 148 cases) across 17 classes using CFP images and metadata, including 26 negative cases, 72 potentially curable, and 50 incurable cases. Patient counts, precision, and recall for each class are displayed alongside the confusion matrix. b, Evaluation of the model on the external test set (n = 455 cases) across 17 classes using CFP images and metadata, including 42 negative cases, 211 potentially curable, and 202 incurable cases. Patient counts, precision, and recall for each class are displayed alongside the confusion matrix. c, Top-N accuracy (N ∈ {3,4,5}) for gene prediction on internal and external test sets based on CFP images and metadata, conducted via 148 and 455 cases, respectively. The bar represents the accuracy and the error bar represents 95% CI. d, Top-N accuracy (N ∈ {3,4,5}) for gene prediction using progressive integration of metadata with CFP images on the internal test set (n = 148 cases). The bar represents the accuracy and the error bar represents 95% CI. e, Top-N accuracy (N ∈ {3,4,5}) for gene prediction using progressive integration of metadata with CFP images on the external test set (n = 455 cases). The bar represents the accuracy and the error bar represents 95% CI. f, Top-N accuracy (N ∈ {3,4,5}) for gene prediction on datasets from Poland (n = 87 cases) and South Korea (n = 164 cases). The bar represents the accuracy and the error bar represents 95% CI.
Extended Data Fig. 2 Performance of Retina4IRD based on OCT and metadata.
a, Evaluation of the model on the internal test set (n = 148 cases) across 17 classes using OCT images and metadata, including 26 negative cases, 72 potentially curable, and 50 incurable cases. Patient counts, precision, and recall for each class are displayed alongside the confusion matrix. b, Evaluation of the model on the external test set (n = 455 cases) across 17 classes using OCT images and metadata, including 42 negative cases, 211 potentially curable, and 202 incurable cases. Patient counts, precision, and recall for each class are displayed alongside the confusion matrix. c, Top-N accuracy (N ∈ {3,4,5}) for gene prediction on internal and external test sets based on OCT images and metadata, conducted via 148 and 455 cases, respectively. The bar represents the accuracy and the error bar represents 95% CI. d, Top-N accuracy (N ∈ {3,4,5}) for gene prediction using progressive integration of metadata with OCT images on the internal test set (n = 148 cases). The bar represents the accuracy and the error bar represents 95% CI. e, Top-N accuracy (N ∈ {3,4,5}) for gene prediction using progressive integration of metadata with OCT images on the external test set (n = 455 cases). The bar represents the accuracy and the error bar represents 95% CI. f, Top-N accuracy (N ∈ {3,4,5}) for gene prediction on datasets from Poland (n = 87 cases) and South Korea (n = 164 cases). The bar represents the accuracy and the error bar represents 95% CI.
Extended Data Fig. 3 Visualization and plausibility of Retina4IRD.
Panels (a)-(f) showed the original OCT images and their corresponding attention maps for USH2A, ABCA4, CYP4V2, CHM, RDH12, and GUCY2D, respectively. The attention maps highlight the regions contributing to the model’s predictions for the six genotypes. g, showed the results of the insertion and deletion experiments performed on the attention maps. The first row shows USH2A, and the second row shows ABCA4.
Extended Data Fig. 4 Top-N Accuracy for each clinical center in FAS.
Panels (a)-(e) present the accuracy for the five main centers in the FAS, labeled as Shanghai General Hospital (n = 154), Eye and ENT Hospital of Fudan University (n = 26), West China Hospital, Sichuan University (n = 22), Beijing Tongren Hospital (n = 31), and Zhongshan Ophthalmic Center, Sun Yat-sen University (n = 52), respectively. The sample sizes for Retina4IRD-assisted group and control group were 79 vs 75 for panel a, 14 vs 12 for panel b, 9 vs 13 for panel c, 13 vs 18 for panel d, and 29 vs 23 for panel e, respectively. Points represent the accuracy. N ∈ {1,2,3,4,5}. FAS, Full Analysis Set.
Extended Data Fig. 5 Retina4lRD enhances physician diagnostic accuracy and downstream management decisions in an RCT.
Panels (a)-(c) present the Correct revisions proportion (CRP), incorrect revision proportion (IRP), and the net benefit proportion (NBP) for physicians after reviewing the Retina4IRD output in the assisted group (n = 148). CRP is defined as the proportion of cases in which the physician’s initial diagnosis was incorrect but was corrected after reviewing the Retina4IRD output. IRP refers to the proportion of cases where the initial diagnosis was correct but became incorrect following review of the Retina4IRD output. NBP is defined as the difference between CRP and IRP. The CRP, IRP and NBP are calculated based on the formula:\({CRP}=\frac{{X}_{1}}{X}\), \({IRP}=\frac{{Y}_{1}}{Y}\), and, \({NBP}={CRP}-{IRP}\),Where the \(X\) denotes the number of initially incorrect annotations by doctors and \({X}_{1}\) is the number of such correct changed cases assisted by Retina4IRD; \(Y\) denotes the number of initially correct annotations by doctors and \({Y}_{1}\) is the number of such incorrectedly changed cases assisted by Retina4IRD. The bars represented the values of CRP, IRP and NBP for panel a, panel b, and panel c respectively. d, Example of a downstream management report in Retina4IRD assisted group. e, Mean composite score for downstream management strategies in the physician-only group versus the Retina4IRD-assisted group in the RCT. The composite strategy was developed based on the top-1 predicted genotype and included four dimensions: Dimension 1, potential eligibility for available therapies or clinical trials; Dimension 2, recommended follow-up interval; Dimension 3, referral for genetic (reproductive) counseling; and Dimension 4, referral for syndromic/systemic evaluation. Each dimension was scored on a 10-point scale, yielding a total score ranging from 0 to 40. The composite strategy score was independently assessed by two experienced IRD experts under masked conditions. These recommendations were applied only to cases with a correct top-1 genotype prediction, as confirmed by NGS. A total of 24 management reports from the Retina4IRD-assisted group and 25 management reports from the control group were included in this analysis. The bars represented the averaged scores for total score and Dimension 1 to 4. The P values, obtained via the two independent samples t-test, were <0.001, <0.001, <0.001, 0.062, and 0.054, respectively.
Supplementary information
Supplementary Information (download PDF )
Supplementary Figs. 1–5, Tables 1–25 and the study protocol for the RCT.
Rights and permissions
Springer Nature or its licensor (e.g. a society or other partner) holds exclusive rights to this article under a publishing agreement with the author(s) or other rightsholder(s); author self-archiving of the accepted manuscript version of this article is solely governed by the terms of such publishing agreement and applicable law.
About this article
Cite this article
Jia, H., Qian, B., Qu, Y. et al. AI-based clinician decision support system for diagnosis of inherited retinal diseases: a multicenter, randomized trial. Nat Med (2026). https://doi.org/10.1038/s41591-026-04545-w
Received:
Accepted:
Published:
Version of record:
DOI: https://doi.org/10.1038/s41591-026-04545-w