From retina to brain: how deep learning closes the gap in silent stroke screeningDownload PDF Download PDF CommentOpen accessPublished: 13 November 2025Minyan Ge1,Yuchun Wang1,2 &Shumao Xu1,2 npj Digital Medicine volume 8, Article number: 655 (2025) Cite this articleSubjectsComputational modelsComputational neuroscienceHigh-throughput screeningMachine learningSilent brain infarctions (SBIs), affecting 20% of adults and increasing stroke risk, evade routine MRI screening. While retinal scans offer a “window to the brain,” prior AI failed to simultaneously detect SBIs and predict strokes. DeepRETStroke overcomes this by analysing eye scans. Trained on ~900,000 images, it uses deep learning combining self-supervised pattern recognition from unlabeled images, semi-supervised SBI detection with limited MRI, and knowledge transfer refinement, transforming eye exams into affordable stroke screenings.Stroke persists as a leading global cause of death and disability, imposing immense societal and healthcare burdens1,2. Early identification of high-risk individuals remains paramount for prevention, yet current strategies reliant on self-reported clinical variables or costly neuroimaging suffer from modest predictive power (C-index 0.58–0.73)3 and limited scalability. This challenge is compounded by silent brain infarctions (SBIs): covert cerebrovascular lesions prevalent in ~20% of asymptomatic adults that triple future stroke risk but evade detection without magnetic resonance imaging (MRI)4. Despite their prognostic significance, routine SBI screening remains impractical, creating a critical prevention gap. In response, the field has turned to artificial intelligence (AI) and non-invasive biomarkers. The retina, an embryological extension of the central nervous system, has emerged as a promising high-resolution window to cerebrovascular health, sharing microvascular characteristics with the brain. While pioneering AI studies have used retinal images to estimate systemic cardiovascular risk factors (e.g., hypertension, cholesterol) or predict general vascular events5,6,7,8, they have largely fallen short of directly detecting subclinical brain pathology like SBIs or translating these findings into personalized, longitudinal stroke risk forecasts9,10. This limitation underscores a fundamental disconnect between technical capability and clinical need: true prevention requires identifying covert cerebrovascular injury before the onset of symptomatic stroke.The recent study in Nature Biomedical Engineering bridges this gap through DeepRETStroke, the first AI system capable of simultaneously detecting SBI and predicting both incident and recurrent stroke risk using solely retinal photographs11. Historically, retinal screening has been inexpensive but insensitive, while MRI offers gold-standard sensitivity at prohibitive cost and limited accessibility. The reported system narrows this gap substantially, achieving 85.2% sensitivity, which closes 78% of the deficit between traditional screening (38.7%) and MRI (96%). This capability arises from a domain-specific foundation model pre-trained on nearly 900,000 unlabeled retinal images from large Chinese cohorts (Shanghai Integration Model, SIM, and China National Diabetic Complications Study, CNDCS) using a masked autoencoder approach. This retinal foundation model (ViT-large) architecture learned universal vascular patterns similar to how language models like ChatGPT master grammar, enabling deep phenotyping of cerebrovascular health (Fig. 1a). The model achieves this via a sophisticated three-stage pipeline (Fig. 1b). First, self-supervised pretraining adapts the RETFound vision transformer architecture, a specialized neural network for medical image analysis, to learn generalized retinal features, establishing a robust visual foundation. Second, the encoder and stroke predictor module are initialized to forecast 5-year incident stroke risk using the Shanghai Diabetes Prevention Program (SDPP) cohort, priming the system for cerebrovascular insights. Third, semi-supervised SBI detection refines the model: the SBI detector is trained on 782 MRI-validated data (SDPP-MRI subset), generating soft labels (model-generated pseudo-labels for unannotated images). These soft labels enable knowledge transfer via joint optimization: the encoder, SBI learner, and stroke predictor are co-trained so that the SBI probability distributions align with actual stroke outcomes. This iterative refinement between SBI detection (Stage 3) and stroke prediction (Stage 2) enhances feature robustness. Finally, the system is fine-tuned on the Nicheng Diabetes Screening Project (NDSP) cohort for recurrent stroke prediction.Fig. 1: Architecture and development pipeline of the DeepRETStroke AI system.a Schematic overview of the DeepRETStroke system. The model encodes a domain-specific foundation representing eye-brain connections, enabling downstream applications such as silent brain infarctions (SBI) detection and future stroke prediction. EHR: electronic health record. b Three-stage pretraining and fine-tuning workflow: Stage 1 (Self-supervised pretraining): The retinal image encoder (adapted from RETFound) learns general features from unlabeled images in the SIM and CNDCS datasets. Stage 2 (Incident stroke prediction): The encoder and Stroke Predictor are trained on the SDPP dataset to forecast stroke risk. Stage 3 (Semi-supervised SBI detection): The SBI Detector is trained on limited MRI-labeled data (SDPP-MRI), then generates soft labels, pseudo-probability distributions rather than binary classifications, for the unlabeled SDPP images. Joint training integrates SBI soft labels with actual stroke outcome data to refine the encoder, SBI Detector, and Stroke Predictor. The pretrained system is fine-tuned on NDSP data (patients with prior stroke) for recurrent stroke prediction. Iterative optimization of SBI detection and stroke prediction (Stages 2–3) enhances model robustness. Adopted with permission from ref. 11 Copyright 2025, Springer Nature.Full size imageValidated through a multi-tier framework across 213,762 images from 12 multiethnic cohorts (China, Singapore, UK, US, Denmark, Malaysia), DeepRETStroke achieved exceptional performance (Table 1). For SBI detection, the fundus model significantly outperformed clinical metadata models, with internal AUCs of 0.797 versus 0.633, and external AUCs of 0.751–0.792 versus 0.537–0.726, reaching sensitivity/specificity of 0.775/0.824 in external cohorts. Incident stroke prediction attained an internal AUC of 0.901 (0.846–0.940) and external AUCs of 0.728–0.895, maintaining robustness over 5 years across diabetic/hypertensive subgroups. Recurrent stroke prediction showed internal/external AUCs of 0.769/0.727 for the fundus model, surpassing metadata models (0.568/0.705), with the performance gap reflecting heightened complexity in patients with prior cerebrovascular damage3,12. In a prospective study of 218 Chinese adults with prior stroke/SBI, DeepRETStroke-guided integrated management (IM) demonstrated real-world impact. Non-IM patients stratified as high-risk by the fundus model had 202.17 recurrent strokes/1000 person-years, while fundus-identified low-risk patients showed significantly fewer events versus metadata-based stratification. DeepRETStroke drove an 82.44% relative reduction in recurrent strokes versus conventional screening (preliminary, small sample size). This performance surpassed traditional risk models, positioning retinal imaging as a scalable window into cerebrovascular risk trajectory.Table 1 Multi-tier validation framework for DeepRETStroke performance assessmentFull size tableHowever, despite DeepRETStroke’s promising clinical potential, unresolved challenges limit its applicability: (1) Generalizability limitations: The model’s heavy reliance on Chinese cohorts (due to scarce global paired retinal-brain imaging datasets) risks underrepresenting genetic, environmental, and healthcare disparities across populations. This may introduce biases in AI-driven stroke prediction, particularly for African, Indigenous, or admixed populations, where cerebrovascular phenotypes may differ. To address generalizability limitations rooted in Chinese cohort dependencies, potential solutions include implementing FAIR-compliant federated learning with multi-ethnic data partnerships, coupled with ethnicity-specific adaptive modules to enhance model performance across diverse populations13. This enables clinics across Africa, Indigenous communities, and admixed populations to refine DeepRETStroke without sharing raw data, with ongoing validation in 10,000 underrepresented participants targeting minimal performance variance across ethnic groups. (2) Oversimplified SBI phenotyping: Treating silent infarctions as binary (present/absent) ignores critical heterogeneity. Lacunar infarcts are typically subcortical and of small-vessel origin, versus cortical SBIs carry divergent stroke recurrence risks14,15, yet retinal signatures for these subtypes remain uncharacterized. Future iterations could develop an open-source SBI atlas combining retinal topography with 3D MRI lesion mapping to enable AI-driven subclassification. Achieving this requires overcoming computational hurdles in cross-modal registration, such as aligning MRI voxel-level annotations (lesion load/spatial distribution) with microvascular retinal features. (3) Unresolved biological plausibility: While gradient-based visualizations highlight salient retinal regions, causal relationships between specific vasculature patterns, such as altered fractal dimensions or arteriolar narrowing and cerebrovascular event,s remain speculative. Retinal changes may partly reflect systemic comorbidities (e.g., hypertension)16 rather than direct neurovascular injury correlates, a concern substantiated by observed discordance of some diabetic patients showing retinal pathology without corresponding silent brain infarctions. This discrepancy suggests retinal alterations may reflect either early cerebrovascular dysfunction preceding infarction or systemic confounders. To definitively validate the “eye-brain window” hypothesis, histopathological validation remains essential, leveraging cross-modal analysis that correlates in vivo retinal patterns with post-mortem cerebral microvasculature to map microinfarcts to specific arteriolar abnormalities. This can potentially integrate the following methodologies: counterfactual models isolating retinal biomarkers from systemic comorbidities like hypertension through perturbation analysis; causal scoring quantifying neurovascular linkages against histopathological evidence; and dynamic angiography capturing real-time hemodynamic precursors such as altered flow velocities preceding clinically evident infarction.By transforming retinal photography, a ubiquitous and low-cost tool, into a scalable cerebrovascular crystal ball, this work redefines stroke prevention paradigms, transcending traditional imaging barriers while illuminating the profound interplay between ocular biomarkers and brain health.Data availabilityNo datasets were generated or analyzed during the current study.ReferencesKohli, S., Agarwal, P., Erekat, A., Nadkarni, G. & Kummer, B. Machine learning to predict penumbra core mismatch in acute ischemic stroke using clinical note data. npj Digit. Med. 8, 340 (2025).Article PubMed PubMed Central Google Scholar Santhanam, N. et al. Machine learning-based forecasting of daily acute ischemic stroke admissions using weather data. npj Digit. Med. 8, 225 (2025).Article PubMed PubMed Central Google Scholar Colangelo, G. et al. Prerisk: A personalized, artificial intelligence–based and statistically–based stroke recurrence predictor for recurrent stroke. Stroke 55, 1200–1209 (2024).Article CAS PubMed Google Scholar Charidimou, A. & Smith, E. E. Cardiovascular management in asymptomatic (silent) cerebral microbleeds and suspected cerebral amyloid angiopathy. Stroke 55, 1101–1112 (2024).Article PubMed Google Scholar Cheung, C. Y. et al. A deep-learning system for the assessment of cardiovascular disease risk via the measurement of retinal-vessel calibre. Nat. Biomed. Eng. 5, 498–508 (2021).Article CAS PubMed Google Scholar Lee, Y. C. et al. Multimodal deep learning of fundus abnormalities and traditional risk factors for cardiovascular risk prediction. npj Digit. Med. 6, 14 (2023).Article PubMed PubMed Central Google Scholar Poplin, R. et al. Prediction of cardiovascular risk factors from retinal fundus photographs via deep learning. Nat. Biomed. Eng. 2, 158–164 (2018).Article PubMed Google Scholar Hu, W. et al. Real-world feasibility, accuracy and acceptability of automated retinal photography and AI-based cardiovascular disease risk assessment in Australian primary care settings: a pragmatic trial. npj Digit. Med. 8, 122 (2025).Article PubMed PubMed Central Google Scholar Lee, H.-J. et al. StrokeClassifier: ischemic stroke etiology classification by ensemble consensus modeling using electronic health records. Npj Digit. Med. 7, 130 (2024).Article PubMed PubMed Central Google Scholar Chew, E. Y. et al. Standardization and clinical applications of retinal imaging biomarkers for cardiovascular disease: a Roadmap from an NHLBI workshop. Nat. Rev. Card. 22, 47–63 (2025).Article Google Scholar Jiang, N. et al. A deep learning system for detecting silent brain infarction and predicting stroke risk. Nat. Biomed. Eng. https://doi.org/10.1038/s41551-025-01413-9 (2025).Fanning, J. P. et al. Perioperative stroke. Nat. Rev. Dis. Prim. 10, 3 (2024).Article PubMed Google Scholar Näher, A.-F. et al. Measuring fairness preferences is important for artificial intelligence in health care. Lancet Digit. Health 6, e302–e304 (2024).Article PubMed Google Scholar Hao, Z. et al. Natural history of silent lacunar infarction: 10-year follow-up of a community-based prospective study of 0.5 million Chinese adults. Lancet Reg. Health 17, 100309 (2021).Google Scholar Huang, Z.-X. et al. AI prediction model for endovascular treatment of vertebrobasilar occlusion with atrial fibrillation. npj Digit. Med. 8, 78 (2025).Article PubMed PubMed Central Google Scholar Cheung, C. Y., Biousse, V., Keane, P. A., Schiffrin, E. L. & Wong, T. Y. Hypertensive eye disease. Nat. Rev. Dis. Prim. 8, 14 (2022).Article PubMed Google Scholar Download referencesAcknowledgementsThe authors gratefully acknowledge the ISTBI at Fudan University for its foundational startup support, particularly during its 10th anniversary, a milestone celebrating a decade of pioneering interdisciplinary research in neuroscience and intelligent technologies. This work also coincides with Fudan University’s 120th anniversary (1905–2025), a historic occasion honoring its enduring legacy in advancing global scientific and educational excellence. S.X. acknowledges the support from the National High-Level Talent Project, National Natural Science Foundation of China (No. 22205254), and CFFF computing platform of Fudan University. The authors thank editors' and reviewers' comments in improving this work.Author informationAuthors and AffiliationsInstitute of Science and Technology for Brain-inspired Intelligence (ISTBI), Fudan University, Shanghai, ChinaMinyan Ge, Yuchun Wang & Shumao XuDepartment of Rehabilitation Medicine, Huashan Hospital, Shanghai, ChinaYuchun Wang & Shumao XuAuthorsMinyan GeView author publicationsSearch author on:PubMed Google ScholarYuchun WangView author publicationsSearch author on:PubMed Google ScholarShumao XuView author publicationsSearch author on:PubMed Google ScholarContributionsConceptualization: M.G., Y.W., S.X.; Writing original draft: M.G., S.X.; Review & editing: M.G., Y.W., S.X.; Supervision: S.X.Corresponding authorCorrespondence to Shumao Xu.Ethics declarationsCompeting interestsThe authors declare no competing interests.Additional informationPublisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.Supplementary informationRightslink® by Copyright Clearance CenterRights and permissionsOpen Access This article is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International License, which permits any non-commercial use, sharing, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if you modified the licensed material. You do not have permission under this licence to share adapted material derived from this article or parts of it. The images or other third party material in this article are included in the article’s Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article’s Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by-nc-nd/4.0/.Reprints and permissionsAbout this articleDownload PDF