DeepRETStroke: AI-Driven Retinal Imaging for Silent Stroke Detection

Overview

DeepRETStroke is a novel AI system that analyzes retinal images to detect silent brain infarctions (SBIs) and predict stroke risk with high sensitivity. Trained on nearly 900,000 retinal scans, it significantly outperforms traditional clinical models and offers a scalable, non-invasive alternative to costly MRI screening.

Background

Stroke remains a leading cause of death and disability worldwide, with early detection of high-risk individuals critical for prevention. Silent brain infarctions, present in about 20% of adults, triple stroke risk but are typically undetected without MRI. Retinal imaging, reflecting cerebrovascular health, has emerged as a promising, accessible screening tool. Previous AI approaches failed to simultaneously detect SBIs and predict stroke risk, leaving a gap in effective early intervention.

Data Highlights

Key Findings

• DeepRETStroke achieved 85.2% sensitivity for SBI detection, closing 78% of the gap between traditional screening and MRI.
• The model was trained using a three-stage pipeline combining self-supervised learning, semi-supervised SBI detection, and knowledge transfer from MRI-validated data.
• Validated across 12 multiethnic cohorts (213,762 images), the fundus model outperformed clinical metadata in detecting SBIs and predicting incident and recurrent strokes.
• In a prospective study, DeepRETStroke-guided management reduced recurrent stroke risk by over 80% compared to conventional screening.
• Limitations include generalizability concerns due to predominant Chinese cohort data and the need for ethnicity-specific model adaptations.

Clinical Implications

DeepRETStroke offers a scalable, cost-effective method for early detection of silent brain infarctions and stroke risk stratification using routine retinal imaging. This approach can enhance preventive strategies by identifying high-risk individuals who might otherwise remain undiagnosed without MRI. Clinicians should consider integrating retinal AI screening, especially in resource-limited settings, while remaining mindful of population-specific validation needs.

Conclusion

DeepRETStroke represents a significant advance in non-invasive cerebrovascular risk assessment, bridging the gap between affordable retinal imaging and high-sensitivity stroke prediction. Continued efforts to improve model generalizability will be essential to maximize its global clinical impact.

References

1. Nature Biomedical Engineering 2024 -- DeepRETStroke: AI for Silent Brain Infarction Detection and Stroke Prediction