An AI-powered screening system is now autonomously diagnosing retinopathy of prematurity (ROP) in Mongolian neonatal intensive care units (NICUs), reducing false negatives by 40% and delivering results in minutes instead of days. The system, developed by eye care nonprofit Orbis International and partner Siloam Vision, runs offline on a $300 NVIDIA Jetson board and is designed for low-resource settings where pediatric ophthalmologists are scarce.
How it works
The system uses a lightweight convolutional neural network trained on 12,000 local retinal scans. Clinicians capture images using a smartphone-attached camera, which are then analyzed by the AI. Results are available within seconds, allowing immediate follow-up for at-risk infants. The AI does not replace specialists but supports their decision-making, particularly in remote areas where telemedicine is used to review images.
Key components:
- Offline operation: No internet required, critical for rural hospitals.
- Telemedicine integration: Images from provincial clinics are sent to central hospitals for specialist review.
- Training: Local doctors are trained to use the cameras and interpret AI-assisted results.
- Portability: The $300 NVIDIA Jetson board makes deployment feasible in low-income regions.
Deployment and impact
Since its launch in 2023, the program has screened over 170 newborns in Mongolia, conducting more than 270 exams. The first babies screened included twins Ariunbaatar and Ariunnandin at the National Center for Maternal and Child Health (NCMCH) in Ulaanbaatar. Their mother, Otgonchimeg, noted the system’s speed compared to the previous process, which relied on a single ophthalmologist and could take days.
The AI system received breakthrough status from the U.S. Food and Drug Administration, marking it as the first of its kind for ROP screening. Expansion is underway, with plans to launch in Bangladesh this month.
Why it matters
Retinopathy of prematurity is the leading cause of childhood blindness globally, with an estimated 32,000 pre-term babies becoming permanently blind or visually impaired each year. Early detection is critical, as the condition can progress rapidly if untreated. In Mongolia, where vast geography and low population density limit access to specialists, the AI system addresses a critical gap. Similar challenges exist in other low- and middle-income countries, where neonatal care improvements have increased survival rates but left many infants at risk due to limited ophthalmology resources.
Tradeoffs
- False positives/negatives: While the system reduces false negatives by 40%, it is not infallible. Clinicians must still review AI-generated results.
- Training requirements: Local staff need training to use the cameras and interpret outputs, though the system is designed to be user-friendly.
- Hardware limitations: The $300 NVIDIA Jetson board is affordable but may still pose a barrier in the poorest settings.
Bottom line
This AI-driven screening system demonstrates how targeted technology can address critical healthcare gaps in low-resource settings. By combining offline AI, telemedicine, and portable hardware, it provides a scalable model for early detection of ROP and other preventable conditions. For hospitals in remote or underserved regions, it offers a practical way to improve outcomes without relying on scarce specialist expertise.