IIT Madras AI Diagnoses Diabetic Retinopathy at 96%
Trained on 2.4 million retinal images and tuned for Indian eyes, the model is built for low-cost smartphone screening.
A team at IIT Madras has published validation results for a deep-learning model that diagnoses diabetic retinopathy — the leading preventable cause of blindness in India — with 96.3 per cent accuracy across a clinical cohort of 48,000 retinal images from four ophthalmology centres in Tamil Nadu and Telangana.
Diabetic retinopathy is damage to the retina's blood vessels caused by diabetes; caught early through screening, most vision loss is preventable.
Key Highlights
- Model DR-Scan Net reached 96.3% accuracy in clinical validation.
- Trained on 2.4 million labelled retinal fundus images.
- Dataset weighted toward Indian eye morphologies.
- Optimised for a smartphone attachment costing about ₹18,000.
- Regulatory clearance expected within 12-18 months.
Built for the Last Mile
The defining choice was optimising the model for affordable smartphone-based retinal cameras rather than costly hospital fundus equipment. Working with a Chennai medical-device startup, the team adapted it for a slit-lamp attachment on standard Android phones, cutting a screening station's hardware cost from about ₹3.5 lakh to ₹18,000.
Of India's estimated 77 million people with diabetes, only a fraction get regular specialist eye exams. The gap is widest in rural and semi-urban areas, where trained community health workers could operate a simplified device if the AI is reliable enough.
The model at a glance
| Detail | Fact |
|---|---|
| Accuracy | 96.3% in clinical validation |
| Training images | 2.4 million labelled fundus images |
| Validation cohort | 48,000 images, four centres |
| Screening station cost | ₹3.5 lakh to ₹18,000 |
| Regulatory timeline | 12-18 months |
Clinical Integration
The team has submitted results to the Central Drugs Standard Control Organisation for medical-device software classification and signed a knowledge-transfer deal with the device startup to commercialise the product. The Indian Council of Medical Research has expressed interest in adding the model to its national diabetic-eye-disease screening protocol, deployed at about 600 district hospitals.
Ophthalmologists caution that AI results must be verified by a specialist before treatment, and that real-world performance needs post-market surveillance. The work sits alongside India's wider deep-tech push, from chip design to space technology.
Frequently Asked Questions
What does DR-Scan Net do?
It diagnoses diabetic retinopathy from retinal images with 96.3% accuracy in clinical validation.
Why tune it for Indian eyes?
Indian eye morphologies and diabetic complication profiles differ from the Western datasets used to train earlier models.
How does it cut cost?
It runs on a smartphone attachment, reducing a screening station's hardware cost to about ₹18,000.
Is it approved for clinical use?
Not yet; classification is under review, with clearance expected in 12-18 months.
Does AI replace the eye specialist?
No. Results must be verified by a specialist before treatment decisions.
Sources
- IIT Madras — Department of Computer Science and Engineering
- Central Drugs Standard Control Organisation
- Indian Council of Medical Research