Never trusted alone
A classical rule-based grader and a neural network grade every image independently. When they agree, the result stands. When they disagree, a human looks.
Smart India Hackathon · SIH26038
Catch diabetic retinopathy before it steals sight.
One fundus photograph in. An international clinical grade, the lesions that drove it, and a clear referral decision out — built for screening camps where an ophthalmologist isn’t down the hall.
- ICDR grades 0–4
- Two independent graders
- Works offline
What our district simulation found
0%
Ophthalmologist utilisation at 100,000 patients a year. This is the queue that breaks first.
0%
Compute utilisation over the day. The AI is not what limits how far screening can scale.
0
Patients screened in one simulated year across eight camps, with every result ready before they leave.
Why DRishti is different
Three ideas a clinician can trust.
Shows its working
Every grade arrives with the lesion overlay, the attention map and the quadrant counts behind it, so a clinician can disagree with a specific finding.
Degrades instead of failing
No trained model, no toolbox, no network, no language model — each one drops a capability and the screening keeps running.
From photograph to referral
Five stages, each one inspectable.
- 01
Capture
A technician photographs both eyes.
- 02
Quality gate
Blur, exposure and field of view checked. Poor images are retaken.
- 03
Segmentation
Disc, fovea, vessels and three lesion types located.
- 04
Two graders
Clinical rules and a CNN grade independently, then compare.
- 05
Report
Grade, overlay and a plain-language PDF for the patient.
The scale we grade on
International Clinical DR severity, 0 to 4.
Grade 2 and above is referable. DRishti is tuned so that its mistakes lean towards referring, because a missed case costs far more than a false alarm.
- 0No apparent DRNo abnormalities
- 1Mild NPDRMicroaneurysms only
- 2Moderate NPDRMore than microaneurysms
- 3Severe NPDRThe 4-2-1 rule is met
- 4Proliferative DRNew vessels or haemorrhage
Referable: grades 2–4
Major feature
We simulated the whole district, not just the model.
Cameras, bandwidth, servers, grader shifts and specialist appointments are modelled as a Simulink flow model and a discrete-event simulation over a full year of screening. Change any resource and watch where the queue forms.
- Finds the bottleneck resource for any programme size
- Shows the daily burst-and-drain of every queue
- Cross-checks the Simulink model against the discrete-event reference
Ready to screen an eye?
Upload a fundus photograph and get a graded, explained result in seconds.