KDU Nethra AI analyses retinal fundus photographs and flags patients who need an ophthalmologist's attention. It was built and validated on thousands of graded images from a Sri Lankan tertiary eye care unit.
Diabetic retinopathy (DR) is a complication of diabetes that damages the blood vessels in the retina, the light-sensitive tissue at the back of the eye. Over time, high blood sugar weakens these vessels, causing them to leak fluid, bleed, or become blocked, and in later stages, to trigger the growth of abnormal new vessels. Left unchecked, this damage is one of the leading causes of preventable blindness in working-age adults.
DR usually develops silently. In its early stages there are often no noticeable symptoms, which is why people with diabetes can have significant retinal damage without realising it. Vision changes such as blurring, dark spots, or difficulty seeing at night tend to appear only once the disease has progressed, by which point treatment options are more limited.
This is what makes regular screening so important. Routine fundus photography can catch the disease early, when it is most treatable, well before a person notices any change in their eyesight. Because diabetes is common and often long-term, screening programmes need to check large numbers of patients regularly, which is exactly the kind of workload an AI-assisted read can help triage.
KDU Nethra AI is a diabetic retinopathy (DR) screening tool developed at University Hospital, Sir John Kotelawala Defence University (UHKDU). It was trained on 7-field montage fundus photographs, each graded against the Early Treatment Diabetic Retinopathy Study (ETDRS) classification by an experienced eye surgeon, so the model's judgment is anchored to real specialist opinion rather than automated labels alone.
Rather than trying to replicate every fine ETDRS distinction, the system is tuned for the decision that matters most in a screening setting: which patients are healthy enough to rescreen later, and which ones need a referral to an eye specialist now. Images pass through a staged pipeline that first detects DR-related lesions, then narrows in on the severity of what it finds.
The result is a two-tier read out (Referable and non-referable DR) designed to support, not replace, clinical decision-making in busy screening programmes.
