Imagine being able to check for a leading cause of blindness using just a smartphone photo. That's the promise of a new AI model designed to detect diabetic retinopathy (DR), a complication of diabetes that damages the retina. In a preclinical study, researchers developed a deep learning method that analyzes eye images taken with smartphones—no expensive equipment needed.
The model, which uses a technique called imbalance-aware optimal transport, was tested on images from low-resource settings. It achieved an AUC of 87%, meaning it's good at distinguishing between healthy eyes and those with DR. Specifically, it correctly identified 89% of actual DR cases (sensitivity) and correctly ruled out disease in 95% of healthy eyes (specificity). These results were consistent across repeated tests.
While these numbers are promising, it's important to note this is a preclinical study—not a clinical trial. The model hasn't been tested in real-world clinics yet. The images came from a specific smartphone dataset, so we don't know how well it would work with other phones or populations. Still, for communities where regular eye exams are hard to come by, this approach could eventually make screening more accessible.