Teaching the Model
At the outset, the team was aided by ophthalmologists at Aravind and Sankara Nethralaya to label the retina images. After a few short months, the model was trained to identify key markers of diabetic retinopathy, such as nerve tissue damage, swelling and hemorrhaging. And with a larger data set, Gulshan was sure that they could make the model even more accurate.
Enter Dr. Jorge Cuadros, head of the Eye Picture Archive Communication System (EyePACS), a telemedicine network connecting patients in rural areas across the United States to opthamologists for diabetic retinopathy scans. But patients seen by EyePACS still have to wait weeks for a graded scan, and Dr. Cuadros was happy to help any effort for a faster diagnosis.
The data EyePACS shared comprised a wide range of patients and was a hundred times as much as the AI team had gathered by that point. That meant a huge labelling workload because each image had to be graded multiple times to compensate for the bias of different graders. ‘The model learns what things they always did consistently’, says Dale Webster, a software engineer at Google. ‘This tends to result in something that’s a bit less biased and a bit more robust.’
To date, close to 100 ophthalmologists have rendered more than 1 million grades for the AI model.