The Importance of Talking to Clinicians and Users

The first algorithm that the team was trying to train was an algorithm that detected the presence or absence of diabetic retinopathy. As we talked to the clinicians a little bit more, they were like, that's not useful. We actually need something that actually discriminates between more than mild diabetic Retinopathy. And it turns out at the end of the day, it was about data quality, right? It was about getting the labels that actually accurately reflected what was in the picture and then getting enough of that kind of high quality data made the biggest difference.