Diabetic retinopathy, damage to the retina caused by diabetes, is one of the conditions in which AI is used to support diagnosis by analysing images of the back of the eye. This study looked at what happens when the data used to train these systems is unbalanced, and tested a way of reducing the resulting bias.
The researchers used a public dataset of retinal images to which clinicians had added skin-tone labels. To recreate a situation of imbalance, they removed from the training data the images of people with darker skin who had diabetic retinopathy requiring referral to a specialist, while keeping these images in the test data. The standard system was 73.0% accurate for people with lighter skin and 60.5% for people with darker skin, a gap of 12.5 percentage points.
They then trained the system with additional synthetic retinal images created by generative AI models. Accuracy became 72.0% for people with lighter skin and 71.5% for people with darker skin, reducing the gap to 0.5 points. The authors conclude that unbalanced training data can lead to unequal accuracy across groups, and that synthetic images may help reduce this bias in the diagnosis of diabetic retinopathy and possibly in other AI systems used in eye care.
Learn more about this study here: https://doi.org/10.1167/tvst.10.2.13
Reference:
Burlina, P., Joshi, N., Paul, W., Pacheco, K.D., & Bressler, N.M. (2021). Addressing Artificial Intelligence Bias in Retinal Diagnostics. Translational Vision Science & Technology, 10(2), 13.