Addressing artificial intelligence bias in retinal diagnostics

AI Generated Summary*

Bias in automated diabetic retinopathy screening was probed by deliberately unbalancing Kaggle EyePACS data (88,692 fundus images). One clinician graded 1555 images by retinal appearance, using pigmentation, vessel caliber and disc size as proxies for lighter or darker skin, with indeterminate cases dropped and remaining labels predicted by a separate network. Referable darker-skin cases were withheld from training yet retained in a balanced 400-image test set. Baseline ResNet50 accuracy was 73.0% for lighter-skin versus 60.5% for darker-skin images (P = 0.008). Adding StyleGAN-generated images of the missing subgroup narrowed that gap to 7.5% and 0.5% for the two debiasing methods, and darker-skin sensitivity rose from 35% to 58% with the latter, described as a proof of concept.

*This summary was generated by AI and is published unedited. Oku does not alter these summaries. It may contain errors or omissions and is provided for general informational purposes only. Oku does not guarantee its accuracy, completeness, or reliability. For authoritative information, please refer to the original, peer-reviewed article.