Source/disclosure information
Issuer:
Disclosure: Koiner reports that he receives personal fees from Boston AI beyond the submitted work. See this study for relevant financial disclosures of all other authors.
Important points:
- AI models have the potential to learn race-related features and infer race from retinal images.
- This increases the risk of racially biased performance on diagnostic tasks.
Research has shown that AI algorithms may be able to infer race not only from color retinal fundus images, but also from grayscale retinal vascular maps, eliminating the possibility of humans inferring race.
The authors expressed concern about potential AI racial biases that could ultimately influence diagnostic and treatment decisions.
Image: Adobe Stock
This study included a sample selected from the Imaging and Informatics Cohort Study in Retinopathy of Prematurity, including 94 infants who were reported as black by their parents/guardians and 94 infants who were reported as white. A total of 151 infants were included. A total of 4,095 color retinal fundus images were collected to train, validate, and test 40 ResNet-18 convolutional neural network (CNN) models. All color retinal fundus images were then segmented into grayscale retinal vascular maps, which were then iteratively transformed by thresholding, binarization, or skeletonization to enable discrimination between black and white races. Potential information about blood vessel pigmentation, size and caliber was excluded.
The model had near-perfect ability to predict self-reported race (SRR) from color retinal fundus images.
“Race itself is a social construct, but it is associated with variations in skin and retinal pigmentation,” the authors write.
However, CNN was able to infer black versus white race with comparable accuracy from grayscale retinal vascular maps, which is not possible for human readers. “Predictive information was retained even in images that appeared devoid of information to the naked eye,” the authors write.
“The results of this diagnostic study suggest that it may be very difficult to remove information related to SRR from fundus photographs. As a result, AI algorithms trained on fundus photographs It can actually result in skewed performance,” they wrote.
