baroodya / skin-tone-bias Goto Github PK

Introducing our project

• Existing dermatology textbooks and literature: underrepresentation of skin conditions on patients w darker skin
• Leads to late diagnoses and 2-3x higher mortality rates for people of color
• Intersection of AI technology and dermatology for accessible solutions to early diagnoses… bUT concern that underrepresentation carries over to datasets.
• This is confirmed but does it necessarily mean that algorithms will certainly perform worse on darker skin?

Most exciting/interesting find

• Yes….but actually no. Preliminary data shows little disparity between lighter and darker skin types. Type VI actually had highest accuracy for 5 of 6 experiments
• However, after more digging, this is likely due to a small data set with a label imbalance
• Could be remedied by (1) more hand labeled data or (2) data augmentation (which we tried and got promising results in terms of reducing disparities, but needs to be explored further)