Skin Type Diversity: a Case Study in Skin Lesion Datasets

Abstract

Abstract Inadequate skin type diversity, leading to racial bias, is a widespread problem in datasets involving human skin. For example, skin lesion datasets used for training deep learning-based models can lead to low accuracy for darker skin types, which are typically under-represented in these datasets. This issue has been discussed in previous works; however,skin type diversity of datasets and reporting of skin types have not been fully assessed. Frequently, ethnicity is used instead of skin type, but ethnicity and skin type are not the same, as many ethnicities can have diverse skin types. Some works define skin types, but do not attempt to assess skin type diversity in datasets. Others, focusing on skin lesions, identify the issue, but also do not measure skin type diversity in the datasets examined. Building on previous works in the area of skin lesion datasets, this review explores the general issue of skin type diversity in datasets by investigating and evaluating skin lesion datasets specifically. The main contributions of this work are: an evaluation of all publicly available skin lesion datasets and their metadata to assess frequency and completeness of reporting of skin type and an investigation into the diversity and representation of specific skin types within these datasets.