Skin of Color Underrepresentation Highlighted in Review of Artificial Intelligence Use in Dermatology

Rebecca Fliorent

Credit: LinkedIn

Issues with the integration of artificial intelligence (AI) for assessment and diagnosis of dermatologic conditions persist for patients with skin of color, according to recent findings, with struggles observed in identification of lesions.¹

These findings and more were the conclusions of a recent literature review, conducted with the purpose of pointing to the issues leading to lack of representation of patients with diverse skin tones within datasets implemented by algorithms.²

This review was led by Rebecca Fliorent, from the Rowan-Virtua School of Osteopathic Medicine in Stratford, New Jersey. Fliorent and colleagues noted the value in improving diagnostic accuracy for as many different skin types as possible.

“The objective of this review is to identify and address gaps in current AI usage in dermatology, specifically regarding individuals with (skin of color), and to propose future steps to enhance its application in dermatologic practice,” Fliorent and colleagues wrote.

Background and Methods

The investigators noted that AI employs sophisticated algorithms and models to learn from different types of information, looking at different patterns with the potential for informed decision-making. The use of AI among dermatologists has been suggested for its implementation for early skin cancer detection, for example.

In a more general sense, the investigators expressed that use of AI had shown promise in providing personalized treatment recommendations by allowing for patient-specific data such as patients’ symptoms, medical history, and responses to different treatments. Thus, AI may lead to more precise diagnostics and improve outcomes among different types of patients.

To begin identifying possible gaps in use of AI for individuals with skin of color, the research team conducted their comprehensive review of existing literature. They implemented the PubMed and Google Scholar research databases.

The team’s review took place in the timeframe between February 2002 - June 2023 and included a wide array of search terms including racial representation, AI, skin cancer, artificial intelligence, skin of color, dermatology, pigmentation, dermatologic screening, disparities in public health, and melanoma.

The investigators’ criteria for research literature inclusion in their study included articles which are written in English and a variety of different types of research like clinical trials, systematic reviews, case reports, and single-center studies done retrospectively. The team excluded studies not matching these descriptions from their review, also working to ensure relevance and reliability in their findings.

Findings

A set of different studies were identified by the investigators over the course of their research, and several of these studies proved to have invaluable information on the topic.

One trial showed that publicly accessible skin image datasets had been limited when they were applied to clinical settings in the real-world, as far as representation of diverse skin tones. Such limitations were described by the research team as stemming from lighting issues, focus accuracy, levels of exposure, aperture, alignment of backgrounds, and the variability of camera shutter speed.

Another study the team looked at had highlighted the failures of AI imaging investigations in skin of color as far as correctly addressing several different CLEAR Checklist elements. The study indicated a lack of skin color information in numerous studies’ processing, given that the regions of patients’ lesions had been specified in under half of the cases.

The investigators were able to identify 10 investigations and 15 AI technologies assessing AI's efficacy in evaluation of images of diverse skin tones.

A substantial number of these investigations pointed to the issue of underrepresentation of images included in datasets, and others applied AI to patient samples that either excluded skin of color or included such patients minimally. A scarcity of representation within the identified datasets and a lack of accuracy with AI technology was pointed to by these studies.

Overall, the research team concluded that the tailoring of AI approaches may be necessary to deal with the complexities of evaluating skin conditions among individuals of diverse skin tones. The goal, they noted, should be mitigating biases inherent in AI datasets to allow for more comprehensive representation of different patient populations.

More inclusive datasets were noted by the team as potentially leading to improved dermatologic care outcomes and to reductions in disparities. The investigators also pointed to the benefits of training dermatologists to capture high-quality lesion images on patients with skin of color.

“Although the mentioned AI programs do not constitute a comprehensive list, many of these programs show promise in supporting clinical decision-making; however, they require modification to become reliable diagnostic aids for SOC patients,” they wrote. “AI companies should disclose information regarding training sets for SOC.”

References

1. ^{Fliorent, R., Fardman, B., Podwojniak, A., Javaid, K., Tan, I.J., Ghani, H., Truong, T.M., Rao, B. and Heath, C. (2024), Artificial intelligence in dermatology: advancements and challenges in skin of color. Int J Dermatol. https://doi.org/10.1111/ijd.17076.}
2. ^{Adamson AS, Smith A. Machine learning and health care disparities in dermatology. JAMA Dermatol. 2018; 154(11): 1247–1248.}