Inclusive by design: Why we must rethink generative <scp>AI</scp> in dermatology

Abstract

Generative AI is a transformative technology with potential uses in health care diagnostics, research and education.1 Specifically, within the field of dermatology, text-to-image generative AI can produce images of skin conditions. However, the range of human skin colours depicted and the accuracy of the generated images remain unassessed. In this issue of the J Eur Acad Dermatol Venereol, Joerg et al.,2 evaluated four AI text-to-image generators (Adobe Firefly Image 3 Model; ChatGPT-4o (DALL-E3); Midjourney Model Version 6; Stable Diffusion 3) over a 24-day period. They assessed the accuracy of dermatological images generated and the diversity of skin tones depicted in relation to text prompts for 20 of the most common skin conditions. Two independent dermatology trainees used the Fitzpatrick skin phototype (FSP) classification system to assess the diversity of skin colours: FSP I–IV and >IV representing light and dark skin, respectively. A separate set of dermatology trainee evaluators assessed the accuracy of the AI-generated images with respect to clinical diagnosis. The researchers found that there was poor accuracy and a lack of skin colour diversity in the dermatological images generated by these four leading AI text-to-image generators. While there are some limitations to the methodology, including the use of FSP to evaluate skin colour and the nature of the dermatoses evaluated, the findings highlight the lack of inclusive datasets used to train these generative AI systems. This mirrors the observed underrepresentation of images of darker skin colours in dermatological textbooks and training resources.3 Moreover, visual FSP assessments are not truly representative of dermatoses seen in those with darkly pigmented skin. It is imperative that leading technology companies seek to curate and use image datasets that are inclusive and representative of the full spectrum of human skin colours seen across the globe. The authors do acknowledge this in comparing the generated data with the distribution of the U.S. ethnic demographic, but this is really insufficient. Objective human skin colour scales may be employed to categorize the skin tones of images used in training datasets. The FSP, although applied in this context, was originally designed as a subjective tool for evaluating skin responses to ultraviolet exposure rather than for this specific application.4 The Monk Skin Tone Scale5 is a new subjective scale for classifying human skin colour, advocated by the technology community. The database used to choose skin colour categories in this scale remains unknown. Furthermore, Monk himself has noted that this scale was developed with the skin colour variations occurring in the Americas in mind. This may limit its relevance or effectiveness when applied to populations worldwide. Other potential objective scales include the Individual typology angle (ITA) and Eumelanin Human Skin Colour Scale (EHSCS).4 The former (ITA) method uses CIE L*a*b* colorimetric parameters, but its categories favour lighter skin tones.4 The latter EHSCS approach is based upon published skin reflectance data but has not been validated as yet.4 Currently, text-to-image generative AI faces limitations, including challenges in accurately producing clinical images and representing the full range of human skin tones. This is a direct result of the lack of inclusivity of the training datasets used for these systems. Dermatologists and the public should be aware of the limitations of generative AI, exercising caution in its use and interpretation. OED: I am a co-author of the publication of the Eumelanin Human Skin Colour Scale referenced in the manuscript. I am also the Principal Investigator and Funder of phase 2 data collection in relation to the Eumelanin Human Skin Colour Scale. RAS: I am a co-author of the publication of the Eumelanin Human Skin Colour Scale referenced in the manuscript. Data sharing is not applicable to this article as no new data were created or analysed in this study.

Ähnliche Arbeiten