AI-supported dermatology: Now for darker skin tones too, thanks to a new data set

(Press-News.org) In many countries in Africa, up to nine out of ten children suffer from a skin problem, and there are far too few local dermatologists. Artificial intelligence could help with diagnosis, but needs to be trained with the relevant images, so researchers have created a new data set for dark skin tones.

Demand is high, the lack of dermatologists acute: in many countries in Africa, there is less than one dermatology specialist per one million people – compared to the World Health Organization (WHO) recommendation of one specialist per 50,000. This lack of specialists is noticeable in rural Africa in particular, with up to 87% of children suffering from untreated skin diseases.

The PASSION project (abbreviation for Pediatric AI Skin Support In Outreach Nations) has been set up with the aim of remedying this problem: a team of researchers from the University of Basel, led by Professor Alexander Navarini, worked with colleagues from Madagascar, Malawi and Guinea to create a foundation for using artificial intelligence (AI) to support dermatological diagnostics in these regions. They are presenting the project at the MICCAI 2024 conference (International Conference on Medical Image Computing and Computer Assisted Intervention) in Marrakesh.

Lack of images of dark skin tones

If AI is to detect eczema from a photo, for example, it first needs to learn what eczema looks like based on hundreds of photos. But the existing photos are primarily of skin problems in light skin types, which have been provided as documentation by clinics in Europe and the USA. The medical shortage in a lot of countries in Africa also means that there is a lack of image material for skin problems in pigmented skin. If AI programs are only trained using photos of light skin, they may be much less effective at diagnosing changes on darker skin tones.

The researchers have therefore created a database of images of very common skin diseases: eczema, fungal infections of the skin and nails, scabies and superficial skin infections with streptococci or staphylococci. This data set can be used to train new AI programs for dermatological diagnostics, but also to test existing AI models for accuracy.

The images were taken, with the patients’ consent, by local dermatologists in Madagascar, Malawi and Guinea from 2020 to 2023. The images were annotated to include the diagnosis and information on the age, gender and body part, and then entered into the anonymized database. The database now contains over 4,200 images of skin changes in around 1300 patients, two thirds of whom are under 18.

Self-diagnosis using a smartphone

“Our vision is that each patient will be able to take a photo of their skin problem themselves free of charge using a smartphone and then upload it. They will then receive a treatment recommendation from AI,” says Navarini. If this method is as accurate as it is hoped, it will be possible to use it in triage and, where appropriate, for initial treatment. Human dermatologists would only be involved if the complaint persists.

“We are currently testing the method step by step as part of a validation study in Madagascar. Once diagnostic accuracy exceeds 80%, we intend to offer the new diagnostic tool with scientific monitoring,” explains Philippe Gottfrois, doctoral student in Navarini’s research group and lead author of the study.

In the next steps, the researchers aim to expand the database to include additional image material, primarily of neglected tropical skin diseases. They hope that AI will be able to narrow the large gap in dermatological care in these regions.

END