Skin Type Determination Using Image Artificial Intelligence

Condition(s) studied

Investigational drug(s) / intervention(s)

Skin imaging and skin characteristic assessment: Participants undergo standardized clinical and dermoscopic skin imaging, skin pigmentation measurements, skin phototype assessments, photodamage assessments, and completion of questionnaires. Data are collected for the development and validation of artificial intelligence algorithms for classification of skin pigmentation, phototype, and photodamage.

Study summary

Skin color, how easily a person burns or tans in the sun (skin phototype), and the amount of chronic sun damage in the skin are important factors in skin health. These characteristics influence a person's risk of skin cancer, how skin diseases appear, how well treatments work, and how accurately doctors and artificial intelligence (AI) systems can diagnose skin conditions. However, current methods for classifying these characteristics are often imprecise and rely heavily on subjective assessments. As a result, both healthcare professionals and patients may incorrectly classify skin type, which can lead to inaccurate risk assessments and less personalized care.

This study aims to develop and validate AI algorithms that can accurately classify skin pigmentation, skin phototype, and accumulated sun damage using photographs of the skin. Unlike existing approaches, the study combines several different methods to create a more objective "ground truth" for training the AI. These methods include skin color measurements using spectrophotometry or colorimetry, assessments using the Monk Skin Tone Scale, questionnaires about sun sensitivity, and clinical evaluations by trained observers. By combining these data sources, the researchers hope to create a more reliable and scientifically robust classification system.

The study will recruit adults aged 18 years and older from several countries, including countries from all continents. Participants will complete a questionnaire about their skin, propensity to burn and sun exposure history. Researchers will then take standardized close-up and dermoscopic images of the skin on the arm and forearm, measure skin pigmentation using objective instruments when available, and assess skin phototype and sun damage. No invasive procedures will be performed, and no personally identifiable information will be collected.

The collected images and measurements will be used to train deep learning AI models. The researchers aim to develop algorithms that can classify skin pigmentation with at least 85% accuracy, skin phototype with at least 75% accuracy, and sun damage with at least 80% accuracy compared with the combined reference assessments. The algorithms will then be tested in independent datasets, including large dermatology image databases from Sweden, to evaluate how well they perform in different populations.

The study has several potential benefits. More accurate classification of skin characteristics could improve personalized skin cancer risk assessments and allow prevention advice to be tailored to individual needs. This may help identify people who would benefit from closer surveillance and stronger sun protection recommendations while avoiding unnecessary restrictions for people at lower risk. Improved classification could also enhance the diagnosis and management of inflammatory skin diseases and skin cancers, which can appear differently in people with different skin tones.

An additional goal is to address known biases in dermatology AI systems, which often perform less accurately in individuals with darker skin. By including participants with a wide range of skin tones and backgrounds, the researchers aim to contribute to the benchmarking of AI-driven medical devices wich hopefully can result in the development of fairer and more equitable AI tools.

The study involves minimal risk. Only photographs of the arm and forearm will be taken, and researchers will avoid capturing tattoos, prominent scars, or other identifying features. All data will be stored securely and only accessible to authorized researchers. The potential benefits of improving skin disease diagnosis, skin cancer prevention, and fairness in medical AI are considered to outweigh the small privacy risks associated with participation.

Eligibility

Primary outcome measure(s)

• Agreement between AI-derived skin pigmentation (tone) classification and objective skin pigmentation measured by colorimetry/ spectrophotometry (Individual Typology Angle, ITA) — At completion of model development and testing (approximately 2029).
  Skin pigmentation will be measured objectively using spectrophotometry/ colorimetry and summarized as the Individual Typology Angle (ITA). AI-derived skin pigmentation classification will be compared with ITA values using correlation and agreement analyses. ITA is considered the primary reference standard for objective assessment of skin pigmentation in the interpretation of AI performance.
• Accuracy of AI-based skin phototype classification — At completion of model development and testing (approximately 2029).
  Accuracy of the deep learning algorithm in classifying skin phototype from clinical and dermoscopic images compared with the reference standard based on the validated Fitzpatrick skin phototype assessment (consisting of 6 categories).
• Agreement between AI-derived skin pigmentation (tone) classification and clinician-assessed Monk Skin Tone Scale category — At completion of model development and testing (approximately 2029).
  Skin pigmentation will be assessed visually by trained investigators using the Monk Skin Tone Scale (categories 1 (fair) to 10 (dark)). AI-derived classifications will be compared with clinician-assigned Monk categories using agreement and correlation analyses. The Monk Skin Tone Scale represents the principal visual reference standard for skin tone classification.
• Agreement between AI-derived photodamage classification and the Clinical Photonumeric Scale for Photodamage Assessment — At completion of model development and testing (approximately 2029).
  Photodamage will be assessed using the validated Clinical Photonumeric Scale (0-3 for 3 defined pigmentation categories) for Photodamage Assessment. AI-derived photodamage classifications will be compared with the photonumeric scale scores using agreement and correlation analyses. This outcome evaluates the agreement between AI-derived classifications and a validated photonumeric clinical assessment of photodamage.

Trial sites (11)