Training AI on Diverse Skin Tones: Fitzpatrick I-VI

As of October 1, 2024.

When a convolutional neural network (CNN) learns to spot melanoma, it learns from images. If those images skew toward lighter skin tones, the model will too. That is not a hypothesis — it is a documented pattern across published dermatology AI (artificial intelligence) literature. The Fitzpatrick skin phototype scale gives researchers a concrete way to measure and correct that imbalance.

This article walks through the Fitzpatrick I-VI classification, explains how machine learning (ML) training pipelines use it, and outlines what dataset diversity actually requires in practice.

What is the Fitzpatrick scale?

Dermatologist Thomas Fitzpatrick introduced this classification in 1975, and it remains the most widely cited skin typing system in both clinical dermatology and computational imaging research. Short answer: The Fitzpatrick scale is a six-category system that groups skin by its response to ultraviolet light — whether it burns, tans, or both. Each type describes a measurable photobiological response, which makes it more useful in AI contexts than self-reported race or ethnicity, because the response pattern correlates with melanin concentration and can be estimated from standard clinical photographs. The scale was originally designed to guide safe UV dosing in phototherapy; its adoption in dermatology AI research came later because it offered a standardized vocabulary for labeling skin diversity in image datasets.

Type	Skin Tone	Hair	Eyes	Burn Behavior	Tan Behavior
I	Very fair, pale white	Blonde or red	Blue, green, or grey	Always burns severely	Never tans
II	Fair white	Blonde or red	Blue, green, hazel	Usually burns	Tans minimally
III	Medium white to olive	Blonde, brown	Hazel, brown	Burns moderately	Tans gradually
IV	Olive to moderate brown	Dark brown	Dark brown	Burns minimally	Tans easily
V	Brown to dark brown	Dark brown or black	Dark brown	Rarely burns	Tans deeply
VI	Deeply pigmented dark brown or black	Black	Dark brown or black	Almost never burns	Tans very deeply

For AI training purposes, dermatologists often group these into three broader bands — light (I-II), medium (III-IV), and dark (V-VI) — to simplify stratification targets. Many researchers argue the six-type breakdown should be preserved because the photobiological differences between Type III and Type IV produce meaningfully different imaging characteristics under standard clinical lighting.

Why does skin tone data diversity matter for AI?

Training data shapes every prediction a model makes. Short answer: AI models generalize to the population they were trained on, and most public dermatology datasets contain a large majority of Types I-III images, which means models perform worse on darker skin tones. A 2022 study on disparities in dermatology AI performance by Daneshjou et al. examined 21 dermatology AI studies and found that only 7 reported patient skin tone at all. Among those that did, the proportion of Fitzpatrick Types V and VI images averaged below 10 percent. Models that see almost no dark-skin training examples during development will have lower sensitivity and specificity when deployed on those patients. Skin conditions like seborrheic dermatitis, post-inflammatory hyperpigmentation, and melanoma all present differently on darker skin, and a model that has rarely seen those presentations will miss them more often.

This is not a minor technical footnote. The Canadian Dermatology Association (CDA) notes that darker-skinned patients already face barriers to timely diagnosis; AI tools that perform poorly on those same patients amplify existing inequities.

Health Canada's 2023 guidance on AI/ML-based Software as a Medical Device (SaMD) explicitly names demographic representativeness as a requirement for training data documentation. That means any Canadian dermatology AI product seeking regulatory clearance must account for Fitzpatrick distribution in its dataset. This requirement reflects a broader pattern in regulatory science: demographic imbalance in training data is now treated as a patient safety issue, not just a model performance concern.

How do AI training pipelines use the Fitzpatrick scale?

Building a balanced dermatology dataset involves three distinct steps: label, measure, correct. Short answer: Teams annotate each training image with a Fitzpatrick type, check whether each type is represented above a minimum threshold, then resample or augment underrepresented types before training begins. The annotation step is the first technical challenge. Fitzpatrick typing from images alone is not trivial; it requires either a clinician label or an automated colorimetry step. The Individual Typology Angle (ITA) method, which calculates skin tone from CIE Lab* color space values, has become a common automated proxy. A 2025 study on automated Fitzpatrick skin tone analysis by Ulrich et al. validated ITA against clinical Fitzpatrick assignments with good agreement for Types I through IV, supporting its use in large-scale dataset annotation pipelines.

DermaDex's computer vision primer covers how CNNs process skin images in more detail. Once images are typed, training teams apply stratified sampling. A target distribution might require that Types V and VI together represent at least 20 percent of all images in a training fold. If the raw dataset falls short, teams use targeted data collection from clinics that serve more diverse populations, augmentation by adjusting brightness and color temperature within clinically valid ranges, and transfer learning to fine-tune a base model on a small high-quality dark-skin dataset after pre-training on a larger but less diverse corpus.

The ISIC (International Skin Imaging Collaboration) Archive is the largest open dermoscopy dataset and contains over 500,000 images. It remains disproportionately weighted toward lighter skin types, which is why research teams building production models do not rely on it alone.

How do I know which Fitzpatrick skin type I am?

Self-assessment is practical and widely used in clinical research. Short answer: You can determine your type by answering a short questionnaire about your skin's typical reaction to about 45-60 minutes of unprotected first sun exposure of the season, covering your natural hair color, eye color, baseline skin color in unexposed areas, and your burn-versus-tan response. Each answer scores points that map to Types I through VI. The AAD (American Academy of Dermatology) provides sun protection guidance that includes Fitzpatrick typing as a tool for safe UV exposure recommendations. For AI research purposes, self-reported Fitzpatrick type is considered adequate for dataset labeling when clinician annotation is not feasible. For clinical decision-making, a dermatologist's in-person assessment is more reliable, particularly for mixed-heritage individuals who may fall between types.

If you want to understand how skin tone affects AI-based diagnostic accuracy, knowing your own Fitzpatrick type is a useful starting point for interpreting any AI tool's reported performance metrics.

Can AI detect my skin type?

Automated classification from photographs is now technically feasible. Short answer: Yes — current AI systems can classify Fitzpatrick type from a standard photograph using colorimetric analysis, with reasonable accuracy for Types I through IV, though accuracy is lower for Types V and VI because those types remain under-represented in most training datasets. A 2025 study showing AI predicts Fitzpatrick skin type from photographs by Draelos et al. demonstrated that a well-trained ML model can predict Fitzpatrick type alongside hyperpigmentation, redness, and wrinkle severity from photographs with clinically useful accuracy when trained on a balanced dataset. That confirms the detection problem is solvable with sufficient diverse training data. The limitation is almost always the data, not the model architecture — CNNs handle colorimetric classification well.

Automated Fitzpatrick classification is now embedded in several commercial skincare apps and is used by AI dermatology platforms to route cases or calibrate confidence thresholds before presenting a differential diagnosis. The remaining challenge is collecting more verified images of Types V and VI from diverse clinical settings, which requires active partnerships between AI developers and clinics serving those patient populations.

Is the Fitzpatrick scale still relevant today?

Critics and supporters both have a point here. Short answer: Yes, though researchers increasingly supplement it with colorimetric tools like the Individual Typology Angle because the scale's six categories are too coarse to capture the full range of human skin variation, particularly at the darker end. The AAD (American Academy of Dermatology) skin cancer resource center for diverse populations while continuing to recommend it as a clinical baseline. Fitzpatrick Types V and VI cover a very wide range of darker tones while Types I through III offer more granularity for lighter skin. Some researchers advocate for scales with more categories, or for replacing categorical typing with continuous colorimetric measures. Despite these limitations, the Fitzpatrick scale remains the standard reference in clinical dermatology, regulatory submissions, and AI dataset documentation globally. The WHO (World Health Organization) and NIH both reference it in skin health guidance, making it the most practical vocabulary for communicating about skin tone diversity in medical AI.

The practical approach used by most teams building AI tools today is to use Fitzpatrick types for stratification while supplementing with ITA scores for finer-grained analysis. For Canadian patients, AI tools intended for clinical use should publish per-type performance metrics so clinicians can see exactly how a model performs across the full I-VI range before deploying it in patient care — a standard increasingly expected under Health Canada's SaMD guidance.

What is meant by Fitzpatrick skin type?

Patients and clinicians use this term interchangeably with skin phototype. Short answer: Fitzpatrick skin type is a six-category classification developed by dermatologist Thomas Fitzpatrick in 1975 that groups skin by its photobiological response to UV exposure — specifically how readily it burns versus how readily it tans. Type I skin always burns and never tans; Type VI skin almost never burns and tans very deeply. In AI and dataset contexts it functions as a demographic proxy for skin color diversity because it correlates with melanin concentration. It is not a racial classification, though it correlates broadly with geographic ancestry. The AAD skin cancer resource center covers practical applications of Fitzpatrick typing for patients, while the dermatology AI research community relies on it primarily as a stratification variable in dataset design and model validation.

Sources

• Daneshjou R et al. "Disparities in dermatology AI performance on a diverse, curated clinical image set." Science Translational Medicine, 2022. PubMed PMID 35960806. https://pubmed.ncbi.nlm.nih.gov/35960806/
• Draelos RL et al. "Artificial Intelligence Predicts Fitzpatrick Skin Type, Pigmentation, Redness, and Wrinkle Severity From Color Photographs of the Face." Journal of Drugs in Dermatology, 2025. PubMed PMID 40135957. https://pubmed.ncbi.nlm.nih.gov/40135957/
• Ulrich P et al. "Beyond Fitzpatrick: automated artificial intelligence-based skin tone analysis in dermatological patients." 2025. PubMed PMID 40537526. https://pubmed.ncbi.nlm.nih.gov/40537526/
• Fitzpatrick TB. "The validity and practicality of sun-reactive skin types I through VI." Archives of Dermatology, 1988. PubMed PMID 3377516. https://pubmed.ncbi.nlm.nih.gov/3377516/
• Health Canada. "Software as a Medical Device: AI/ML-based SaMD Guidance Document." 2023. https://www.canada.ca/en/health-canada/services/drugs-health-products/medical-devices/application-information/guidance-documents/software-medical-device-guidance.html
• American Academy of Dermatology. "Skin Cancer Resource Center." https://www.aad.org/public/diseases/skin-cancer
• American Academy of Dermatology. "Sun Protection." https://www.aad.org/public/everyday-care/sun-protection