A team at Technical University of Munich develops an AI pipeline combining DensePose and OpenFace to compute Individual Typology Angle (ITA) from CIELAB color values, automatically mapping images to Monk and Fitzpatrick skin tone scales for teledermatology and clinical research.
Key points
DensePose and OpenFace segment forearm and nasal bridge pixels, convert RGB to CIELAB, and compute mean ITA per image.
ITA values map to Monk (10-tone) and Fitzpatrick (6-type) scales via established thresholds, offering continuous-to-categorical classification.
Algorithm achieves 89–92% accuracy on clinical images with balanced accuracy of 66–68% on Monk scale, while Fitzpatrick performance remains below 20%.
Why it matters:
This approach standardizes skin tone assessment, enabling inclusive teledermatology diagnostics and large-scale epidemiological studies across diverse populations.
Q&A
What is the Individual Typology Angle?
How do DensePose and OpenFace aid skin tone analysis?
What distinguishes the Monk Skin Tone Scale?
Why does the algorithm perform better on AI-generated images?
Read full article
Academy
CIELAB Color Space
CIELAB is a three-dimensional color model developed by the International Commission on Illumination (CIE) to represent all perceivable colors. It separates color information into three coordinates: L* for lightness (ranging from 0 for black to 100 for white), a* for the green–red axis, and b* for the blue–yellow axis. Unlike RGB or CMYK, CIELAB is designed to approximate human vision, making equal numerical changes correspond to similar perceived color shifts. This uniformity is critical for scientific applications such as skin tone measurement, where subtle pigment differences must be quantified consistently.
Individual Typology Angle (ITA)
The Individual Typology Angle (ITA) is a standardized metric that translates CIELAB values into a continuous skin tone variable. ITA is calculated by taking the arctangent of the ratio ((L* – 50)/b*) and converting from radians to degrees: ITA = arctan((L* – 50)/b*) × 180/π. A higher ITA indicates lighter skin tones, while lower values correspond to darker tones. By mapping each pixel’s CIELAB data into a single angle, ITA offers an objective, reproducible measure that overcomes the subjectivity of visual scales.
Skin Tone Classification Scales
Skin tone assessment often uses categorical scales for clinical and research contexts. The Fitzpatrick scale defines six types based on UV sensitivity and tanning behavior, but its uneven categories limit representation of darker tones and have inherent subjectivity. The newer Monk Skin Tone Scale comprises ten evenly spaced categories derived from ITA thresholds, explicitly aiming for inclusivity across the global spectrum. Each category covers an equal ITA interval, reducing bias and offering finer resolution for diverse populations.
Automated Skin Tone Analysis with AI
Recent advances integrate computer vision and color science to automate skin tone classification. Frameworks like DensePose use convolutional neural networks to segment anatomical regions—in this case, the lower forearm—by generating labeled masks. Similarly, OpenFace identifies facial landmarks and isolates the nasal bridge. Segmented pixels convert from RGB to CIELAB and then into ITA, allowing for fully automatic mapping into categorical scales. This AI pipeline minimizes manual intervention, speeds up large‐scale image analysis, and supports teledermatology and personalized treatment by ensuring standardized, objective skin tone data.
Applications and Implications
- Teledermatology: Enables remote diagnosis and monitoring across skin tones without manual input.
- Epidemiology: Facilitates large‐scale studies on disease prevalence by skin tone, helping address health disparities.
- Personalized Medicine: Supports treatment planning that accounts for pigment‐based variations in drug response and dermatological presentation.
Key Takeaway: Combining AI segmentation with CIELAB‐derived ITA and an inclusive categorical scale leads to robust, bias‐resistant skin tone classification suitable for diverse clinical and research settings.
