Limitations of ITA for Skin Type Estimation Under Uncontrolled Imaging Conditions

Accurately assessing skin color diversity is essential for evaluating whether image datasets without explicit skin type labels are sufficiently diverse for training deep learning models. Traditional methods, such as the Fitzpatrick scale, categorize skin types into six discrete classes but may not fully capture the complexity of skin color and its interaction with light under varying conditions. Continuous and quantitative approaches have been proposed to better represent skin color variation, but their reliability under uncontrolled imaging conditions remains unclear.

This study evaluated individual typology angle (ITA), an image-derived color or lightness metric commonly mapped to Fitzpatrick skin type (FST) categories. A dermatologist-labeled dataset (PAD-UFES-20) was used to analyze the performance of ITA. A skin patch image dataset was derived from PAD-UFES-20, and skin color features were extracted for evaluation.

The results showed substantial overlap between ITA distributions across FST categories and poor agreement with dermatologist-assigned labels. Fixed-threshold ITA classification achieved 22.9% accuracy, below the majority-class baseline accuracy of 51.3%, and failed to correctly classify the darkest skin type (FST VI). Lighting variation, threshold misalignment, overlapping ITA distributions, and dataset imbalance contributed to unstable ITA behavior across skin types.

Under uncontrolled imaging conditions, fixed-threshold ITA did not provide reliable Fitzpatrick skin type classification and should not be interpreted as a validated skin type measurement method. These findings demonstrate that ITA is highly sensitive to lighting variation and does not reliably correspond to dermatologist-assigned FST labels. The results highlight the limitations of using ITA as a proxy for skin type assessment in image datasets.