Ensuring skin type diversity is essential for AI‐based skin analysis applications, as it affects the accuracy of tasks such as detecting skin conditions and evaluating treatment outcomes [1, 2]. Studies like Gender Shades have highlighted how disparities in dataset representation can lead to significant performance gaps in AI systems, particularly for underrepresented groups such as individuals with darker skin types [3]. Reliable measurement of skin type diversity is crucial for creating datasets that allow AI models to generalize across different skin types. In dermatology, the Fitzpatrick skin type (FST) is the most used system to classify skin type, providing a standard for healthcare professionals and aesthetic practitioners [4]. This system involves a face‐to‐face meeting with dermatologists [5, 6]. The Fitzpatrick scale, which categorizes skin based on its color and reaction to initial sun exposure resulting in erythema (burning or tanning), was originally developed by Thomas B. Fitzpatrick in 1975 [7]. This classification system is subjective, as it relies on self‐reporting or dermatologist assessment rather than quantitative metrics. These six skin types are shown in Figure 1.

An alternative extension to the Fitzpatrick classification is the Monk Skin Tone (MST) [8] scale, a 10‐point system developed by Ellis Monk to more comprehensively represent the diversity of human skin tones in fairness‐focused AI applications. However, like Fitzpatrick, it discretizes the continuous skin color spectrum into fixed categories and therefore cannot fully capture the nuanced variations observed in real‐world skin colors.

Prior research has shown that using automated methods, skin type assessment can be done directly from images [9, 10]. However, there are some limitations when detecting exact skin types using automatic methods; for instance, the methods face challenges in determining skin type in images taken under different lighting conditions and may not perform well for all skin type groups. Therefore, a comprehensive review of existing skin type measurement methods used in previous studies was conducted to identify which methods were most frequently used and whether the limitations were considered for each method.

An overview of studies that have used different skin type measurement methods is presented in Table 1. For example, the importance of accounting for varying lighting conditions in skin type measurements lies in its direct impact on skin color perception, texture representation, and the overall quality of captured images. Fitzpatrick labels can also give insight into the different skin types utilized to develop these methods. Table 1 shows a comparative overview of skin type classification and measurement methods used in peer‐reviewed publications. Papers were selected as follows: A systematic search was performed on the PubMed, IEEE Xplore, and Google Scholar databases using the search terms “Fitzpatrick skin type”, “illumination variation”, and “skin type measurement”. The comparison included only peer‐reviewed journal papers and peer‐reviewed conference papers from those search results. The investigation revealed that various methods are used to measure skin type, ranging from physical devices like spectrophotometers to computational approaches, including AI‐based methods, ITA (individual typology angle), and others.

As Table 1 shows, ITA is the dominant skin type measurement method in existing literature. This method measures skin pigmentation in the L^*a^*b^* color space, while the discrete and subjective Fitzpatrick categories are based on the pigmentation and response to UV exposure. Although ITA provides a continuous image‐derived representation of skin color, it can also be mapped onto discrete categories, like Fitzpatrick classification system. Some previous research has treated ITA classifications as synonymous with those of Fitzpatrick [65, 66]. ITA is not a classification method itself, but it is a feature used in skin type assessment, representing the angle between the color and luminance vectors of a person's skin color in the L^*a^*b^* color space, to estimate image‐derived skin color characteristics based on the calculated angle.

However, other existing skin type measurement methods, as outlined in Table 1, are physically based techniques, which require in‐person assessment and are therefore not suitable for evaluation through digital images. For example, skin reflectance estimation based on dichromatic separation (SREDS) was proposed to measure skin color from a single image by decomposing the light reflected from the skin into two components [38]. Three datasets, the CMU Multi‐PIE face database, MEDS‐II, and Morph‐II—were used to evaluate the superior performance of the new model compared to ITA and Relative Skin Reflectance (RSR). Although this method outperforms other methods in uncontrolled lighting conditions, it still has difficulties with variations in lighting conditions. In [43], face area lightness measures (FALMs) are calculated to estimate skin color from images using color analysis or FSTs. The estimated FALMs from the images are then compared to ground truth skin measurements from a calibrated dermatological device specifically designed to measure skin lightness, showing a large difference between these two measures resulting from camera, background, and environmental conditions.

In the case of deep learning‐based methods, which typically require large, diverse datasets and often do not use reliable or consistent Fitzpatrick labels. In many prior works, dermatologist‐assigned FST labels were either unavailable or generated using non‐expert annotators, which undermines the reliability of comparisons. Chang et al. [30] introduce a convolutional neural network model to classify their created skin spectra dataset into the six FSTs, assessing sun exposure reactions. Their performance evaluation includes a comparison with an artificial neural network model. Mbatha et al. [61]. developed a neural network model trained on human‐annotated data under 15 different lighting conditions to predict MST values with improved robustness. Their work demonstrates how modern machine learning techniques can produce consistent estimates across variable environments. However, such approaches typically require large, well‐labeled datasets and careful control of image acquisition, which may not always be available in real‐world or legacy image collections. The MST labels used for training and evaluation were manually assigned by the participants themselves. The classification of facial features extracted from images into different skin types of normal, oily, dry, and combination is accomplished using a support vector machine with the use of discrete wavelet transform, contrast, and local binary pattern [34]. A multi‐task cascaded convolutional neural network is used to classify skin images into skin types, oily or dry [40]. In another study, a biological approach is adopted, focusing on the detection of skin type through the assessment of Sebum levels [23].

Corbin and Marques [52] used ITA with four different approaches to estimate FST in three skin lesion image datasets: ISIC 2018, ISIC 2017, and ISIC 2016 [67], none of which provide information on skin types. Their study compares different methods for ITA‐based estimation and skin type classification, highlighting the impact of skin lesion size on their results. They proposed new methods based on AI models and ITA to address the lack of skin type information in skin image datasets lacking this metadata. Their research aims to address the underrepresentation of darker skin types in public skin lesion image datasets to produce fair models.

Kinyanjui et al. [31] used ITA to assign FST categories in other dermatology datasets that do not include Fitzpatrick skin type information: SD‐198 [68] and ISIC‐2018 [67]. They highlighted the limitation of the underrepresentation of darker skin populations in publicly available skin disease datasets. The skin type classification results showed that the majority of images belonged to lighter skin types. Campiche et al. [33]. used ITA and IWA_Newtone [69] to show the significant difference in skin type between these two groups of lighter‐ and darker‐skinned women. These two methods were compared in terms of skin color and signs of aging using the ColorFace imaging system. IWA_Newtone was developed by Newtone Technologies and incorporates components of luminance and chrominance in its assessment. There was a lack of some skin types, including lighter skin types and darker skin types.

As it was highlighted in the study by Alipour et al. [70]., the availability of large and diverse datasets is a major challenge in skin type research. There are two publicly available FST‐labeled datasets (e.g., PAD‐UFES‐20 and Fitzpatrick17k) that exhibit significant class imbalance and potential bias; using them to train deep learning models may reproduce and reinforce these biases. As a result, applying such models to these datasets may unintentionally propagate biased labeling and worsen the fairness issues the field aims to address. However, dermatologist‐assigned FST labels derived from the PAD‐UFES‐20 dataset were used in our study, providing a more consistent basis for evaluation.

According to the explanations above regarding skin type measurement methods, most existing literature either overlooks the influence of varying lighting conditions on their results or fails to assess their methodologies across all six skin types. Therefore, it is necessary to assess the performance of skin type measurement methods in these challenging scenarios to understand how they respond to such conditions and identify their weaknesses. Among these methods, ITA has emerged as a widely used image‐derived color/lightness metric commonly mapped to Fitzpatrick skin type categories, particularly in studies involving image datasets without expert‐assigned skin type labels. However, its reliability under uncontrolled lighting conditions and imbalanced data remains insufficiently understood. Our study does not introduce a new measurement method but rather investigates ITA's limitations in accurately measuring skin type diversity and offers new insights into its practical performance. Unlike previous studies that addressed general biases and proposed broad solutions in skin type measurement methods, our work delves into the limitations of this method itself on an image dataset with Fitzpatrick labels. It is investigated whether ITA can reliably characterize dataset‐level skin color variation under uncontrolled imaging conditions. Due to the limited availability of datasets with expert‐assigned Fitzpatrick skin type labels, our study focused on PAD‐UFES‐20, which also offers clear separation between lesions and surrounding healthy skin, facilitating reliable skin patch extraction. This level of detailed analysis has not been a focus of previous research.

The major contributions in this work include:

A systematic review of skin type measurement methods, highlighting the widespread use of ITA as a proxy for image‐derived skin color in prior research.

The performance of ITA for the skin patch image dataset is evaluated using the method shown in Figure 2. This evaluation includes analysis of the distribution of ITA values across the dataset and within all six FST categories, as well as a comparison between ITA‐based predictions and dermatologist‐assigned FST labels. Although ITA is not inherently a skin type classifier, it is analyzed in a classification‐like setting to examine its behavior when mapped to Fitzpatrick categories. The dataset's color and lightness distributions were analyzed to gain insights into the variations of color and lightness across different skin types in the images. The approach involves several key steps, including dataset preparation, color space conversion, and ITA‐based FST classifications.

The experiments were conducted using Python 3.11.0 and various libraries such as OpenCV and Matplotlib. The skin patch image dataset was outlined in PNG format with image dimensions ranging from 59×67 to 2929×593. In section 3.1, the analysis of images in terms of color and lightness is assessed. In Section 3.2, the overall distribution of ITA values is analyzed to highlight skin color imbalances in the dataset. Section 3.3 examines ITA distributions within each Fitzpatrick skin type (FST) category. Section 3.4 explores the influence of ITA parameters (L^*, b^*) on their values. Section 3.5 compares ITA predictions with expert‐assigned FST labels using a confusion matrix. Finally, Section 3.6 presents examples of misclassified images and analyzes how lighting variation impacts ITA‐based classification accuracy. The code and dataset can be found here: https://github.com/NedaAlip/Assessment‐of‐skin‐type‐measurement‐method/tree/main

Our study focused on evaluating ITA's robustness and reliability as a continuous quantitative metric of skin color analysis in image datasets. Given the limited access to datasets with expert‐labeled Fitzpatrick types, PAD‐UFES‐20 was chosen for its well‐separated lesion and healthy skin areas, allowing for better skin patch segmentation. The number of images labeled as skin types V or VI in the PAD‐UFES‐20 dataset (and hence in our skin patch image dataset) is very low. The distribution of ITA values across lighter skin types (higher ITA values) and darker skin types (lower ITA values) reflects the over‐representation of light skin types and under‐representation of dark skin types. Also, the skin images in PAD‐UFES‐20 present a wide range of lighting conditions, including shadowed areas, uneven lighting, and local overexposure. The results indicated that while the ITA method provides measurable differences in lightness across the dataset, its ability to assign FST categories accurately based on FST was limited not only by dataset imbalance, but also by substantial overlap between ITA distributions, predictions misalignment, lighting sensitivity, and failure to correctly classify the darkest skin type (FST VI).

The histogram distribution of ITA values for each FST showed that ITA does not conform well to the predefined FST ranges, particularly for darker types. The broader range of ITA values compared to the FST thresholds suggests that ITA produces broader continuous variation than discrete FST thresholds. Lighting variations further exacerbate ITA‐based classifier's limitations, as shown in Figure 15 and Figure 16, where the same skin type appears differently under different lighting conditions.

Figure 15(a)) illustrates that under varying lighting conditions; different lighter skin types may exhibit similar reflection and absorption properties. This results in more overlapping values within ITA method's parameter space, leading to misperceptions of true skin color and confusion with other light skin types. The higher rate of misclassifications between adjacent lighter skin types I and II in the confusion matrix highlights this issue. Figure 15(b)) illustrates why ITA‐based method classifies the same skin types across different thresholds, instead of classifying them within the same threshold range. It suggests that ITA‐based classifier is primarily influenced by changes in lightness (L^*), although it is also affected by variations in chrominance (b^*), depending on the extent of those variations. Figure 16(a)) and Figure 16(b)) illustrate how lighting variations can alter the perceived skin color, leading to misclassification of skin types. Accurate lighting calibration and standardized imaging conditions are essential to minimize such inconsistencies and improve the reliability of skin type measurement methods.

Figures 15 and 16 demonstrated how lighting conditions, such as overexposure, shadowing, and uneven lighting conditions, alter the perceived color and lightness of skin, leading to misclassification. Lighter skin types (I and II), which reflect more diffuse light, exhibited relatively stable ITA values under variable lighting. In contrast, darker skin types, which absorb more light due to higher melanin concentration, showed greater variability in ITA measurements, particularly in the L^* channel of the L^*a^*b^* color space [77]. This confirms that ITA is disproportionately affected by lighting when evaluating darker skin types.

The results also indicate that lighting variations disproportionately affect ITA measurements for darker skin types (V‐VI). Because ITA is driven by the L^* component of the L^*a^*b^* color space, any reduction in lightness caused by shadows, uneven lighting, or optical absorption is amplified in darker skin, where L^* values are already low. As a result, small lighting changes can cause large shifts in ITA values, making the method less stable for highly pigmented skin. This instability is clearly visible in Figure 11 (e–f), where the ITA histograms for types V‐VI show wide dispersion and poor alignment with the expected threshold ranges. These findings highlight that ITA mapping is inherently less robust for darker skin because reduced reflectance and higher melanin absorption magnify lighting‐dependent fluctuations in L^*. Results show that ITA is more sensitive to lighting variability in darker skin types, which limits its reliability when used in uncontrolled imaging environments.

The lightness histogram exhibits a broader distribution, reflecting variability in lighting conditions and highlighting ITA's greater sensitivity to lightness over chrominance. The low number of data points in the bins for very dark and very light values suggests the need for more samples to improve the analysis of these skin types. Table 2 indicates a clear dataset imbalance, with very few samples for skin types V and VI. This limits the strength of the conclusions that can be drawn for darker skin types and reduces the reliability of distributional estimates for these groups. However, the unstable ITA distributions observed for darker skin types should be interpreted as the combined result of small sample size, lighting sensitivity, overlapping ITA distributions, and mismatch between fixed ITA thresholds and dermatologist‐assigned FST labels, rather than as a consequence of dataset imbalance alone. As shown in Table 4, skin types V and VI exhibit wider 95% confidence intervals compared to lighter skin types. This reflects the combined effect of small sample sizes and greater lighting sensitivity in darker skin types. The broad intervals indicate that ITA‐based predictions for darker skins are less stable and less reliable. Also, overlapping ITA values across different FSTs in the histogram showed that FST skin type classification based only on ITA is not reliable.

Lightness also decreases in the darkest skin type of VI, while chrominance values show little variation, indicating limited color diversity and nuanced differences in representation between different skin types in the dataset. This lack of color diversity reflects the inherent similarity in chrominance values across all skin types, rather than the dominance of any specific skin type in the dataset.

ITA values were compared against Fitzpatrick labels assigned by dermatologists, which are discrete categories based on in‐person assessment. ITA showed very low classification accuracy when evaluated against subjective and discrete dermatologist‐assigned FST labels, which were used as the ground truth for the analysis. It does not accurately classify images into actual skin types as determined by dermatologist‐assigned labels. The inherent difference between these two continuous and discrete methods limits the direct alignment between ITA and FST. ITA's limitations in aligning with predefined FST thresholds highlight the need for alternative or enhanced metrics that are both robust to lighting changes and sensitive to subtle color variations, especially in darker skin colors. It shows that relying on fixed, discrete thresholds (such as those used in the Fitzpatrick system) may not be sufficient for identifying nuanced skin color variation. This limitation is particularly relevant for dataset fairness assessment, as underestimation or misclassification of darker skin types can cause bias in AI models. This highlights the need for alternative or complementary approaches capable of capturing the complexity of skin color diversity in image datasets.

Beyond ITA, other approaches have been proposed for skin type assessment. Physical and colorimetric methods, such as spectrophotometer‐based melanin or erythema indices, can provide accurate and lighting‐independent measurements; however, they are impractical for large‐scale image datasets due to their cost, acquisition time, and the need for controlled measurement conditions. More recent AI‐based methods may provide flexible and data‐driven approaches for skin color estimation; however, their performance also depends on the availability of sufficiently diverse and reliably labeled datasets, particularly for underrepresented darker skin types. These limitations help explain why ITA has been commonly used in prior work; however, its application should be treated with caution, particularly under varying lighting conditions, and only with explicit awareness of its inherent limitations, rather than as a standalone solution.

Threshold recalibration, lighting normalization, incorporating greater diversity, particularly within underrepresented skin colors, and preprocessing approaches may be explored in future work; however, the current results do not demonstrate that such modifications would be sufficient to achieve reliable Fitzpatrick skin type classification. PAD‐UFES‐20 was the only publicly available dataset that provided both dermatologist‐assigned Fitzpatrick labels and high‐resolution skin lesion images suitable for isolating skin regions. This issue restricted our ability to evaluate ITA across multiple datasets. It is considered a limitation, and it should be addressed in future work through collecting additional samples from darker skin types to ensure more balanced representation. Controlled lighting conditions and lighting normalization techniques would be valuable in future evaluations to minimize lighting‐related variability. Several preprocessing techniques have been proposed to address lighting variability in skin image analysis, including color constancy algorithms [78, 79], histogram equalization [80], and GAN‐based illumination normalization [81]. These methods may help reduce lighting artifacts and may improve the consistency of color‐based metrics such as ITA. In future work, integrating such preprocessing steps before ITA computation may allow for more robust and fair assessments.

Our comparison between ITA‐based predictions and dermatologist‐assigned FST labels revealed inconsistencies, largely due to dataset limitations such as class imbalance, rigid Fitzpatrick thresholds, and the sensitivity of ITA to lighting conditions. Although ITA provides a continuous representation of skin color, this is insufficient for reliable alignment with discrete dermatologist‐assigned FST labels. These results indicate that ITA is not suitable for reliable FST classification under real‐world imaging conditions. While ITA provides continuous image‐derived color representation, it cannot reliably characterize dataset‐level skin color diversity or fairness under uncontrolled imaging conditions. To fairly evaluate the performance of that, future work will focus on collecting a diverse dataset with controlled lighting conditions to support accurate evaluation. Such a dataset may support future investigation of threshold recalibration and evaluation of alternative continuous skin color metrics. Also, new ITA thresholds designed specifically for darker skin types should be explored, and ITA should be compared against emerging metrics and spectrophotometric ground‐truth measurements. Finally, using a calibrated imaging system would enable more reliable cross‐dataset comparisons and support the development of lighting‐aware skin tone assessment models. As part of future work, a new dataset has been designed as part of this work, with continuous skin color labels obtained using a colorimeter as a reference, along with controlled and well‐documented lighting conditions.

ITA has been used in dermatology and AI‐based skin analysis studies as an image‐derived skin color or lightness metric commonly mapped to Fitzpatrick categories [35, 50]. Recent studies also highlight the persistent under‐representation of darker skin in dermatology datasets and AI‐generated images [1, 82]. Our findings support these concerns, showing that reliance on ITA alone may introduce systematic bias, particularly when datasets lack adequate dark skin representation. The results further demonstrate that ITA is unreliable for FST classification under uncontrolled imaging conditions and should not be interpreted as evidence of reliable skin type assessment performance.

To assess skin type diversity in datasets used to train AI models and for capturing the detailed color gradations of different skin types, the Fitzpatrick labeling system, based on fixed and rigid thresholds, may not be sufficiently informative. Instead, there is a need to investigate more robust and validated approaches capable of capturing subtle variations in skin color and provide a more nuanced understanding of dataset diversity. The findings suggest that rigid categorical classifications alone may be insufficient for representing the full variability of skin color in image datasets. Given the importance of image‐derived skin color assessment methods in detecting any under‐representation within datasets, it is essential to ensure their performance under varying lighting conditions. This may help identify under‐representation before datasets are used to train AI models to ensure fair and unbiased performance of these models across all skin colors. The results indicate that fixed‐threshold ITA does not provide sufficiently reliable performance for FST classification under uncontrolled imaging conditions. It emphasizes the importance of developing and validating more robust approaches for image‐based skin color assessment, particularly under variable lighting conditions and across diverse skin colors.

This publication has emanated from research supported by a grant from Science Foundation Ireland under Grant number 18/CRT/6222.

This study did not involve any new data collection involving human participants. All datasets used (PAD‐UFES‐20 and Fitzpatrick17k) are publicly available. https://doi.org/10.13039/501100001602

The authors declare no conflicts of interest.