The Effect of Skin Tone Classification on Bias in Bruise Detection

The Effect of Skin Tone Classification on Bias in Bruise Detection | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Research Article The Effect of Skin Tone Classification on Bias in Bruise Detection Dharmi Desai, Amin Nayebi, Mehrdad Ghyabi, David Lattanzi, Katherine Scafide, and 1 more This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-6875703/v1 This work is licensed under a CC BY 4.0 License Status: Under Review Version 1 posted 12 You are reading this latest preprint version Abstract Background AI-based bruise detection tools have shown promise in injury detection but often underperform on darker skin tones. Most fairness evaluations in dermatology AI rely on a single method of skin tone categorization. In biased models, the choice of skin tone scale plays a critical role, not only in grouping individuals, but also in making disparities visible. A well-designed scale should reveal, rather than mask, differences in model performance across skin tones. This study evaluated how different skin tone classification methods affect the accuracy, average precision, and fairness measures in bruise detection. Methods Using a dataset of 11,766 bruise and non-bruise images collected under white and alternate light sources, six skin tone labeling methods were applied: Individual Typology Angle (ITA), Fitzpatrick scale (full and binarized), Monk Skin Tone (MST), and two clustering approaches (chromatic and luminance-based). Chromatic clusters were created using the raw L* (lightness), a* (red-green), and b* (yellow-blue) values, whereas luminance clustering used the L* values only. A YOLOv5 model was re-trained for bruise detection, and model accuracy, average precision, and fairness were assessed on the test set using Demographic Parity Difference (DPD) and Equal Opportunity Difference (EOD) metrics across these six skin tone groupings. Results Performance and fairness varied widely depending on the skin tone classification method used. Granular scales such as the Monk Skin Tone (MST) scale and the chromatic cluster-based scale revealed greater variability in both performance and fairness across the skin tone groups, especially in the mid-to-dark range. In contrast, simpler and broader scales such as luminance-based clustering and binary Fitzpatrick showed more stable trends, but they may have hidden important differences between skin tones. Conclusions Skin tone classification plays a key role in how both performance and fairness are evaluated in bruise detection models. Granular skin tone scales such as MST and chromatic clustering may not show the highest performance, but reveal disparities more clearly, whereas broader scales may mask them despite performing well. Addressing these biases requires careful selection of skin tone grouping methods for evaluation. Skin tone bias Bruise detection AI fairness Demographic parity Equal Opportunity Object detection Figures Figure 1 Figure 2 Figure 3 Figure 4 Figure 5 Figure 6 Figure 7 Figure 8 Figure 9 Figure 10 Figure 11 Figure 12 Introduction Despite advances in artificial intelligence (AI) based medical imaging, patients with darker skin tones remain at a higher risk of being overlooked [ 1 – 3 ]. An injury caused by the same mechanism, at the same time, and at the same location can look completely different on different skin tones. AI tools designed to detect physical injuries often rely on visual features that may or may not appear the same across all skin tones (Fig. 1 ). As a result, these tools may make inaccurate detections. One of the reasons for such bias is the limited representation of skin tones in training datasets [ 4 ]. This kind of representation bias is well-known in dermatology-focused AI, where models trained mostly on lighter or darker skin often struggle to perform well on the other [ 5 , 6 ]. Similar disparities have also been reported in non-AI technologies, such as pulse oximetry, where oxygen saturation levels are systematically overestimated in patients with darker skin due to differences in light absorption [ 7 , 8 ]. When these tools are not evaluated across diverse populations, they can increase disparities in care and reduce confidence in the use of AI in clinical settings. This challenge is especially evident in the context of bruise detection, because bruises are often subtle and vary in appearance depending on skin pigmentation. On lighter skin, bruises may appear red, purple, or yellow as they heal. On darker skin, bruises often appear deep purple or brown and may blend with the surrounding skin, making them harder to detect by both humans and AI tools [ 9 ]. In our previous s work, we found that bruises are harder to see on darker skin and to address this, we incorporated images captured under alternate light source (ALS) conditions to enhance bruise visibility during data collection [ 10 ]. ALS is a method that uses controlled wavelengths of light (ultraviolet, visible, or infrared ranges) to enhance visibility of bruises that are difficult to see under natural lighting [ 11 ]. However, improving visibility is only one part of the solution. AI systems depend on labeled examples to learn patterns, but what is labeled depends on what is visible to human annotators. If bruises are less visible on certain skin tones, they may be under-annotated or missed entirely during dataset creation. In turn, the model may not learn to detect these cases, which further increases disparities in performance. In addition to labeling bruises, fairness evaluations also depend on how skin tone is recorded. Not surprisingly, skin tone is not always measured in a consistent or objective way. In many datasets it is estimated either by using L*a*b* values ( L represents lightness, while a and b represent the red-green and blue-yellow color components respectively) from a colorimeter, or by comparing it to a predefined skin tone scale [ 12 ]. These measurements can change based on lighting conditions or camera quality and often depend on the individual labeling them. Without a standard method, skin tone annotations may be inaccurate and introduce bias into the dataset. This is important when trying to measure fairness because model performance is often compared across skin tone groups. If those groups are defined in a vague or inconsistent way, they can hide true differences in performance or make small gaps look larger than they are. Sometimes what looks like bias in the model may be the result of how skin tone was labeled. To evaluate fairness properly, careful consideration must be given to how skin tones are defined and grouped. There is no single correct way to define those groups. Each method groups skin tones differently and that can lead to different conclusions about fairness. Given that the underlying model can exhibit bias, this raises an important question in bruise detection: Which skin tone classification is the most effective at revealing those disparities? This study explores that question by comparing not only model accuracy, but also fairness evaluation results across two distinct approaches to skin tone classification: using existing skin tone scales and data-driven groups constructed by clustering. The central idea is that a good classification scale should make disparities visible rather than hide them. By examining the same model through multiple classification frameworks, the study aims to better understand how skin tone definitions influence bias in AI based bruise detection. Individual Typology Angle (ITA) Skin Tone Classification The Individual Typology Angle (ITA) is a continuous measure used to classify skin tone based on values from the CIELAB (Commission Internationale de l’Eclairage L*a*b*) color space, which is widely used in imaging and color science [ 13 ]. CIELAB separates color into three components: L* for lightness, a* for the red-green axis, and b* for the yellow-blue axis (Fig. 2 , left). ITA relies specifically on the L* and b* values to calculate an angle that reflects how light or dark the skin appears (Fig. 2 , right). The higher the angle, the lighter the skin and vice versa. ITA offers a numeric and objective way to classify skin tone without relying on human perception. In the study, L* and b* values were extracted from a selected area of exposed non-bruised skin in each participant and were used to compute an ITA score using the standard formula, Eq. (1). Once calculated, each image was placed into one of six tone categories based on thresholds widely used in prior studies: very light (> 55°), light (41°- 55°), intermediate (28°- 41°), tan (10°- 28°), brown (-30°- 10°), and dark ( < − 30°) [ 14 ]. ITA = \(\:\text{arctan}[\frac{({L}^{*}-\:50)}{{b}^{*}}]*\frac{180}{\pi\:}\) . (1) Fitzpatrick Skin Tone Classification The Fitzpatrick scale was first introduced in 1975 primarily to describe how the skin responds to sunlight, and to help dermatologists make decisions in phototherapy and other dermatological treatments by estimating how likely a person’s skin is to burn or tan when exposed to the sun [ 15 ]. It includes six categories, starting with very light skin that burns easily (Type I) and ending with highly pigmented skin that rarely burns (Type VI) (Fig. 3 ). Even though it was meant to measure UV sensitivity, not to represent how the skin looks, multiple researchers have often used it in medical AI studies to group people by skin tone. This is due to its wide availability and its established use in dermatology, even though it does not capture the full range of visual skin color variation. One major criticism of the Fitzpatrick scale is that it was developed based on a limited and non-diverse population [ 16 ]. More specifically, it primarily reflects the skin characteristics of lighter-skinned individuals. As a result, the darker categories (Types V and VI) were added after the original classification and were not developed with the same level of empirical validation [ 17 ]. This has led to concerns that the scale compresses a wide range of darker skin tones into just one or two categories, which may overlook important visual and clinical differences between individuals and introduce unintended bias. For example, in a survey of more than 2,000 Black adults, 59 percent said they could not find a skin tone on the Fitzpatrick scale that matched their own [ 18 ]. Furthermore, the scale is often used in a subjective way, based on how someone sees themselves or how others see them, rather than through any consistent or objective method. These inconsistencies in labeling can result in misclassification, especially in bruise datasets where lighting, image quality, and camera settings can influence how the bruise appears. In fairness evaluations, relying on the Fitzpatrick scale may therefore lead to inaccurate assessments, as it may mask disparities that exist within the broad groupings of darker skin. For example, a model may perform well on medium-dark skin but poorly on very dark skin, yet both might fall under the same Fitzpatrick type. This lack of granularity may limit the scale’s usefulness in settings where precise evaluation of model performance across skin tones is needed. Monk Skin Tone Classification The Monk Skin Tone (MST) scale was recently developed to offer a more inclusive way to represent skin tone, especially in fields like computer vision and AI [ 19 ]. It was created by Dr. Ellis Monk, a sociologist at Harvard, in collaboration with Google Research. The scale includes ten distinct skin tones, either as spheres/orbs with gradients or flat patches, that range from very light to very dark, based on how people actually see and interpret skin color in images (Fig. 4 ). Unlike older frameworks such as the Fitzpatrick scale, the MST scale was built around visual perception. Its goal is to better reflect the diversity of real-world populations, especially those who have often been underrepresented or oversimplified in datasets. What makes the MST scale different is that it does not treat skin tone as a biological or clinical variable, but instead focuses on how skin looks in everyday settings, whether in photos, on screens, or in videos [ 20 ]. It was developed through both sociological research and public perception testing to make sure the ten skin tones are clear, distinct, and reflective of how people describe skin color in everyday life. This makes it especially useful in fairness research, where subtle differences in appearance can affect how an AI model performs. In this study, MST values were assigned to each subject based on pixel-level color matching between the subject’s skin tone and the reference swatches from the MST spheres. These assignments were then used to group participants for fairness evaluations, allowing to look more closely at whether the model performed differently across the full range of skin tones. Methods To evaluate algorithmic fairness in bruise detection across diverse skin tones, six skin tone classification frameworks were employed: ITA categories, binarized and full-category Fitzpatrick Skin Types (FST), the Monk Skin Tone (MST) scale, data-driven clustering using L*a*b* color values (chromatic), and clustering using only the L* (luminance) component. Binarized Fitzpatrick (grouped types I-III as light skin, IV-VI as dark) was chosen as an additional approach since it has been commonly used in literature for object detection tasks and also with fairness metrics [ 22 , 23 ]. By applying both chromatic and luminance-based clustering, the aim was to compare the impact of classification framework choice on model accuracy and measured fairness outcomes. These frameworks were used to stratify this study’s dataset of annotated bruise images, allowing the assessment of group-level performance disparities using standard fairness metrics. Data Sources This study utilizes a bruise dataset collected previously with support from the National Institute of Justice (NIJ; Award# 2016-DN-BX-0147). A total of 11,766 bruise and non-bruise images from 118 distinct participants were used, capturing 246 distinct bruises (Table 1 ). The images were captured in a controlled laboratory setting, where bruises were deliberately induced using either a paintball impact or a dropped weight to simulate real-world blunt force trauma [ 10 ]. Each injury was photographed at several time points to document how the bruise changed over time. The imaging protocol included both white lighting and ALS conditions. Each image was annotated by three independent and trained human labelers, and any potential discrepancies were resolved by a team leader. The final annotations included bounding boxes around the bruises, which were used as the ground truth for model training. All annotations were stored in a structured database for further analysis. For model development and evaluation, the dataset was split into training and testing sets in a 85:15 ratio. In addition to images, structured data was also collected for each subject, including age, gender, injury location, and skin tone. Skin tone categories in the original dataset was based on colorimetric measurements from uninjured adjacent skin, reported in L*a*b* color space [ 10 ]. Subjects were then assigned to one of six skin tone categories using their calculated ITA values: very light (> 55°), light (41–55°), intermediate (28–41°), tan (10–28°), brown (− 30–10°), and dark ( ≤ − 30°). This approach enabled a consistent, measurement-based classification across the full range of pigmentation. All personally identifiable information was removed prior to analysis in accordance with HIPAA regulations. The resulting database included both annotated images and their associated metadata. A custom platform with a user interface and API is currently being developed by the team to enable multi-criteria searches and AI training of bruise images, aiming to become the largest ALS-based bruise image repository [ 24 ]. Table 1 Test set distribution by skin color, gender, and age group. Category Subjects n (%) Injuries n (%) Images n (%) Test Train Test Train Test Train Skin Color Type I – Very Light 15 (12.9) 15 (12.7) 16 (13.1) 16 (12.9) 289 (16.4) 1477 (14.8) Type II - Light 23 (19.8) 23 (19.5) 25 (20.5) 27 (21.8) 373 (21.1) 2241 (22.4) Type III - Intermediate 21 (18.1) 21 (17.8) 23 (18.9) 21 (16.9) 335 (19.0) 1864 (18.6) Type IV - Tan 23 (19.8) 24 (20.3) 23 (18.9) 24 (19.4) 300 (16.9) 1876 (18.8) Type V - Brown 17 (14.7) 17 (14.4) 17 (13.9) 17 (13.7) 259 (14.7) 1369 (13.7) Type VI - Dark 17 (14.7) 18 (15.3) 18 (14.7) 19 (15.3) 210 (11.9) 1173 (11.7) Gender Male 34 (29.3) 35 (29.7) 34 (27.9) 35 (28.2) 482 (27.3) 2780 (27.8) Female 82 (70.7) 83 (70.3) 88 (72.1) 89 (71.8) 1284 (72.7) 7220 (72.2) Age group (years) 0–18 25 (21.6) 25 (21.2) 26 (21.3) 25 (20.2) 466 (26.4) 2460 (24.6) 19–35 80 (68.9) 82 (69.5) 84 (68.9) 86 (69.3) 1164 (65.9) 6723 (67.2) 36–50 9 (7.8) 9 (7.6) 10 (8.2) 11 (8.9) 128 (7.2) 748 (7.5) 51–65 2 (1.7) 2 (1.7) 2 (1.6) 2 (1.6) 8 (0.5) 69 (0.7) Total 116 (100) 118 (100) 122 (100) 124 (100) 1766 (100) 10,000 (100) Bruise Detection Model The bruise detection model was trained using the YOLOv5 (You Only Look Once Version 5) architecture, which is a real-time object detection model that evaluates the entire image in one pass [ 25 ]. By partitioning the input into a grid and predicting object locations and class labels simultaneously, YOLO avoids the sequential steps used in earlier pipelines, offering improved speed and adaptability to visual variability in bruising. Given the dataset size, training the entire model was not feasible. Instead, transfer learning was applied. This meant that the backbone of a pre-trained YOLOv5 model was kept fixed while the detection-specific layers were fine-tuned on the dataset [ 26 ]. This approach allowed the model to build on general visual features while learning specific cues relevant to bruise identification. Each image was processed to extract the ground-truth bounding boxes by converting YOLO's normalized coordinate format into absolute pixel dimensions. Following this, images were normalized and transformed into tensor representations compatible with the model input requirements. The trained YOLO network then produced predicted bruise masks. Bounding boxes were then produced during model training, along with associated confidence and the intersection over union (IoU) scores for each detection. The confidence score indicates how strongly the model believes a detected region contains a bruise while IoU measures the overlap between the predicted box and the ground truth [ 27 , 28 ]. The YOLOv5 model was selected based on a detailed investigation of several object detection architectures, which is out of scope for this publication. Skin Tone Categorization and Mapping To enable comparisons across other labeling systems, Fitzpatrick and MST classifications were added using a nearest-neighbor approach based on the L*a*b* color space (Fig. 5 ). For each scale, representative skin tone swatches were collected, i.e. Fitzpatrick types I–VI and MST orbs 1–10, and extracted RGB pixel values from each swatch image. These RGB values were normalized and converted to CIELAB format using a standard perceptual transformation implemented in the color space conversion package in Python v 3.11.9 [ 29 , 30 ]. Each subject’s L*a*b* values were then pulled from existing metadata and compared to the full set of reference tones using Euclidean distance. In cases where multiple swatches were similarly close in the L*a*b* space, ties were resolved by assigning the label with the lowest average distance among its nearest neighbors (k = 10). Fitzpatrick values were then mapped to its respective categories, while MST values were kept in their original numeric form. ITA categories were already present in the dataset and were used without modification. Skin Tone Clustering Rather than relying on predefined skin tone classifications, which may not reflect the full variation in a bruise dataset, groupings were derived directly from the skin color data. This allowed us to examine whether the model behaved differently across empirically defined clusters, based on how the skin actually appeared, rather than how it was categorized. For this, two clustering methods were used: Chromatic Clustering : This approach grouped participants based on their full color profile using all three components of the L*a*b* color space. This captured both brightness and undertone. The goal here was to identify whether combinations of hue and luminance of the participant’s skin tone played a role in bruise visibility and whether the model's detections varied across those combinations. Since bruise appearance is not just about how dark it is, but also how its color blends with or stands out from the surrounding skin, clustering by full color information provided a way to account for that complexity. Luminance Clustering : This second approach included only the L* channel from the CIELAB color wheel, which represents brightness. This was used to isolate whether variation in lightness alone was sufficient to explain changes in model performance. This approach was driven by the observation that L* showed the most variability in our dataset and largely defined the clustering structure, whereas a* and b* contributed little separation. By removing chromatic components and focusing solely on luminance, detection disparities could be tested to determine whether they were driven by brightness differences alone. Both clustering approaches were first performed on the full training set, using only participants with complete L*a*b* data in order to generate stable and representative cluster definitions. For each method, agglomerative hierarchical clustering was used and the optimal number of clusters was selected based on silhouette score analysis [ 31 ]. The average silhouette scores were then computed for each of clustering methods across a range of cluster counts. The score is higher when points are well matched to their own cluster and far from others. The number of clusters (k) was systematically varied from two to ten, and the optimal value was identified by selecting the k that yielded the highest silhouette score. Once clusters were established, the same clustering assignments were applied to the held-out test set by mapping test set points to the nearest training-derived cluster centroids in the corresponding feature space. This ensured that the test set was grouped consistently with the clusters identified with the training data. Model Evaluation and Fairness Metrics Once the clusters were established, accuracy and model fairness was evaluated across skin tone groupings using two widely used metrics, Demographic Parity Difference (DPD) and Equal Opportunity Difference (EOD) (Table 2 ). These metrics were chosen because they specifically assess differences in model predictions across groups and were suitable for the distribution of outcomes in our dataset [ 32 , 33 ]. In the fairest case for these metrics, the observed difference is zero, indicating no measurable disparity across skin tone groups. Other fairness metrics such as Equalized Odds, and performance metrics such as precision, recall, and AUC, were not used due to the limited number of non-bruise images in the test set, which prevented reliable comparisons across skin tone groups. Fairness evaluations were conducted across six skin tone classification methods: ITA, Fitzpatrick, Fitzpatrick binarized, MST, and the two clustering-based groupings (chromatic and luminance). Each image was assigned to a skin tone group based on its labeling method, and those groupings were held constant during evaluation. For the fairness and accuracy metrics, predictions were binarized using fixed thresholds. A prediction ( \(\:\widehat{Y})\:\) was considered valid if the confidence score was at least 0.4 and the IoU was 0.6. These thresholds were selected based on prior analysis, which showed that values in the range of 0.2–0.6 for confidence and 0.5–0.7 for IoU offered the best balance between performance and fairness. Standardizing thresholds allowed us to compare fairness outcomes consistently across all grouping methods. In contrast, average precision (AP) was also calculated without applying a confidence threshold. AP is a standard performance metric used commonly in object detection tasks. It measures the area under the precision-recall curve for a single class, summarizing how well the model balances precision and recall across all confidence levels [ 34 ]. Mean average precision (mAP) further aggregates AP across all detection tasks to give a summary measure of performance. A skin tone scale was considered good if it showed a fluctuation between skin tones for the considered metrics. The goal was not to optimize the model, but to find a skin tone scale that “breaks” the model by showing potential bias. Thus, the larger the fluctuation in accuracy and fairness metrics, the more appropriate the skin tone scale is for evaluating models. Table 2 Fairness metric formulas and descriptions used to assess group-level disparities in bruise detection. Metric Description Formula Demographic Parity Difference (DPD) Difference in the rate of positive predictions across groups, regardless of actual outcome. \(\:\text{D}\text{P}\:\stackrel{\scriptscriptstyle\text{def}}{=}{P\left(\widehat{Y}=1∣A=a\right)\:-P\left(\widehat{\:Y}=1\right)}^{*}\) Equal Opportunity Difference (EOD) Difference in true positive rates across groups, conditional on the outcome being positive. \(\:\text{E}\text{O}\:\stackrel{\scriptscriptstyle\text{def}}{=}{P\left(\widehat{Y}=1∣A=a\right)\:-P\left(\widehat{\:Y}=1\right)}^{*}\) * \(\:\widehat{Y}\) denotes the model’s predicted label (1 = bruise detected); \(\:Y\) denotes the ground truth label (1 = actual bruise present). A = a indicates that the individual belongs to skin tone group a . P represents conditional probability. Results After mapping the data to its respective Fitzpatrick and MST scales, the distribution of skin tone labels differed noticeably across systems (Fig. 6 ). The raw ITA labels were spread more evenly, with many images falling into Light, Intermediate, and Tan categories. MST labels were clustered at the lighter end, mostly in Types 2 and 3, with few images in darker categories. No images in the dataset were labeled as MST Types 9 or 10. Only a small number of images were labeled in the darker MST categories. The Fitzpatrick distribution showed a similar skew, with most images assigned to Type I, fewer in Types IV and V, and none in Type VI. Clustering Chromatic clustering based on L*, a*, and b* values produced the highest silhouette score, 0.62, at k = 2 (Fig. 7 ). Scores decreased steadily as the number of clusters increased, with silhouette values dropping to 0.38 by k = 10. By k = 7, the silhouette score had dropped to 0.41 and did not show further improvement beyond that point. As a result, seven clusters were selected as the final grouping for analysis. In contrast, clustering based on L* alone resulted in consistently higher silhouette scores across the same range of cluster counts. From the silhouette score graph, k = 4 is where the plot decreases or flattens with minimal drop-off through adjacent values. The corresponding dendrogram showed clear hierarchical separation, with branch splits occurring at higher distances and greater vertical spacing between cluster boundaries. In the dendrogram constructed from the hierarchical linkage matrix, the default visualization in the Python package separated the data into three main clusters, corresponding to the largest linkage distances. However, finer-grained splits at lower distances were evident throughout the structure, and the clusters were chosen based on the silhouette score metric (Fig. 8 ). To visualize the skin tone distributions captured by both clustering methods, the average L*a*b* values were calculated for each group. These values were then converted into representative color swatches, forming a continuous scale from lighter to darker tones. The resulting palettes illustrates the range of skin tones present in the dataset as identified by the clustering process (Fig. 9 ). In Fig. 10 , the distributions of all six classification methods in L*a*b* space is shown in panels A to F. Across all labeling systems, the data points form a narrow, curved band that runs primarily along the L* axis. Depending on the classification method, there is little to no variance between the chromatic dimensions (a* and b*). In comparing the classification methods, it was also found that the same image was often assigned to different skin tone categories depending on which system was used. Each system organizes skin tones in a way that mirrors variation in brightness more than color. For example, an image labeled as “intermediate” in the ITA scale could be placed in a different group under MST or fall near a boundary in chromatic clustering. These inconsistencies were common and raised concerns about the alignment between classification frameworks. This further supported the need for data-driven grouping strategies and justified the inclusion of L*-based clustering alongside the chromatic approach. Since the dataset with the raw ITA values already showed a strong tendency to separate by L* (Panel A), the study also tested whether brightness alone could form meaningful and consistent groupings. Performance Metrics For most scales, accuracy was higher in lighter skin tone groups and lower in darker ones (Table 3 ). In the Monk Skin Tone scale, accuracy dropped from 0.90 in group 1 to 0.78 in group 8, with some fluctuation in later groups. The full Fitzpatrick scale showed a similar pattern, with accuracy falling from 0.86 in type I to 0.67 in type V. When the Fitzpatrick scale was binarized into light and dark groups, the light group still performed better (0.84 vs. 0.72). The ITA scale also showed a decline, from 0.91 in very light tones to 0.73 in dark tones. Chromatic clusters had more variation, with lower accuracy in some mid and darker clusters like C2, C5, and C6. Luminance clusters showed the opposite trend with the lightest cluster having the lowest accuracy. Overall, the model tended to be more accurate on lighter tones, and performance dropped for darker skin across most scales. Table 3 Skin tone group accuracies for each classification method at IoU ≥ 0.6 and confidence ≥ 0.4. Scale Group Accuracy Individual Typology Angle Very Light 0.91 Light 0.84 Intermediate 0.85 Tan 0.84 Brown 0.76 Dark 0.73 Fitzpatrick I 0.86 II 0.80 III 0.83 IV 0.74 V 0.67 Fitzpatrick Binarized Light 0.84 Dark 0.72 Monk Skin Tone 1 0.90 2 0.87 3 0.82 4 0.75 5 0.67 6 7 8 0.93 0.71 0.78 Chromatic Clusters C0 0.85 C1 0.83 C2 0.74 C3 0.86 C4 0.89 C5 0.71 C6 0.74 Luminance Clusters L0 0.72 L1 0.83 L2 0.87 L3 0.81 These results can be further validated by the AP results below (Fig. 11 ). Patterns in average precision generally mirrored the accuracy trends. Most classification methods showed higher AP values for lighter skin tone groups and a gradual decline for darker ones. The Fitzpatrick and ITA categories followed a similar pattern, with the darkest tones showing the lowest AP. Although binarized Fitzpatrick produced more stable results, the light group still outperformed the dark group. MST and Chromatic clusters showed greater fluctuation, particularly in mid-range tones. In contrast, luminance clusters showed less variation overall but showed an increase towards the darkest cluster. Fairness Metrics Fairness evaluation was assessed using DPD and EOD gaps across all six skin tone classification methods: ITA, Fitzpatrick, MST, binarized Fitzpatrick, chromatic clusters, and luminance-based clusters. In this context, a "gap" refers to the difference between a specific group's prediction rate and the overall average prediction rate across all skin tone groups. While each system defines skin tone categories differently, they all follow the same basic structure by organizing individuals from lighter to darker tones. For comparability, they were aligned along a shared axis. All groups were mapped to a normalized skin tone scale, ranging from 0 (lightest) to 1 (darkest). Across both DPD and EOD, a noticeable decline in gaps appeared towards the end/darker tones of the normalized skin tone scale for several classification methods (Fig. 11 ). The graphs show that fairness outcomes are not always linear across the skin tone spectrum and that mid-tone groups may not experience the same levels of disparity as lighter or darker groups. Gap values decreased from lighter to darker groups in both metrics, with the most negative values often occurring at the darkest end of the scale, particularly for Fitzpatrick. MST and chromatic clusters showed the largest drops in the mid-to-dark range. In contrast, both luminance clustering and ITA showed relatively smoother trends across, with gap values decreasing more gradually and remaining relatively close to zero across most tone groups. Luminance clustering showed a decline in fairness gaps for darker skin tones despite showing the highest accuracy and AP for that group. Binary Fitzpatrick, due to its two-group structure, produced stable values across the tone scale, with both DPD and EOD gaps remaining moderate. Discussion This study evaluated the impact of skin tone classification on model performance and fairness by holding the model and data constant while varying the labeling scale. Most of the classification methods, specifically ITA, MST, Fitzpatrick full and binarized, and chromatic clustering, showed higher accuracy and AP in lighter skin tone groups and lower in darker ones. In scales with more granularity, like the MST, the decrease in accuracy and AP was not smooth. The only exception was luminance clustering, where they were lower in lighter groups and improved in the darker ones. Despite this, fairness dropped for darker skin tones. This difference may be because luminance clustering groups skin tones based only on brightness, ignoring color differences that affect how bruises appear. These shifts show that even with the same model and data, the way skin tone is defined can change model performance and that matters when evaluating its fairness. Relying only on performance metrics such as accuracy may not always mean that fairness would be high. Testing and selecting the correct scale considering both aspects, accuracy and fairness, becomes essential in bruise model and bias detection. Model accuracy, AP, and both DPD and EOD fairness metrics revealed that detection disparities tend to worsen as skin tone darkens under all scales, and especially under classification methods with coarse or poorly aligned groupings like ITA and Fitzpatrick full and binarized. The closer the fairness metric values are to zero, the fairer the model is. When lower DPD values appear for darker skin tones, it indicates that the model is less likely to make positive predictions in those groups compared to the overall average. EOD, which measures correct bruise detections among cases where a bruise is actually present, showed similar gaps. Lower EOD values for darker tones suggest that even when bruises exist, the model is more likely to miss them. This is particularly important in clinical contexts, where missed detections can result in bruises going undocumented, potentially affecting follow-up care and decision-making. The classifications with the largest fluctuations, MST and chromatic clustering, categorize skin into more specific skin tone groups. These finer groupings do not cause disparities, but they do expose them more clearly by identifying groups for which models underperform. In contrast, binarized Fitzpatrick produced more stable DPD and EOD gaps. But this stability may be misleading as broader groupings may average out performance differences within each group, masking disparities that exist between medium and very dark tones. Likewise, although luminance clustering produced relatively smooth trends, its reliance on lightness (L*) alone, without incorporating chromatic dimensions (a* and b*), may oversimplify how skin tone affects bruise visibility, which may potentially limit its fairness in more diverse populations. In most classification methods, both DPD and EOD appeared more stable for mid-range skin tones. This suggests that the model may be most stable for medium tones under all scales, while its predictions become less consistent at the lighter and darker ends. Instead of a simple increase in bias with darker skin, the model appears to struggle more at both extremes, regardless of the scale. With positive values towards lighter skin tones, lighter skin receives more favorable treatment than average, whereas darker skin receives less favorable. The results also show a clear tradeoff between how detailed a skin tone scale is and how practical it is to use. Scales with more categories, like MST or chromatic clusters, helped show exactly where the model’s accuracy dropped. But they also made the results more variable and harder to interpret. Simpler groupings, like binary Fitzpatrick or luminance clusters, gave more stable results, but likely missed differences by grouping skin tones together. This means that using broader categories can make a model look more consistent and fairer than it really is, especially if it's underperforming for specific groups. To improve fairness, future efforts should focus on both better skin tone representation in datasets and refining how skin tone groupings are defined and used in evaluation. This includes more detailed annotations for images where bruises are harder to see, especially on darker skin. Limitations This study has several limitations that should be considered when interpreting the results. First, the dataset remains under development, and while it currently includes over 11,000 bruise and non-bruise images, fewer images fell in the very dark spectrum of the skin tone scales. As a result, no images were mapped to MST 9 or 10 and Fitzpatrick Type VI skin tones. Additional data collection has been ongoing to address this imbalance. Second, mapping the dataset across the different labeling systems produced inconsistent results. For example, the same image would receive two very different skin tone labels depending on the method used. This problem is also reported in [ 35 ]. Hence, each pixel was matched to determine the closest match with our data. This variation may make it difficult to establish a single ground truth, which may also impact fairness differently. Additionally, the dataset was collected in a controlled setting with deliberate inducing of bruises. This may not capture the complex environments found in clinical settings. Lighting, camera differences, and natural variation in injury appearance could all influence model performance outside the lab. Testing in real-world conditions is needed to confirm whether these fairness patterns hold. Finally, and most importantly, the true box annotations for the bruises depend on if its visible to humans. Since the data was collected in the lab, the collection team was aware if there is a bruise or not. But when the tool is applied to real-world settings, bruises on dark skin might be missed due to visibility and not captured at all. Future work should focus on improving data diversity, labeling and annotation consistency, and real-world testing to ensure more equitable bruise detection. Conclusion This study highlights the critical role that skin tone classification plays in evaluating model performance and fairness in AI-based bruise detection systems. The findings make it clear that the choice of labeling framework significantly shapes how disparities are revealed and addressed. When more coarse classification methods, such as the MST and chromatic clusters scale, were used, performance gaps became more visible across the full scale. In contrast, broader grouping strategies, including binary Fitzpatrick and luminance-based clustering, appeared to minimize these gaps, not necessarily by improving model performance, but perhaps by averaging over variation that may reflect meaningful clinical differences. The focus of this paper was not to reduce bias, but to select the best approach to detect it. Most fairness research in medical AI focuses on mitigation, resampling, reweighting, or using ensemble models to equalize outcomes across groups [ 36 ]. These methods assume that group labels are already valid and meaningful. But if the categories themselves are poorly defined or too coarse, they can hide real differences in model behavior. Our results show that depending on which skin tone scale is used, model performance can appear stable or vary significantly across groups. This suggests that fairness interventions that don’t consider how skin tone is defined may miss important disparities altogether. By systematically comparing six skin tone classification methods, this study provides evidence that fairness evaluations must be grounded in thoughtful and transparent choices about how subgroups are defined. Abbreviations ALS Alternate Light Source ITA Individual Typology Angle CIELAB Commission Internationale de l’Eclairage L*a*b* MST Monk Skin Tone DPD Demographic Parity Difference EOD Equal Opportunity Difference IoU Intersection over Union AP Average Precision mAP Mean Average Precision FST Fitzpatrick Skin Type RGB Red, Green, Blue HIPAA Health Insurance Portability and Accountability Act YOLO You Only Look Once. Declarations Ethics approval and consent to participate The George Mason University IRB (# STUDY00000129 ) determined that the proposed activity is not research involving human subjects as defined by the U.S. DHHS and FDA regulations. The informed consent is waived by the IRB (# STUDY00000129) . This study complies with the principles of the Declaration of Helsinki. Consent for publication Not applicable Availability of data and material The datasets during and/or analyzed during the current study is available from the corresponding author on reasonable request. Competing interests The authors declare that they have no competing interests. Funding The project was supported by the Artificial Intelligence/Machine Learning Consortium to Advance Health Equity and Researcher Diversity (AIM-AHEAD), funded by the National Institutes of Health (Grant #1OT2OD032581-02-811). Author’s contributions D.D. is the primary author of the manuscript. J.W. contributed to the writing and supervised the development of the methods and experimental design. K.S. contributed to the writing and interpretation of data. M.G. contributed to the writing and developed the bruise detection model. D.L. supported the implementation of the deep learning components. A.N. contributed to the design and analysis of fairness metrics. Acknowledgments Not applicable. Authors’ information 1 Health Informatics Program, College of Public Health, George Mason University, Fairfax, VA, USA. 2 Department of Civil, Environmental, and Infrastructure Engineering, George Mason University, Fairfax, VA, USA. 3 Department of Nursing, College of Public Health, George Mason University, Fairfax, VA, USA. References O’Malley A, Veenhuizen M, Ahmed A. Ensuring appropriate representation in AI-generated medical imagery: A Methodological Approach to Address Skin Tone Bias (Preprint). JMIR AI. 2024 Mar 11; Bevan P, Atapour-Abarghouei A. Detecting Melanoma Fairly: Skin Tone Detection and Debiasing for Skin Lesion Classification [Internet]. Available from: https://arxiv.org/pdf/2202.02832. Daneshjou R, Vodrahalli K, Novoa RA, Jenkins M, Liang W, Rotemberg V, et al. Disparities in dermatology AI performance on a diverse, curated clinical image set. Science Advances. 2022 Aug 12;8(32). Norori N, Hu Q, Aellen FM, Faraci FD, Tzovara A. Addressing bias in big data and AI for health care: A call for open science. Patterns [Internet]. 2021 Oct 8;2(10):100347. Available from: https://pmc.ncbi.nlm.nih.gov/articles/PMC8515002/. Towards fairness in AI for melanoma detection: systemic review and recommendations [Internet]. [cited 2025 Mar 13]. Available from: https://arxiv.org/html/2411.12846v1. Groh M, Harris C, Soenksen L, Lau F, Han R, Kim A, et al. Evaluating deep neural networks trained on clinical images in dermatology with the Fitzpatrick 17k dataset [Internet]. arXiv; 2021 [cited 2025 Mar 13]. Available from: http://arxiv.org/abs/2104.09957. Al-Halawani R, Charlton PH, Qassem M, Kyriacou PA. A review of the effect of skin pigmentation on pulse oximeter accuracy. Physiol Meas. 2023;44(5):05TR01. Available from: https://pmc.ncbi.nlm.nih.gov/articles/PMC10391744/. Sjoding MW, Dickson RP, Iwashyna TJ, Gay SE, Valley TS. Racial bias in pulse oximetry measurement. N Engl J Med. 2020;383(25):2477–8. Available from: https://www.nejm.org/doi/full/10.1056/NEJMc2029240. King LM. An overview of bruises [Internet]. WebMD; 2024 May [cited 2025 Mar 13]. Available from: https://www.webmd.com/skin-problems-and-treatments/bruises-article. Scafide KN, Sheridan DJ, Downing NR, Hayat MJ. Detection of Inflicted Bruises by Alternate Light: Results of a Randomized Controlled Trial. Journal of Forensic Sciences. 2020 Feb 3;65(4):1191–8. Scafide KN, Downing NR, Kutahyalioglu NS, Sheridan DJ, Neil, Hayat MJ. Predicting alternate light absorption in areas of trauma based on degree of skin pigmentation: Not all wavelengths are equal. Forensic Science International. 2022 Jul 1;339:111410–0. Jiang L, Wang H, Gao C, Zhang X, Xiao K, Melgosa M, et al. Skin color measurements before and after two weeks of sun exposure. Vision Research. 2022 Mar; 192:107976. Ly BCK, Dyer EB, Feig JL, Chien AL, Del Bino S. Research Techniques Made Simple: Cutaneous Colorimetry: A Reliable Technique for Objective Skin Color Measurement. Journal of Investigative Dermatology [Internet]. 2020 Jan;140(1):3-12.e1. Available from: https://www.researchgate.net/publication/338303610_Research_Techniques_Made_Simple_Cutaneous_ Colorimetry_A_Reliable_Technique_for_Objective_Skin_Color_Measurement. Osto M, Hamzavi IH, Lim HW, Kohli I. Individual Typology Angle and Fitzpatrick Skin Phototypes are Not Equivalent in Photodermatology. Photochemistry and Photobiology. 2021 Nov 26; Ware OR, Dawson JE, Shinohara MM, Taylor SC. Racial limitations of fitzpatrick skin type. Cutis [Internet]. 2020 Feb 1;105(2):77–80. Available from: https://pubmed.ncbi.nlm.nih.gov/32186531/. Hecht M. What Are the Fitzpatrick Skin Types? [Internet]. Healthline. Healthline Media; 2019. Available from: https://www.healthline.com/health/beauty-skin-care/fitzpatrick-skin-types. Narla S, Heath CR, Alexis A, Silverberg JI. Racial disparities in dermatology. Archives of Dermatological Research. 2022 Dec 12;315(5). Pichon LC, Landrine H, Corral I, Hao Y, Mayer JA, Hoerster KD. Measuring skin cancer risk in African Americans: is the Fitzpatrick Skin Type Classification Scale culturally sensitive? Ethnicity & Disease [Internet]. 2010;20(2):174–9. Available from: https://pubmed.ncbi.nlm.nih.gov/20503899/. Monk E. The Monk Skin Tone Scale [Internet]. SocArXiv; 2023. Available from: osf.io/preprints/socarxiv/pdf4c_v1. Schumann C, Olanubi, Gbolahan O, Wright A, Jr M, Heldreth C, Ricco S. Consensus and Subjectivity of Skin Tone Annotation for ML Fairness [Internet]. arXiv.org. 2023. Available from: https://arxiv.org/abs/2305.09073. Skin Tone Research @ Google [Internet]. skintone.google. Available from: https://skintone.google/get-started. Pakzad A, Abhishek K, Hamarneh G. CIRCLe: Color Invariant Representation Learning for Unbiased Classification of Skin Lesions [Internet]. arXiv.org. 2022. Available from: https://arxiv.org/abs/2208.13528. Wilson B, Hoffman J, Morgenstern J. Predictive Inequity in Object Detection [Internet]. 2019 Feb. Available from: https://arxiv.org/pdf/1902.11097. Wojtusiak J, Qodrati M, Markiewicz M, Aminfar K, Lattanzi D, Scafide K. Interdisciplinary platform for bruise image research. Presented at: AMIA 2024 Annual Symposium; 2024 Nov 9–13; San Francisco, CA. Available from: https://knowledge.amia.org/A2024/pdf/a2024a332/a2024fl332. YOLO models for Object Detection Explained [Yolov8 Updated] [Internet]. encord.com. 2024. Available from: https://encord.com/blog/yolo-object-detection-guide/. Ultralytics. YOLO11 NEW [Internet]. [cited 2025 Mar 13]. Available from: https://docs.ultralytics.com/models/yolo11. Shroff M. Object detection with deep learning 2023: beginner’s friendly key terms explanation [Internet]. Medium; 2023 [cited 2025 Mar 13]. Available from: https://medium.com/@shroffmegha6695/object-detection-with-deep-learning-beginners-friendly-key-terms-explanation-d4fb594fea83. Pal M, Pokhriyal S, Sikdar S, Ganguly N. Ensuring generalized fairness in batch classification. Sci Rep. 2023;13(1):18892. Available from: https://www.nature.com/articles/s41598-023-45943-1. Python Release Python 3.11.9 [Internet]. Python.org. Available from: https://www.python.org/downloads/release/python-3119/. skimage.color — skimage 0.25.2 documentation [Internet]. Scikit-image.org. 2025 [cited 2025 Jun 11]. Available from: https://scikit-image.org/docs/0.25.x/api/skimage.color.html#skimage.color.rgb2lab. Dinh DT, Fujinami T, Huynh VN. Estimating the Optimal Number of Clusters in Categorical Data Clustering by Silhouette Coefficient. Communications in Computer and Information Science [Internet]. 2019 [cited 2021 Mar 10];1–17. Available from: https://link.springer.com/chapter/10.1007%2F978-981-15-1209-4_1. Lei H, Gohari A, Farnia F. On the Inductive Biases of Demographic Parity-based Fair Learning Algorithms [Internet]. arXiv; 2024 [cited 2025 Mar 13]. Available from: http://arxiv.org/abs/2402.18129. Hardt M, Price E, Price E, Srebro N. Equality of Opportunity in Supervised Learning. In: Advances in Neural Information Processing Systems [Internet]. Curran Associates, Inc.; 2016 [cited 2025 Mar 13]. Available from: https://proceedings.neurips.cc/paper/2016/hash/9d2682367c3935defcb1f9e247a97c0d-Abstract.html. Henderson P, Ferrari V. End-to-end training of object class detectors for mean average precision. arXiv:160703476 [cs] [Internet]. 2017 Mar 16 [cited 2022 May 16]; Available from: https://arxiv.org/abs/1607.03476. Cheng WC. Monastic Color Reproduction: A Software Tool for Printing and Assessing the Monk Skin Tone Scale. Journal of Imaging Science and Technology. 2024 Sep 1;68(5):1–13. An J, Ying L, Zhu Y. Why resampling outperforms reweighting for correcting sampling bias with stochastic gradients [Internet]. arXiv.org. 2020 [cited 2025 Jun 11]. Available from: https://arxiv.org/abs/2009.13447. Additional Declarations No competing interests reported. Cite Share Download PDF Status: Under Review Version 1 posted Editorial decision: Revision requested 10 Apr, 2026 Reviews received at journal 24 Nov, 2025 Reviews received at journal 23 Nov, 2025 Reviews received at journal 23 Nov, 2025 Reviewers agreed at journal 20 Nov, 2025 Reviewers agreed at journal 18 Nov, 2025 Reviewers agreed at journal 13 Nov, 2025 Editor invited by journal 17 Oct, 2025 Reviewers invited by journal 23 Jun, 2025 Editor assigned by journal 23 Jun, 2025 Submission checks completed at journal 20 Jun, 2025 First submitted to journal 20 Jun, 2025 You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {"props":{"pageProps":{"initialData":{"identity":"rs-6875703","acceptedTermsAndConditions":true,"allowDirectSubmit":false,"archivedVersions":[],"articleType":"Research Article","associatedPublications":[],"authors":[{"id":475237736,"identity":"1090ed48-9ec2-43b4-a3e7-34b1fc108071","order_by":0,"name":"Dharmi Desai","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAAAuklEQVRIiWNgGAWjYDCCAwlAwsBGDsJjI15LmjGpWhgOJzYQrYXvePLhDz8K0tLnh50xYPhQdpiwFskzz9IkewxscjfezjFgnHGOCC0GN3LMGHgM0nI3zs4xYOZtI0pL/uePfwwOpxuCtPwlTksOgzSPweEEeWmgFkZitAD9YiYtY5BmuEE6reBgz7l0wlqAIfb445s/NvLys5M3PvhRZk1YC8KFB4BxRIJ6IJBvIE39KBgFo2AUjCAAADhoPykHRsJSAAAAAElFTkSuQmCC","orcid":"","institution":"George Mason University","correspondingAuthor":true,"prefix":"","firstName":"Dharmi","middleName":"","lastName":"Desai","suffix":""},{"id":475237737,"identity":"9b41a6f6-17c5-47c9-88e5-87606a3ec45a","order_by":1,"name":"Amin Nayebi","email":"","orcid":"","institution":"George Mason University","correspondingAuthor":false,"prefix":"","firstName":"Amin","middleName":"","lastName":"Nayebi","suffix":""},{"id":475237738,"identity":"9e120681-ea67-4180-b878-3384dfbb09a3","order_by":2,"name":"Mehrdad Ghyabi","email":"","orcid":"","institution":"George Mason University","correspondingAuthor":false,"prefix":"","firstName":"Mehrdad","middleName":"","lastName":"Ghyabi","suffix":""},{"id":475237739,"identity":"efe8a43a-052c-4f25-bf18-e71cc962e446","order_by":3,"name":"David Lattanzi","email":"","orcid":"","institution":"George Mason University","correspondingAuthor":false,"prefix":"","firstName":"David","middleName":"","lastName":"Lattanzi","suffix":""},{"id":475237740,"identity":"9e129f3b-39d6-4c1a-9713-c8b9d6882ce2","order_by":4,"name":"Katherine Scafide","email":"","orcid":"","institution":"George Mason University","correspondingAuthor":false,"prefix":"","firstName":"Katherine","middleName":"","lastName":"Scafide","suffix":""},{"id":475237741,"identity":"32a63485-a974-4dc5-b6fc-15cccd73b1e0","order_by":5,"name":"Janusz Wojtusiak","email":"","orcid":"","institution":"George Mason University","correspondingAuthor":false,"prefix":"","firstName":"Janusz","middleName":"","lastName":"Wojtusiak","suffix":""}],"badges":[],"createdAt":"2025-06-12 02:23:05","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-6875703/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-6875703/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":85643150,"identity":"480226d7-5cf9-4f75-b1cd-873e3e2766d4","added_by":"auto","created_at":"2025-06-30 08:06:43","extension":"png","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":188990,"visible":true,"origin":"","legend":"\u003cp\u003eThe same bruising mechanism and timing used in natural light, but inconsistent detection across skin tones (light skin on left, dark on right). Green box – true bruise location, blue box – location the model predicts.\u003c/p\u003e","description":"","filename":"1.png","url":"https://assets-eu.researchsquare.com/files/rs-6875703/v1/644674e8bc41bb2ec93b8b3b.png"},{"id":85644565,"identity":"099e930b-565c-47c9-8b37-2da70cbc9cb9","added_by":"auto","created_at":"2025-06-30 08:14:43","extension":"png","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":59714,"visible":true,"origin":"","legend":"\u003cp\u003eITA-based classification of skin tone using CIELAB values. (A) CIELAB color space axes. (B) ITA angle thresholds mapped to six skin tone categories based on L* and b* values.\u003c/p\u003e","description":"","filename":"2.png","url":"https://assets-eu.researchsquare.com/files/rs-6875703/v1/e115d93ebf9a03b7dd49bc30.png"},{"id":85643151,"identity":"e9d27c87-63cf-42a8-bd75-46c500915ce4","added_by":"auto","created_at":"2025-06-30 08:06:43","extension":"png","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":44793,"visible":true,"origin":"","legend":"\u003cp\u003eThe Fitzpatrick Skin Tone Scale.\u003c/p\u003e","description":"","filename":"3.png","url":"https://assets-eu.researchsquare.com/files/rs-6875703/v1/3f3e774f38405da846572d1f.png"},{"id":85645736,"identity":"cbebae49-1168-4007-aed4-693572d4e7d1","added_by":"auto","created_at":"2025-06-30 08:30:43","extension":"png","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":94446,"visible":true,"origin":"","legend":"\u003cp\u003eThe Monk Skin Tone Orbs (based on data provided by Google Research [21]).\u003c/p\u003e","description":"","filename":"4.png","url":"https://assets-eu.researchsquare.com/files/rs-6875703/v1/ea620b05a36a549d97cff763.png"},{"id":85643157,"identity":"6b0a7a05-0f6e-4097-ae01-89a19f5718d8","added_by":"auto","created_at":"2025-06-30 08:06:43","extension":"png","order_by":5,"title":"Figure 5","display":"","copyAsset":false,"role":"figure","size":142703,"visible":true,"origin":"","legend":"\u003cp\u003eA representation of the skin tone mapping using L*a*b* color space. The input image’s (classified as ITA Brown on left) L*a*b* values (green) is matched to MST 8 (middle) and Fitzpatrick 3 (right) using nearest-neighbor comparisons. Top 10 neighbors are shown; best match is in blue.\u003c/p\u003e","description":"","filename":"5.png","url":"https://assets-eu.researchsquare.com/files/rs-6875703/v1/acf1308714015cae2c5e922e.png"},{"id":85643155,"identity":"ae790207-5a00-4de2-8844-9509ed3dc019","added_by":"auto","created_at":"2025-06-30 08:06:43","extension":"png","order_by":6,"title":"Figure 6","display":"","copyAsset":false,"role":"figure","size":22024,"visible":true,"origin":"","legend":"\u003cp\u003eTest set skin tone label distributions by ITA (left), MST (center), and Fitzpatrick (right) after pixel-level mapping.\u003c/p\u003e","description":"","filename":"6.png","url":"https://assets-eu.researchsquare.com/files/rs-6875703/v1/dfaacca074bf3822f6a1f1de.png"},{"id":85643152,"identity":"a165a32a-5c99-4562-b17b-b98482411335","added_by":"auto","created_at":"2025-06-30 08:06:43","extension":"png","order_by":7,"title":"Figure 7","display":"","copyAsset":false,"role":"figure","size":27572,"visible":true,"origin":"","legend":"\u003cp\u003eAverage silhouette scores for chromatic L*, a*, b* clustering and L*-only clustering.\u003c/p\u003e","description":"","filename":"7.png","url":"https://assets-eu.researchsquare.com/files/rs-6875703/v1/a0bfd3d52751b6bc27334ee0.png"},{"id":85644570,"identity":"9d802d89-28fc-4876-849d-7fc5485060ae","added_by":"auto","created_at":"2025-06-30 08:14:43","extension":"png","order_by":8,"title":"Figure 8","display":"","copyAsset":false,"role":"figure","size":41648,"visible":true,"origin":"","legend":"\u003cp\u003eDendrogram for chromatic clustering with \u003cem\u003ek = 7\u003c/em\u003e clusters (left) and for luminicance clustering with \u003cem\u003ek = 4\u003c/em\u003e clusters (right).\u003c/p\u003e","description":"","filename":"8.png","url":"https://assets-eu.researchsquare.com/files/rs-6875703/v1/b06d85969a8f420b4f989a21.png"},{"id":85644858,"identity":"6547cd4b-da71-4899-a851-693af0a1f45e","added_by":"auto","created_at":"2025-06-30 08:22:43","extension":"png","order_by":9,"title":"Figure 9","display":"","copyAsset":false,"role":"figure","size":9843,"visible":true,"origin":"","legend":"\u003cp\u003eVisual representation of the (top) seven chromatic clusters (C0–C6) and (bottom) four luminance clusters (L0–L3), based on average L*a*b* values within each group.\u003c/p\u003e","description":"","filename":"9.png","url":"https://assets-eu.researchsquare.com/files/rs-6875703/v1/346bcc536993749caaeaedb0.png"},{"id":85644576,"identity":"789a1fb5-33ec-4c30-9d34-32614bc9a6e3","added_by":"auto","created_at":"2025-06-30 08:14:44","extension":"png","order_by":10,"title":"Figure 10","display":"","copyAsset":false,"role":"figure","size":156780,"visible":true,"origin":"","legend":"\u003cp\u003e3D L*a*b* color space representations of skin tone classifications in the test dataset using (A) ITA categories, (B) Fitzpatrick, (C) Binarized Fitzpatrick, (D) Monk Skin Tone (E) Chromatic Clusters, (F) Luminicance Clusters. Each point represents the L*a*b* value assigned to a single image in the test set.\u003c/p\u003e","description":"","filename":"10.png","url":"https://assets-eu.researchsquare.com/files/rs-6875703/v1/5cd0bf4017816db357b386da.png"},{"id":85643169,"identity":"c689e45e-5139-4de5-a0b2-1330d6564d21","added_by":"auto","created_at":"2025-06-30 08:06:43","extension":"png","order_by":11,"title":"Figure 11","display":"","copyAsset":false,"role":"figure","size":116810,"visible":true,"origin":"","legend":"\u003cp\u003ePrecision-recall curves by skin tone group for each classification method (IoU ≥ 0.6), with average precision (AP) scores shown in the legend.\u003c/p\u003e","description":"","filename":"11.png","url":"https://assets-eu.researchsquare.com/files/rs-6875703/v1/500f1278977f59d53f7c3098.png"},{"id":85643166,"identity":"7b0ba98a-01f1-4b7b-aae3-71c553ad303f","added_by":"auto","created_at":"2025-06-30 08:06:43","extension":"png","order_by":12,"title":"Figure 12","display":"","copyAsset":false,"role":"figure","size":93800,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eFigure 11. \u003c/strong\u003eFairness metrics results across skin tone classifications. DPD on the left, EOD on the right.\u003c/p\u003e","description":"","filename":"12.png","url":"https://assets-eu.researchsquare.com/files/rs-6875703/v1/04ab582c42b33b21ff60dfb5.png"},{"id":85646374,"identity":"28fd9536-2403-4795-97d1-1f3639c0e1cb","added_by":"auto","created_at":"2025-06-30 08:38:45","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":1909834,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-6875703/v1/d614194d-549b-43f5-8cb7-7cdb757ebd1d.pdf"}],"financialInterests":"No competing interests reported.","formattedTitle":"The Effect of Skin Tone Classification on Bias in Bruise Detection","fulltext":[{"header":"Introduction","content":"\u003cp\u003eDespite advances in artificial intelligence (AI) based medical imaging, patients with darker skin tones remain at a higher risk of being overlooked [\u003cspan additionalcitationids=\"CR2\" citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e]. An injury caused by the same mechanism, at the same time, and at the same location can look completely different on different skin tones. AI tools designed to detect physical injuries often rely on visual features that may or may not appear the same across all skin tones (Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003e). As a result, these tools may make inaccurate detections. One of the reasons for such bias is the limited representation of skin tones in training datasets [\u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e4\u003c/span\u003e]. This kind of representation bias is well-known in dermatology-focused AI, where models trained mostly on lighter or darker skin often struggle to perform well on the other [\u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e5\u003c/span\u003e, \u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e6\u003c/span\u003e]. Similar disparities have also been reported in non-AI technologies, such as pulse oximetry, where oxygen saturation levels are systematically overestimated in patients with darker skin due to differences in light absorption [\u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e7\u003c/span\u003e, \u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e8\u003c/span\u003e]. When these tools are not evaluated across diverse populations, they can increase disparities in care and reduce confidence in the use of AI in clinical settings.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eThis challenge is especially evident in the context of bruise detection, because bruises are often subtle and vary in appearance depending on skin pigmentation. On lighter skin, bruises may appear red, purple, or yellow as they heal. On darker skin, bruises often appear deep purple or brown and may blend with the surrounding skin, making them harder to detect by both humans and AI tools [\u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e9\u003c/span\u003e]. In our previous\u003c/p\u003e \u003cp\u003es work, we found that bruises are harder to see on darker skin and to address this, we incorporated images captured under alternate light source (ALS) conditions to enhance bruise visibility during data collection [\u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e10\u003c/span\u003e]. ALS is a method that uses controlled wavelengths of light (ultraviolet, visible, or infrared ranges) to enhance visibility of bruises that are difficult to see under natural lighting [\u003cspan citationid=\"CR11\" class=\"CitationRef\"\u003e11\u003c/span\u003e]. However, improving visibility is only one part of the solution. AI systems depend on labeled examples to learn patterns, but what is labeled depends on what is visible to human annotators. If bruises are less visible on certain skin tones, they may be under-annotated or missed entirely during dataset creation. In turn, the model may not learn to detect these cases, which further increases disparities in performance.\u003c/p\u003e \u003cp\u003eIn addition to labeling bruises, fairness evaluations also depend on how skin tone is recorded. Not surprisingly, skin tone is not always measured in a consistent or objective way. In many datasets it is estimated either by using L*a*b* values (\u003cem\u003eL\u003c/em\u003e represents lightness, while \u003cem\u003ea\u003c/em\u003e and \u003cem\u003eb\u003c/em\u003e represent the red-green and blue-yellow color components respectively) from a colorimeter, or by comparing it to a predefined skin tone scale [\u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e12\u003c/span\u003e]. These measurements can change based on lighting conditions or camera quality and often depend on the individual labeling them. Without a standard method, skin tone annotations may be inaccurate and introduce bias into the dataset. This is important when trying to measure fairness because model performance is often compared across skin tone groups. If those groups are defined in a vague or inconsistent way, they can hide true differences in performance or make small gaps look larger than they are. Sometimes what looks like bias in the model may be the result of how skin tone was labeled. To evaluate fairness properly, careful consideration must be given to how skin tones are defined and grouped. There is no single correct way to define those groups. Each method groups skin tones differently and that can lead to different conclusions about fairness. Given that the underlying model can exhibit bias, this raises an important question in bruise detection: \u003cem\u003eWhich skin tone classification is the most effective at revealing those disparities?\u003c/em\u003e This study explores that question by comparing not only model accuracy, but also fairness evaluation results across two distinct approaches to skin tone classification: using existing skin tone scales and data-driven groups constructed by clustering. The central idea is that a good classification scale should make disparities visible rather than hide them. By examining the same model through multiple classification frameworks, the study aims to better understand how skin tone definitions influence bias in AI based bruise detection.\u003c/p\u003e\n\u003ch3\u003eIndividual Typology Angle (ITA) Skin Tone Classification\u003c/h3\u003e\n\u003cp\u003eThe Individual Typology Angle (ITA) is a continuous measure used to classify skin tone based on values from the CIELAB (Commission Internationale de l\u0026rsquo;Eclairage L*a*b*) color space, which is widely used in imaging and color science [\u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e13\u003c/span\u003e]. CIELAB separates color into three components: L* for lightness, a* for the red-green axis, and b* for the yellow-blue axis (Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003e, left). ITA relies specifically on the L* and b* values to calculate an angle that reflects how light or dark the skin appears (Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003e, right). The higher the angle, the lighter the skin and vice versa. ITA offers a numeric and objective way to classify skin tone without relying on human perception. In the study, L* and b* values were extracted from a selected area of exposed non-bruised skin in each participant and were used to compute an ITA score using the standard formula, Eq.\u0026nbsp;(1). Once calculated, each image was placed into one of six tone categories based on thresholds widely used in prior studies: very light (\u0026gt;\u0026thinsp;55\u0026deg;), light (41\u0026deg;- 55\u0026deg;), intermediate (28\u0026deg;- 41\u0026deg;), tan (10\u0026deg;- 28\u0026deg;), brown (-30\u0026deg;- 10\u0026deg;), and dark (\u0026thinsp;\u0026lt;\u0026thinsp;\u0026minus;\u0026thinsp;30\u0026deg;) [\u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e14\u003c/span\u003e].\u003c/p\u003e \u003cp\u003eITA \u003cem\u003e=\u003c/em\u003e \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:\\text{arctan}[\\frac{({L}^{*}-\\:50)}{{b}^{*}}]*\\frac{180}{\\pi\\:}\\)\u003c/span\u003e\u003c/span\u003e. (1)\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cdiv id=\"Sec3\" class=\"Section2\"\u003e \u003ch2\u003eFitzpatrick Skin Tone Classification\u003c/h2\u003e \u003cp\u003eThe Fitzpatrick scale was first introduced in 1975 primarily to describe how the skin responds to sunlight, and to help dermatologists make decisions in phototherapy and other dermatological treatments by estimating how likely a person\u0026rsquo;s skin is to burn or tan when exposed to the sun [\u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e15\u003c/span\u003e]. It includes six categories, starting with very light skin that burns easily (Type I) and ending with highly pigmented skin that rarely burns (Type VI) (Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003e). Even though it was meant to measure UV sensitivity, not to represent how the skin looks, multiple researchers have often used it in medical AI studies to group people by skin tone. This is due to its wide availability and its established use in dermatology, even though it does not capture the full range of visual skin color variation.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eOne major criticism of the Fitzpatrick scale is that it was developed based on a limited and non-diverse population [\u003cspan citationid=\"CR16\" class=\"CitationRef\"\u003e16\u003c/span\u003e]. More specifically, it primarily reflects the skin characteristics of lighter-skinned individuals. As a result, the darker categories (Types V and VI) were added after the original classification and were not developed with the same level of empirical validation [\u003cspan citationid=\"CR17\" class=\"CitationRef\"\u003e17\u003c/span\u003e]. This has led to concerns that the scale compresses a wide range of darker skin tones into just one or two categories, which may overlook important visual and clinical differences between individuals and introduce unintended bias. For example, in a survey of more than 2,000 Black adults, 59 percent said they could not find a skin tone on the Fitzpatrick scale that matched their own [\u003cspan citationid=\"CR18\" class=\"CitationRef\"\u003e18\u003c/span\u003e]. Furthermore, the scale is often used in a subjective way, based on how someone sees themselves or how others see them, rather than through any consistent or objective method. These inconsistencies in labeling can result in misclassification, especially in bruise datasets where lighting, image quality, and camera settings can influence how the bruise appears. In fairness evaluations, relying on the Fitzpatrick scale may therefore lead to inaccurate assessments, as it may mask disparities that exist within the broad groupings of darker skin. For example, a model may perform well on medium-dark skin but poorly on very dark skin, yet both might fall under the same Fitzpatrick type. This lack of granularity may limit the scale\u0026rsquo;s usefulness in settings where precise evaluation of model performance across skin tones is needed.\u003c/p\u003e \u003c/div\u003e\n\u003ch3\u003eMonk Skin Tone Classification\u003c/h3\u003e\n\u003cp\u003eThe Monk Skin Tone (MST) scale was recently developed to offer a more inclusive way to represent skin tone, especially in fields like computer vision and AI [\u003cspan citationid=\"CR19\" class=\"CitationRef\"\u003e19\u003c/span\u003e]. It was created by Dr. Ellis Monk, a sociologist at Harvard, in collaboration with Google Research. The scale includes ten distinct skin tones, either as spheres/orbs with gradients or flat patches, that range from very light to very dark, based on how people actually see and interpret skin color in images (Fig.\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003e). Unlike older frameworks such as the Fitzpatrick scale, the MST scale was built around visual perception. Its goal is to better reflect the diversity of real-world populations, especially those who have often been underrepresented or oversimplified in datasets.\u003c/p\u003e \u003cp\u003eWhat makes the MST scale different is that it does not treat skin tone as a biological or clinical variable, but instead focuses on how skin looks in everyday settings, whether in photos, on screens, or in videos [\u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e20\u003c/span\u003e]. It was developed through both sociological research and public perception testing to make sure the ten skin tones are clear, distinct, and reflective of how people describe skin color in everyday life. This makes it especially useful in fairness research, where subtle differences in appearance can affect how an AI model performs. In this study, MST values were assigned to each subject based on pixel-level color matching between the subject\u0026rsquo;s skin tone and the reference swatches from the MST spheres. These assignments were then used to group participants for fairness evaluations, allowing to look more closely at whether the model performed differently across the full range of skin tones.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e"},{"header":"Methods","content":"\u003cp\u003eTo evaluate algorithmic fairness in bruise detection across diverse skin tones, six skin tone classification frameworks were employed: ITA categories, binarized and full-category Fitzpatrick Skin Types (FST), the Monk Skin Tone (MST) scale, data-driven clustering using L*a*b* color values (chromatic), and clustering using only the L* (luminance) component. Binarized Fitzpatrick (grouped types I-III as light skin, IV-VI as dark) was chosen as an additional approach since it has been commonly used in literature for object detection tasks and also with fairness metrics [\u003cspan citationid=\"CR22\" class=\"CitationRef\"\u003e22\u003c/span\u003e, \u003cspan citationid=\"CR23\" class=\"CitationRef\"\u003e23\u003c/span\u003e]. By applying both chromatic and luminance-based clustering, the aim was to compare the impact of classification framework choice on model accuracy and measured fairness outcomes. These frameworks were used to stratify this study\u0026rsquo;s dataset of annotated bruise images, allowing the assessment of group-level performance disparities using standard fairness metrics.\u003c/p\u003e\n\u003ch3\u003eData Sources\u003c/h3\u003e\n\u003cp\u003eThis study utilizes a bruise dataset collected previously with support from the National Institute of Justice (NIJ; Award# 2016-DN-BX-0147). A total of 11,766 bruise and non-bruise images from 118 distinct participants were used, capturing 246 distinct bruises (Table\u0026nbsp;\u003cspan refid=\"Tab1\" class=\"InternalRef\"\u003e1\u003c/span\u003e). The images were captured in a controlled laboratory setting, where bruises were deliberately induced using either a paintball impact or a dropped weight to simulate real-world blunt force trauma [\u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e10\u003c/span\u003e]. Each injury was photographed at several time points to document how the bruise changed over time. The imaging protocol included both white lighting and ALS conditions. Each image was annotated by three independent and trained human labelers, and any potential discrepancies were resolved by a team leader. The final annotations included bounding boxes around the bruises, which were used as the ground truth for model training. All annotations were stored in a structured database for further analysis. For model development and evaluation, the dataset was split into training and testing sets in a 85:15 ratio.\u003c/p\u003e \u003cp\u003eIn addition to images, structured data was also collected for each subject, including age, gender, injury location, and skin tone. Skin tone categories in the original dataset was based on colorimetric measurements from uninjured adjacent skin, reported in L*a*b* color space [\u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e10\u003c/span\u003e]. Subjects were then assigned to one of six skin tone categories using their calculated ITA values: very light (\u0026gt;\u0026thinsp;55\u0026deg;), light (41\u0026ndash;55\u0026deg;), intermediate (28\u0026ndash;41\u0026deg;), tan (10\u0026ndash;28\u0026deg;), brown (\u0026minus;\u0026thinsp;30\u0026ndash;10\u0026deg;), and dark (\u0026thinsp;\u0026le;\u0026thinsp;\u0026minus;\u0026thinsp;30\u0026deg;). This approach enabled a consistent, measurement-based classification across the full range of pigmentation. All personally identifiable information was removed prior to analysis in accordance with HIPAA regulations. The resulting database included both annotated images and their associated metadata. A custom platform with a user interface and API is currently being developed by the team to enable multi-criteria searches and AI training of bruise images, aiming to become the largest ALS-based bruise image repository [\u003cspan citationid=\"CR24\" class=\"CitationRef\"\u003e24\u003c/span\u003e].\u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab1\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 1\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003eTest set distribution by skin color, gender, and age group.\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"4\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eCategory\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eSubjects n (%)\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eInjuries n (%)\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003eImages n (%)\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e\u0026nbsp;\u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eTest Train\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eTest Train\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003eTest Train\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eSkin Color\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eType I \u0026ndash; Very Light\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e15 (12.9) 15 (12.7)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e16 (13.1) 16 (12.9)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e289 (16.4) 1477 (14.8)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eType II - Light\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e23 (19.8) 23 (19.5)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e25 (20.5) 27 (21.8)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e373 (21.1) 2241 (22.4)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eType III - Intermediate\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e21 (18.1) 21 (17.8)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e23 (18.9) 21 (16.9)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e335 (19.0) 1864 (18.6)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eType IV - Tan\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e23 (19.8) 24 (20.3)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e23 (18.9) 24 (19.4)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e300 (16.9) 1876 (18.8)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eType V - Brown\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e17 (14.7) 17 (14.4)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e17 (13.9) 17 (13.7)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e259 (14.7) 1369 (13.7)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eType VI - Dark\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e17 (14.7) 18 (15.3)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e18 (14.7) 19 (15.3)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e210 (11.9) 1173 (11.7)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eGender\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eMale\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e34 (29.3) 35 (29.7)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e34 (27.9) 35 (28.2)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e482 (27.3) 2780 (27.8)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eFemale\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e82 (70.7) 83 (70.3)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e88 (72.1) 89 (71.8)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e1284 (72.7) 7220 (72.2)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eAge group (years)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e0\u0026ndash;18\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e25 (21.6) 25 (21.2)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e26 (21.3) 25 (20.2)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e466 (26.4) 2460 (24.6)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e19\u0026ndash;35\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e80 (68.9) 82 (69.5)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e84 (68.9) 86 (69.3)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e1164 (65.9) 6723 (67.2)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e36\u0026ndash;50\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e9 (7.8) 9 (7.6)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e10 (8.2) 11 (8.9)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e128 (7.2) 748 (7.5)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e51\u0026ndash;65\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e2 (1.7) 2 (1.7)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e2 (1.6) 2 (1.6)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e8 (0.5) 69 (0.7)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eTotal\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e\u003cb\u003e116 (100) 118 (100)\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e\u003cb\u003e122 (100) 124 (100)\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e\u003cb\u003e1766 (100) 10,000 (100)\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e\n\u003ch3\u003eBruise Detection Model\u003c/h3\u003e\n\u003cp\u003eThe bruise detection model was trained using the YOLOv5 (You Only Look Once Version 5) architecture, which is a real-time object detection model that evaluates the entire image in one pass [\u003cspan citationid=\"CR25\" class=\"CitationRef\"\u003e25\u003c/span\u003e]. By partitioning the input into a grid and predicting object locations and class labels simultaneously, YOLO avoids the sequential steps used in earlier pipelines, offering improved speed and adaptability to visual variability in bruising. Given the dataset size, training the entire model was not feasible. Instead, transfer learning was applied. This meant that the backbone of a pre-trained YOLOv5 model was kept fixed while the detection-specific layers were fine-tuned on the dataset [\u003cspan citationid=\"CR26\" class=\"CitationRef\"\u003e26\u003c/span\u003e]. This approach allowed the model to build on general visual features while learning specific cues relevant to bruise identification. Each image was processed to extract the ground-truth bounding boxes by converting YOLO's normalized coordinate format into absolute pixel dimensions. Following this, images were normalized and transformed into tensor representations compatible with the model input requirements. The trained YOLO network then produced predicted bruise masks. Bounding boxes were then produced during model training, along with associated confidence and the intersection over union (IoU) scores for each detection. The confidence score indicates how strongly the model believes a detected region contains a bruise while IoU measures the overlap between the predicted box and the ground truth [\u003cspan citationid=\"CR27\" class=\"CitationRef\"\u003e27\u003c/span\u003e, \u003cspan citationid=\"CR28\" class=\"CitationRef\"\u003e28\u003c/span\u003e]. The YOLOv5 model was selected based on a detailed investigation of several object detection architectures, which is out of scope for this publication.\u003c/p\u003e \u003cdiv id=\"Sec8\" class=\"Section2\"\u003e \u003ch2\u003eSkin Tone Categorization and Mapping\u003c/h2\u003e \u003cp\u003eTo enable comparisons across other labeling systems, Fitzpatrick and MST classifications were added using a nearest-neighbor approach based on the L*a*b* color space (Fig.\u0026nbsp;\u003cspan refid=\"Fig5\" class=\"InternalRef\"\u003e5\u003c/span\u003e). For each scale, representative skin tone swatches were collected, i.e. Fitzpatrick types I\u0026ndash;VI and MST orbs 1\u0026ndash;10, and extracted RGB pixel values from each swatch image. These RGB values were normalized and converted to CIELAB format using a standard perceptual transformation implemented in the color space conversion package in Python v 3.11.9 [\u003cspan citationid=\"CR29\" class=\"CitationRef\"\u003e29\u003c/span\u003e, \u003cspan citationid=\"CR30\" class=\"CitationRef\"\u003e30\u003c/span\u003e]. Each subject\u0026rsquo;s L*a*b* values were then pulled from existing metadata and compared to the full set of reference tones using Euclidean distance. In cases where multiple swatches were similarly close in the L*a*b* space, ties were resolved by assigning the label with the lowest average distance among its nearest neighbors (k\u0026thinsp;=\u0026thinsp;10). Fitzpatrick values were then mapped to its respective categories, while MST values were kept in their original numeric form. ITA categories were already present in the dataset and were used without modification.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003c/div\u003e\n\u003ch3\u003eSkin Tone Clustering\u003c/h3\u003e\n\u003cp\u003eRather than relying on predefined skin tone classifications, which may not reflect the full variation in a bruise dataset, groupings were derived directly from the skin color data. This allowed us to examine whether the model behaved differently across empirically defined clusters, based on how the skin \u003cem\u003eactually\u003c/em\u003e appeared, rather than how it was categorized. For this, two clustering methods were used:\u003c/p\u003e \u003cp\u003e \u003col\u003e \u003cspan\u003e \u003cli\u003e \u003cp\u003e \u003cb\u003eChromatic Clustering\u003c/b\u003e: This approach grouped participants based on their full color profile using all three components of the L*a*b* color space. This captured both brightness and undertone. The goal here was to identify whether combinations of hue and luminance of the participant\u0026rsquo;s skin tone played a role in bruise visibility and whether the model's detections varied across those combinations. Since bruise appearance is not just about how dark it is, but also how its color blends with or stands out from the surrounding skin, clustering by full color information provided a way to account for that complexity.\u003c/p\u003e \u003c/li\u003e \u003c/span\u003e \u003cspan\u003e \u003cli\u003e \u003cp\u003e \u003cb\u003eLuminance Clustering\u003c/b\u003e: This second approach included only the L* channel from the CIELAB color wheel, which represents brightness. This was used to isolate whether variation in lightness alone was sufficient to explain changes in model performance. This approach was driven by the observation that L* showed the most variability in our dataset and largely defined the clustering structure, whereas a* and b* contributed little separation. By removing chromatic components and focusing solely on luminance, detection disparities could be tested to determine whether they were driven by brightness differences alone.\u003c/p\u003e \u003c/li\u003e \u003c/span\u003e \u003c/ol\u003e \u003c/p\u003e \u003cp\u003eBoth clustering approaches were first performed on the full training set, using only participants with complete L*a*b* data in order to generate stable and representative cluster definitions. For each method, agglomerative hierarchical clustering was used and the optimal number of clusters was selected based on silhouette score analysis [\u003cspan citationid=\"CR31\" class=\"CitationRef\"\u003e31\u003c/span\u003e]. The average silhouette scores were then computed for each of clustering methods across a range of cluster counts. The score is higher when points are well matched to their own cluster and far from others. The number of clusters (k) was systematically varied from two to ten, and the optimal value was identified by selecting the k that yielded the highest silhouette score. Once clusters were established, the same clustering assignments were applied to the held-out test set by mapping test set points to the nearest training-derived cluster centroids in the corresponding feature space. This ensured that the test set was grouped consistently with the clusters identified with the training data.\u003c/p\u003e\n\u003ch3\u003eModel Evaluation and Fairness Metrics\u003c/h3\u003e\n\u003cp\u003eOnce the clusters were established, accuracy and model fairness was evaluated across skin tone groupings using two widely used metrics, Demographic Parity Difference (DPD) and Equal Opportunity Difference (EOD) (Table\u0026nbsp;\u003cspan refid=\"Tab2\" class=\"InternalRef\"\u003e2\u003c/span\u003e). These metrics were chosen because they specifically assess differences in model predictions across groups and were suitable for the distribution of outcomes in our dataset [\u003cspan citationid=\"CR32\" class=\"CitationRef\"\u003e32\u003c/span\u003e, \u003cspan citationid=\"CR33\" class=\"CitationRef\"\u003e33\u003c/span\u003e]. In the fairest case for these metrics, the observed difference is zero, indicating no measurable disparity across skin tone groups. Other fairness metrics such as Equalized Odds, and performance metrics such as precision, recall, and AUC, were not used due to the limited number of non-bruise images in the test set, which prevented reliable comparisons across skin tone groups.\u003c/p\u003e \u003cp\u003eFairness evaluations were conducted across six skin tone classification methods: ITA, Fitzpatrick, Fitzpatrick binarized, MST, and the two clustering-based groupings (chromatic and luminance). Each image was assigned to a skin tone group based on its labeling method, and those groupings were held constant during evaluation. For the fairness and accuracy metrics, predictions were binarized using fixed thresholds. A prediction (\u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:\\widehat{Y})\\:\\)\u003c/span\u003e\u003c/span\u003ewas considered valid if the confidence score was at least 0.4 and the IoU was 0.6. These thresholds were selected based on prior analysis, which showed that values in the range of 0.2\u0026ndash;0.6 for confidence and 0.5\u0026ndash;0.7 for IoU offered the best balance between performance and fairness. Standardizing thresholds allowed us to compare fairness outcomes consistently across all grouping methods. In contrast, average precision (AP) was also calculated without applying a confidence threshold. AP is a standard performance metric used commonly in object detection tasks. It measures the area under the precision-recall curve for a single class, summarizing how well the model balances precision and recall across all confidence levels [\u003cspan citationid=\"CR34\" class=\"CitationRef\"\u003e34\u003c/span\u003e]. Mean average precision (mAP) further aggregates AP across all detection tasks to give a summary measure of performance.\u003c/p\u003e \u003cp\u003eA skin tone scale was considered good if it showed a fluctuation between skin tones for the considered metrics. The goal was not to optimize the model, but to find a skin tone scale that \u0026ldquo;breaks\u0026rdquo; the model by showing potential bias. Thus, the larger the fluctuation in accuracy and fairness metrics, the more appropriate the skin tone scale is for evaluating models.\u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab2\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 2\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003eFairness metric formulas and descriptions used to assess group-level disparities in bruise detection.\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"3\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eMetric\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eDescription\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eFormula\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eDemographic Parity Difference (DPD)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eDifference in the rate of positive predictions across groups, regardless of actual outcome.\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e\u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:\\text{D}\\text{P}\\:\\stackrel{\\scriptscriptstyle\\text{def}}{=}{P\\left(\\widehat{Y}=1∣A=a\\right)\\:-P\\left(\\widehat{\\:Y}=1\\right)}^{*}\\)\u003c/span\u003e\u003c/span\u003e\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eEqual Opportunity Difference (EOD)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eDifference in true positive rates across groups, conditional on the outcome being positive.\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e\u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:\\text{E}\\text{O}\\:\\stackrel{\\scriptscriptstyle\\text{def}}{=}{P\\left(\\widehat{Y}=1∣A=a\\right)\\:-P\\left(\\widehat{\\:Y}=1\\right)}^{*}\\)\u003c/span\u003e\u003c/span\u003e\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003cp\u003e* \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:\\widehat{Y}\\)\u003c/span\u003e\u003c/span\u003e denotes the model\u0026rsquo;s predicted label (1\u0026thinsp;=\u0026thinsp;bruise detected); \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:Y\\)\u003c/span\u003e\u003c/span\u003e denotes the ground truth label (1\u0026thinsp;=\u0026thinsp;actual bruise present). A\u0026thinsp;=\u0026thinsp;a indicates that the individual belongs to skin tone group \u003cem\u003ea\u003c/em\u003e. P represents conditional probability.\u003c/p\u003e"},{"header":"Results","content":"\u003cp\u003eAfter mapping the data to its respective Fitzpatrick and MST scales, the distribution of skin tone labels differed noticeably across systems (Fig.\u0026nbsp;\u003cspan refid=\"Fig6\" class=\"InternalRef\"\u003e6\u003c/span\u003e). The raw ITA labels were spread more evenly, with many images falling into Light, Intermediate, and Tan categories. MST labels were clustered at the lighter end, mostly in Types 2 and 3, with few images in darker categories. No images in the dataset were labeled as MST Types 9 or 10. Only a small number of images were labeled in the darker MST categories. The Fitzpatrick distribution showed a similar skew, with most images assigned to Type I, fewer in Types IV and V, and none in Type VI.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cdiv id=\"Sec12\" class=\"Section2\"\u003e \u003ch2\u003eClustering\u003c/h2\u003e \u003cp\u003eChromatic clustering based on L*, a*, and b* values produced the highest silhouette score, 0.62, at k\u0026thinsp;=\u0026thinsp;2 (Fig.\u0026nbsp;\u003cspan refid=\"Fig7\" class=\"InternalRef\"\u003e7\u003c/span\u003e). Scores decreased steadily as the number of clusters increased, with silhouette values dropping to 0.38 by k\u0026thinsp;=\u0026thinsp;10. By k\u0026thinsp;=\u0026thinsp;7, the silhouette score had dropped to 0.41 and did not show further improvement beyond that point. As a result, seven clusters were selected as the final grouping for analysis. In contrast, clustering based on L* alone resulted in consistently higher silhouette scores across the same range of cluster counts. From the silhouette score graph, k\u0026thinsp;=\u0026thinsp;4 is where the plot decreases or flattens with minimal drop-off through adjacent values. The corresponding dendrogram showed clear hierarchical separation, with branch splits occurring at higher distances and greater vertical spacing between cluster boundaries. In the dendrogram constructed from the hierarchical linkage matrix, the default visualization in the Python package separated the data into three main clusters, corresponding to the largest linkage distances. However, finer-grained splits at lower distances were evident throughout the structure, and the clusters were chosen based on the silhouette score metric (Fig.\u0026nbsp;\u003cspan refid=\"Fig8\" class=\"InternalRef\"\u003e8\u003c/span\u003e). To visualize the skin tone distributions captured by both clustering methods, the average L*a*b* values were calculated for each group. These values were then converted into representative color swatches, forming a continuous scale from lighter to darker tones. The resulting palettes illustrates the range of skin tones present in the dataset as identified by the clustering process (Fig.\u0026nbsp;\u003cspan refid=\"Fig9\" class=\"InternalRef\"\u003e9\u003c/span\u003e).\u003c/p\u003e \u003cp\u003eIn Fig.\u0026nbsp;\u003cspan refid=\"Fig10\" class=\"InternalRef\"\u003e10\u003c/span\u003e, the distributions of all six classification methods in L*a*b* space is shown in panels A to F. Across all labeling systems, the data points form a narrow, curved band that runs primarily along the L* axis. Depending on the classification method, there is little to no variance between the chromatic dimensions (a* and b*). In comparing the classification methods, it was also found that the same image was often assigned to different skin tone categories depending on which system was used. Each system organizes skin tones in a way that mirrors variation in brightness more than color. For example, an image labeled as \u0026ldquo;intermediate\u0026rdquo; in the ITA scale could be placed in a different group under MST or fall near a boundary in chromatic clustering. These inconsistencies were common and raised concerns about the alignment between classification frameworks. This further supported the need for data-driven grouping strategies and justified the inclusion of L*-based clustering alongside the chromatic approach. Since the dataset with the raw ITA values already showed a strong tendency to separate by L* (Panel A), the study also tested whether brightness alone could form meaningful and consistent groupings.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec13\" class=\"Section2\"\u003e \u003ch2\u003ePerformance Metrics\u003c/h2\u003e \u003cp\u003eFor most scales, accuracy was higher in lighter skin tone groups and lower in darker ones (Table\u0026nbsp;\u003cspan refid=\"Tab3\" class=\"InternalRef\"\u003e3\u003c/span\u003e). In the Monk Skin Tone scale, accuracy dropped from 0.90 in group 1 to 0.78 in group 8, with some fluctuation in later groups. The full Fitzpatrick scale showed a similar pattern, with accuracy falling from 0.86 in type I to 0.67 in type V. When the Fitzpatrick scale was binarized into light and dark groups, the light group still performed better (0.84 vs. 0.72). The ITA scale also showed a decline, from 0.91 in very light tones to 0.73 in dark tones. Chromatic clusters had more variation, with lower accuracy in some mid and darker clusters like C2, C5, and C6. Luminance clusters showed the opposite trend with the lightest cluster having the lowest accuracy. Overall, the model tended to be more accurate on lighter tones, and performance dropped for darker skin across most scales.\u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab3\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 3\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003eSkin tone group accuracies for each classification method at IoU\u0026thinsp;\u0026ge;\u0026thinsp;0.6 and confidence\u0026thinsp;\u0026ge;\u0026thinsp;0.4.\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"3\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eScale\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eGroup\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eAccuracy\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eIndividual Typology Angle\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eVery Light\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.91\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eLight\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.84\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eIntermediate\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.85\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eTan\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.84\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eBrown\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e\u003cb\u003e0.76\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eDark\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e\u003cb\u003e0.73\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eFitzpatrick\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eI\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.86\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eII\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.80\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eIII\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.83\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eIV\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.74\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eV\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e\u003cb\u003e0.67\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eFitzpatrick Binarized\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eLight\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.84\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eDark\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e\u003cb\u003e0.72\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eMonk Skin Tone\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e1\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.90\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e2\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.87\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e3\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.82\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e4\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.75\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e5\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.67\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e6\u003c/p\u003e \u003cp\u003e7\u003c/p\u003e \u003cp\u003e8\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.93\u003c/p\u003e \u003cp\u003e\u003cb\u003e0.71\u003c/b\u003e\u003c/p\u003e \u003cp\u003e0.78\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eChromatic Clusters\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eC0\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.85\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eC1\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.83\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eC2\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.74\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eC3\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.86\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eC4\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.89\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eC5\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e\u003cb\u003e0.71\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eC6\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.74\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eLuminance Clusters\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eL0\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e\u003cb\u003e0.72\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eL1\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.83\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eL2\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.87\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eL3\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.81\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003cp\u003eThese results can be further validated by the AP results below (Fig.\u0026nbsp;\u003cspan refid=\"Fig11\" class=\"InternalRef\"\u003e11\u003c/span\u003e). Patterns in average precision generally mirrored the accuracy trends. Most classification methods showed higher AP values for lighter skin tone groups and a gradual decline for darker ones. The Fitzpatrick and ITA categories followed a similar pattern, with the darkest tones showing the lowest AP. Although binarized Fitzpatrick produced more stable results, the light group still outperformed the dark group. MST and Chromatic clusters showed greater fluctuation, particularly in mid-range tones. In contrast, luminance clusters showed less variation overall but showed an increase towards the darkest cluster.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec14\" class=\"Section2\"\u003e \u003ch2\u003eFairness Metrics\u003c/h2\u003e \u003cp\u003eFairness evaluation was assessed using DPD and EOD gaps across all six skin tone classification methods: ITA, Fitzpatrick, MST, binarized Fitzpatrick, chromatic clusters, and luminance-based clusters. In this context, a \"gap\" refers to the difference between a specific group's prediction rate and the overall average prediction rate across all skin tone groups. While each system defines skin tone categories differently, they all follow the same basic structure by organizing individuals from lighter to darker tones. For comparability, they were aligned along a shared axis. All groups were mapped to a normalized skin tone scale, ranging from 0 (lightest) to 1 (darkest).\u003c/p\u003e \u003cp\u003eAcross both DPD and EOD, a noticeable decline in gaps appeared towards the end/darker tones of the normalized skin tone scale for several classification methods (Fig.\u0026nbsp;\u003cspan refid=\"Fig11\" class=\"InternalRef\"\u003e11\u003c/span\u003e). The graphs show that fairness outcomes are not always linear across the skin tone spectrum and that mid-tone groups may not experience the same levels of disparity as lighter or darker groups. Gap values decreased from lighter to darker groups in both metrics, with the most negative values often occurring at the darkest end of the scale, particularly for Fitzpatrick. MST and chromatic clusters showed the largest drops in the mid-to-dark range. In contrast, both luminance clustering and ITA showed relatively smoother trends across, with gap values decreasing more gradually and remaining relatively close to zero across most tone groups. Luminance clustering showed a decline in fairness gaps for darker skin tones despite showing the highest accuracy and AP for that group. Binary Fitzpatrick, due to its two-group structure, produced stable values across the tone scale, with both DPD and EOD gaps remaining moderate.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003c/div\u003e"},{"header":"Discussion","content":"\u003cp\u003eThis study evaluated the impact of skin tone classification on model performance and fairness by holding the model and data constant while varying the labeling scale. Most of the classification methods, specifically ITA, MST, Fitzpatrick full and binarized, and chromatic clustering, showed higher accuracy and AP in lighter skin tone groups and lower in darker ones. In scales with more granularity, like the MST, the decrease in accuracy and AP was not smooth. The only exception was luminance clustering, where they were lower in lighter groups and improved in the darker ones. Despite this, fairness dropped for darker skin tones. This difference may be because luminance clustering groups skin tones based only on brightness, ignoring color differences that affect how bruises appear. These shifts show that even with the same model and data, the way skin tone is defined can change model performance and that matters when evaluating its fairness. Relying only on performance metrics such as accuracy may not always mean that fairness would be high. Testing and selecting the correct scale considering both aspects, accuracy and fairness, becomes essential in bruise model and bias detection.\u003c/p\u003e \u003cp\u003eModel accuracy, AP, and both DPD and EOD fairness metrics revealed that detection disparities tend to worsen as skin tone darkens under all scales, and especially under classification methods with coarse or poorly aligned groupings like ITA and Fitzpatrick full and binarized. The closer the fairness metric values are to zero, the fairer the model is. When lower DPD values appear for darker skin tones, it indicates that the model is less likely to make positive predictions in those groups compared to the overall average. EOD, which measures correct bruise detections among cases where a bruise is actually present, showed similar gaps. Lower EOD values for darker tones suggest that even when bruises exist, the model is more likely to miss them. This is particularly important in clinical contexts, where missed detections can result in bruises going undocumented, potentially affecting follow-up care and decision-making.\u003c/p\u003e \u003cp\u003eThe classifications with the largest fluctuations, MST and chromatic clustering, categorize skin into more specific skin tone groups. These finer groupings do not cause disparities, but they do expose them more clearly by identifying groups for which models underperform. In contrast, binarized Fitzpatrick produced more stable DPD and EOD gaps. But this stability may be misleading as broader groupings may average out performance differences within each group, masking disparities that exist between medium and very dark tones. Likewise, although luminance clustering produced relatively smooth trends, its reliance on lightness (L*) alone, without incorporating chromatic dimensions (a* and b*), may oversimplify how skin tone affects bruise visibility, which may potentially limit its fairness in more diverse populations.\u003c/p\u003e \u003cp\u003eIn most classification methods, both DPD and EOD appeared more stable for mid-range skin tones. This suggests that the model may be most stable for medium tones under all scales, while its predictions become less consistent at the lighter and darker ends. Instead of a simple increase in bias with darker skin, the model appears to struggle more at both extremes, regardless of the scale. With positive values towards lighter skin tones, lighter skin receives more favorable treatment than average, whereas darker skin receives less favorable. The results also show a clear tradeoff between how detailed a skin tone scale is and how practical it is to use. Scales with more categories, like MST or chromatic clusters, helped show exactly where the model\u0026rsquo;s accuracy dropped. But they also made the results more variable and harder to interpret. Simpler groupings, like binary Fitzpatrick or luminance clusters, gave more stable results, but likely missed differences by grouping skin tones together. This means that using broader categories can make a model look more consistent and fairer than it really is, especially if it's underperforming for specific groups. To improve fairness, future efforts should focus on both better skin tone representation in datasets and refining how skin tone groupings are defined and used in evaluation. This includes more detailed annotations for images where bruises are harder to see, especially on darker skin.\u003c/p\u003e \u003cdiv id=\"Sec16\" class=\"Section2\"\u003e \u003ch2\u003eLimitations\u003c/h2\u003e \u003cp\u003eThis study has several limitations that should be considered when interpreting the results. First, the dataset remains under development, and while it currently includes over 11,000 bruise and non-bruise images, fewer images fell in the very dark spectrum of the skin tone scales. As a result, no images were mapped to MST 9 or 10 and Fitzpatrick Type VI skin tones. Additional data collection has been ongoing to address this imbalance. Second, mapping the dataset across the different labeling systems produced inconsistent results. For example, the same image would receive two very different skin tone labels depending on the method used. This problem is also reported in [\u003cspan citationid=\"CR35\" class=\"CitationRef\"\u003e35\u003c/span\u003e]. Hence, each pixel was matched to determine the closest match with our data. This variation may make it difficult to establish a single ground truth, which may also impact fairness differently.\u003c/p\u003e \u003cp\u003eAdditionally, the dataset was collected in a controlled setting with deliberate inducing of bruises. This may not capture the complex environments found in clinical settings. Lighting, camera differences, and natural variation in injury appearance could all influence model performance outside the lab. Testing in real-world conditions is needed to confirm whether these fairness patterns hold. Finally, and most importantly, the true box annotations for the bruises depend on if its visible to humans. Since the data was collected in the lab, the collection team was aware if there is a bruise or not. But when the tool is applied to real-world settings, bruises on dark skin might be missed due to visibility and not captured at all. Future work should focus on improving data diversity, labeling and annotation consistency, and real-world testing to ensure more equitable bruise detection.\u003c/p\u003e \u003c/div\u003e"},{"header":"Conclusion","content":"\u003cp\u003eThis study highlights the critical role that skin tone classification plays in evaluating model performance and fairness in AI-based bruise detection systems. The findings make it clear that the choice of labeling framework significantly shapes how disparities are revealed and addressed. When more coarse classification methods, such as the MST and chromatic clusters scale, were used, performance gaps became more visible across the full scale. In contrast, broader grouping strategies, including binary Fitzpatrick and luminance-based clustering, appeared to minimize these gaps, not necessarily by improving model performance, but perhaps by averaging over variation that may reflect meaningful clinical differences.\u003c/p\u003e \u003cp\u003eThe focus of this paper was not to reduce bias, but to select the best approach to detect it. Most fairness research in medical AI focuses on mitigation, resampling, reweighting, or using ensemble models to equalize outcomes across groups [\u003cspan citationid=\"CR36\" class=\"CitationRef\"\u003e36\u003c/span\u003e]. These methods assume that group labels are already valid and meaningful. But if the categories themselves are poorly defined or too coarse, they can hide real differences in model behavior. Our results show that depending on which skin tone scale is used, model performance can appear stable or vary significantly across groups. This suggests that fairness interventions that don\u0026rsquo;t consider how skin tone is defined may miss important disparities altogether. By systematically comparing six skin tone classification methods, this study provides evidence that fairness evaluations must be grounded in thoughtful and transparent choices about how subgroups are defined.\u003c/p\u003e"},{"header":"Abbreviations","content":"\u003cdiv class=\"DefinitionList\"\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003eALS\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003eAlternate Light Source\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003eITA\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003eIndividual Typology Angle\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003eCIELAB\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003eCommission Internationale de l\u0026rsquo;Eclairage L*a*b*\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003eMST\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003eMonk Skin Tone\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003eDPD\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003eDemographic Parity Difference\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003eEOD\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003eEqual Opportunity Difference\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003eIoU\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003eIntersection over Union\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003eAP\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003eAverage Precision\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003emAP\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003eMean Average Precision\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003eFST\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003eFitzpatrick Skin Type\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003eRGB\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003eRed, Green, Blue\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003eHIPAA\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003eHealth Insurance Portability and Accountability Act\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003eYOLO\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003eYou Only Look Once.\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003c/div\u003e"},{"header":"Declarations","content":"\u003cp\u003e\u003cstrong\u003eEthics approval and consent to participate\u0026nbsp;\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe George Mason University IRB (#\u003cem\u003eSTUDY00000129\u003c/em\u003e) determined that the proposed activity is not research involving human subjects as defined by the U.S. DHHS and FDA regulations. The informed consent is waived by the IRB (#\u003cem\u003eSTUDY00000129)\u003c/em\u003e. This study complies with the principles of the Declaration of Helsinki.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eConsent for publication\u003c/strong\u003e\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eNot applicable\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eAvailability of data and material\u003c/strong\u003e\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eThe datasets during and/or analyzed during the current study is available from the corresponding author on reasonable request.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eCompeting interests\u0026nbsp;\u003c/strong\u003e\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eThe authors declare that they have no competing interests.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eFunding\u003c/strong\u003e\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eThe project was supported by the Artificial Intelligence/Machine Learning Consortium to Advance Health Equity and Researcher Diversity (AIM-AHEAD), funded by the National Institutes of Health (Grant #1OT2OD032581-02-811).\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eAuthor\u0026rsquo;s contributions\u0026nbsp;\u003c/strong\u003e\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eD.D. is the primary author of the manuscript. J.W. contributed to the writing and supervised the development of the methods and experimental design. K.S. contributed to the writing and interpretation of data. M.G. contributed to the writing and developed the bruise detection model. D.L. supported the implementation of the deep learning components. A.N. contributed to the design and analysis of fairness metrics.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eAcknowledgments\u003c/strong\u003e\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eNot applicable. \u0026nbsp;\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eAuthors\u0026rsquo; information\u0026nbsp;\u003c/strong\u003e\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e\u003csup\u003e1\u003c/sup\u003eHealth Informatics Program, College of Public Health, George Mason University, Fairfax, VA, USA.\u003c/p\u003e\n\u003cp\u003e\u003csup\u003e2\u003c/sup\u003eDepartment of Civil, Environmental, and Infrastructure Engineering, George Mason University, Fairfax, VA, USA.\u003c/p\u003e\n\u003cp\u003e\u003csup\u003e3\u003c/sup\u003eDepartment of Nursing, College of Public Health, George Mason University, Fairfax, VA, USA.\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\n\u003cli\u003eO\u0026rsquo;Malley A, Veenhuizen M, Ahmed A. Ensuring appropriate representation in AI-generated medical imagery: A Methodological Approach to Address Skin Tone Bias (Preprint). JMIR AI. 2024 Mar 11;\u003c/li\u003e\n\u003cli\u003eBevan P, Atapour-Abarghouei A. Detecting Melanoma Fairly: Skin Tone Detection and Debiasing for Skin Lesion Classification [Internet]. Available from: https://arxiv.org/pdf/2202.02832.\u003c/li\u003e\n\u003cli\u003eDaneshjou R, Vodrahalli K, Novoa RA, Jenkins M, Liang W, Rotemberg V, et al. Disparities in dermatology AI performance on a diverse, curated clinical image set. Science Advances. 2022 Aug 12;8(32).\u003c/li\u003e\n\u003cli\u003eNorori N, Hu Q, Aellen FM, Faraci FD, Tzovara A. Addressing bias in big data and AI for health care: A call for open science. Patterns [Internet]. 2021 Oct 8;2(10):100347. Available from: https://pmc.ncbi.nlm.nih.gov/articles/PMC8515002/.\u003c/li\u003e\n\u003cli\u003eTowards fairness in AI for melanoma detection: systemic review and recommendations [Internet]. [cited 2025 Mar 13]. Available from: https://arxiv.org/html/2411.12846v1. \u003c/li\u003e\n\u003cli\u003eGroh M, Harris C, Soenksen L, Lau F, Han R, Kim A, et al. Evaluating deep neural networks trained on clinical images in dermatology with the Fitzpatrick 17k dataset [Internet]. arXiv; 2021 [cited 2025 Mar 13]. Available from: http://arxiv.org/abs/2104.09957. \u003c/li\u003e\n\u003cli\u003eAl-Halawani R, Charlton PH, Qassem M, Kyriacou PA. A review of the effect of skin pigmentation on pulse oximeter accuracy. Physiol Meas. 2023;44(5):05TR01. Available from: https://pmc.ncbi.nlm.nih.gov/articles/PMC10391744/. \u003c/li\u003e\n\u003cli\u003eSjoding MW, Dickson RP, Iwashyna TJ, Gay SE, Valley TS. Racial bias in pulse oximetry measurement. N Engl J Med. 2020;383(25):2477\u0026ndash;8. Available from: https://www.nejm.org/doi/full/10.1056/NEJMc2029240. \u003c/li\u003e\n\u003cli\u003eKing LM. An overview of bruises [Internet]. WebMD; 2024 May [cited 2025 Mar 13]. Available from: https://www.webmd.com/skin-problems-and-treatments/bruises-article. \u003c/li\u003e\n\u003cli\u003eScafide KN, Sheridan DJ, Downing NR, Hayat MJ. Detection of Inflicted Bruises by Alternate Light: Results of a Randomized Controlled Trial. Journal of Forensic Sciences. 2020 Feb 3;65(4):1191\u0026ndash;8.\u003c/li\u003e\n\u003cli\u003eScafide KN, Downing NR, Kutahyalioglu NS, Sheridan DJ, Neil, Hayat MJ. Predicting alternate light absorption in areas of trauma based on degree of skin pigmentation: Not all wavelengths are equal. Forensic Science International. 2022 Jul 1;339:111410\u0026ndash;0.\u003c/li\u003e\n\u003cli\u003eJiang L, Wang H, Gao C, Zhang X, Xiao K, Melgosa M, et al. Skin color measurements before and after two weeks of sun exposure. Vision Research. 2022 Mar; 192:107976.\u003c/li\u003e\n\u003cli\u003eLy BCK, Dyer EB, Feig JL, Chien AL, Del Bino S. Research Techniques Made Simple: Cutaneous Colorimetry: A Reliable Technique for Objective Skin Color Measurement. Journal of Investigative Dermatology [Internet]. 2020 Jan;140(1):3-12.e1. Available from: https://www.researchgate.net/publication/338303610_Research_Techniques_Made_Simple_Cutaneous_\u003cbr\u003eColorimetry_A_Reliable_Technique_for_Objective_Skin_Color_Measurement.\u003c/li\u003e\n\u003cli\u003eOsto M, Hamzavi IH, Lim HW, Kohli I. Individual Typology Angle and Fitzpatrick Skin Phototypes are Not Equivalent in Photodermatology. Photochemistry and Photobiology. 2021 Nov 26;\u003c/li\u003e\n\u003cli\u003eWare OR, Dawson JE, Shinohara MM, Taylor SC. Racial limitations of fitzpatrick skin type. Cutis [Internet]. 2020 Feb 1;105(2):77\u0026ndash;80. Available from: https://pubmed.ncbi.nlm.nih.gov/32186531/.\u003c/li\u003e\n\u003cli\u003eHecht M. What Are the Fitzpatrick Skin Types? [Internet]. Healthline. Healthline Media; 2019. Available from: https://www.healthline.com/health/beauty-skin-care/fitzpatrick-skin-types.\u003c/li\u003e\n\u003cli\u003eNarla S, Heath CR, Alexis A, Silverberg JI. Racial disparities in dermatology. Archives of Dermatological Research. 2022 Dec 12;315(5).\u003c/li\u003e\n\u003cli\u003ePichon LC, Landrine H, Corral I, Hao Y, Mayer JA, Hoerster KD. Measuring skin cancer risk in African Americans: is the Fitzpatrick Skin Type Classification Scale culturally sensitive? Ethnicity \u0026amp; Disease [Internet]. 2010;20(2):174\u0026ndash;9. Available from: https://pubmed.ncbi.nlm.nih.gov/20503899/.\u003c/li\u003e\n\u003cli\u003eMonk E. The Monk Skin Tone Scale [Internet]. SocArXiv; 2023. Available from: osf.io/preprints/socarxiv/pdf4c_v1. \u003c/li\u003e\n\u003cli\u003eSchumann C, Olanubi, Gbolahan O, Wright A, Jr M, Heldreth C, Ricco S. Consensus and Subjectivity of Skin Tone Annotation for ML Fairness [Internet]. arXiv.org. 2023. Available from: https://arxiv.org/abs/2305.09073.\u003c/li\u003e\n\u003cli\u003eSkin Tone Research @ Google [Internet]. skintone.google. Available from: https://skintone.google/get-started.\u003c/li\u003e\n\u003cli\u003ePakzad A, Abhishek K, Hamarneh G. CIRCLe: Color Invariant Representation Learning for Unbiased Classification of Skin Lesions [Internet]. arXiv.org. 2022. Available from: https://arxiv.org/abs/2208.13528.\u003c/li\u003e\n\u003cli\u003eWilson B, Hoffman J, Morgenstern J. Predictive Inequity in Object Detection [Internet]. 2019 Feb. Available from: https://arxiv.org/pdf/1902.11097.\u003c/li\u003e\n\u003cli\u003eWojtusiak J, Qodrati M, Markiewicz M, Aminfar K, Lattanzi D, Scafide K. Interdisciplinary platform for bruise image research. Presented at: AMIA 2024 Annual Symposium; 2024 Nov 9\u0026ndash;13; San Francisco, CA. Available from: https://knowledge.amia.org/A2024/pdf/a2024a332/a2024fl332. \u003c/li\u003e\n\u003cli\u003eYOLO models for Object Detection Explained [Yolov8 Updated] [Internet]. encord.com. 2024. Available from: https://encord.com/blog/yolo-object-detection-guide/.\u003c/li\u003e\n\u003cli\u003eUltralytics. YOLO11 NEW [Internet]. [cited 2025 Mar 13]. Available from: https://docs.ultralytics.com/models/yolo11. \u003c/li\u003e\n\u003cli\u003eShroff M. Object detection with deep learning 2023: beginner\u0026rsquo;s friendly key terms explanation [Internet]. Medium; 2023 [cited 2025 Mar 13]. Available from: https://medium.com/@shroffmegha6695/object-detection-with-deep-learning-beginners-friendly-key-terms-explanation-d4fb594fea83. \u003c/li\u003e\n\u003cli\u003ePal M, Pokhriyal S, Sikdar S, Ganguly N. Ensuring generalized fairness in batch classification. Sci Rep. 2023;13(1):18892. Available from: https://www.nature.com/articles/s41598-023-45943-1. \u003c/li\u003e\n\u003cli\u003ePython Release Python 3.11.9 [Internet]. Python.org. Available from: https://www.python.org/downloads/release/python-3119/.\u003c/li\u003e\n\u003cli\u003eskimage.color \u0026mdash; skimage 0.25.2 documentation [Internet]. Scikit-image.org. 2025 [cited 2025 Jun 11]. Available from: https://scikit-image.org/docs/0.25.x/api/skimage.color.html#skimage.color.rgb2lab.\u003c/li\u003e\n\u003cli\u003eDinh DT, Fujinami T, Huynh VN. Estimating the Optimal Number of Clusters in Categorical Data Clustering by Silhouette Coefficient. Communications in Computer and Information Science [Internet]. 2019 [cited 2021 Mar 10];1\u0026ndash;17. Available from: https://link.springer.com/chapter/10.1007%2F978-981-15-1209-4_1.\u003c/li\u003e\n\u003cli\u003eLei H, Gohari A, Farnia F. On the Inductive Biases of Demographic Parity-based Fair Learning Algorithms [Internet]. arXiv; 2024 [cited 2025 Mar 13]. Available from: http://arxiv.org/abs/2402.18129.\u003c/li\u003e\n\u003cli\u003eHardt M, Price E, Price E, Srebro N. Equality of Opportunity in Supervised Learning. In: Advances in Neural Information Processing Systems [Internet]. Curran Associates, Inc.; 2016 [cited 2025 Mar 13]. Available from: https://proceedings.neurips.cc/paper/2016/hash/9d2682367c3935defcb1f9e247a97c0d-Abstract.html.\u003c/li\u003e\n\u003cli\u003eHenderson P, Ferrari V. End-to-end training of object class detectors for mean average precision. arXiv:160703476 [cs] [Internet]. 2017 Mar 16 [cited 2022 May 16]; Available from: https://arxiv.org/abs/1607.03476.\u003c/li\u003e\n\u003cli\u003eCheng WC. Monastic Color Reproduction: A Software Tool for Printing and Assessing the Monk Skin Tone Scale. Journal of Imaging Science and Technology. 2024 Sep 1;68(5):1\u0026ndash;13.\u003c/li\u003e\n\u003cli\u003eAn J, Ying L, Zhu Y. Why resampling outperforms reweighting for correcting sampling bias with stochastic gradients [Internet]. arXiv.org. 2020 [cited 2025 Jun 11]. Available from: https://arxiv.org/abs/2009.13447.\u003c/li\u003e\n\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":false,"hideJournal":false,"highlight":"","institution":"","isAcceptedByJournal":false,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"bmc-medical-informatics-and-decision-making","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"midm","sideBox":"Learn more about [BMC Medical Informatics and Decision Making](http://bmcmedinformdecismak.biomedcentral.com/)","snPcode":"","submissionUrl":"https://www.editorialmanager.com/midm/default.aspx","title":"BMC Medical Informatics and Decision Making","twitterHandle":"BMC_series","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"em","reportingPortfolio":"BMC Series","inReviewEnabled":true,"inReviewRevisionsEnabled":true},"keywords":"Skin tone bias, Bruise detection, AI fairness, Demographic parity, Equal Opportunity, Object detection","lastPublishedDoi":"10.21203/rs.3.rs-6875703/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-6875703/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003ch2\u003eBackground\u003c/h2\u003e \u003cp\u003eAI-based bruise detection tools have shown promise in injury detection but often underperform on darker skin tones. Most fairness evaluations in dermatology AI rely on a single method of skin tone categorization. In biased models, the choice of skin tone scale plays a critical role, not only in grouping individuals, but also in making disparities visible. A well-designed scale should reveal, rather than mask, differences in model performance across skin tones. This study evaluated how different skin tone classification methods affect the accuracy, average precision, and fairness measures in bruise detection.\u003c/p\u003e\u003ch2\u003eMethods\u003c/h2\u003e \u003cp\u003eUsing a dataset of 11,766 bruise and non-bruise images collected under white and alternate light sources, six skin tone labeling methods were applied: Individual Typology Angle (ITA), Fitzpatrick scale (full and binarized), Monk Skin Tone (MST), and two clustering approaches (chromatic and luminance-based). Chromatic clusters were created using the raw L* (lightness), a* (red-green), and b* (yellow-blue) values, whereas luminance clustering used the L* values only. A YOLOv5 model was re-trained for bruise detection, and model accuracy, average precision, and fairness were assessed on the test set using Demographic Parity Difference (DPD) and Equal Opportunity Difference (EOD) metrics across these six skin tone groupings.\u003c/p\u003e\u003ch2\u003eResults\u003c/h2\u003e \u003cp\u003ePerformance and fairness varied widely depending on the skin tone classification method used. Granular scales such as the Monk Skin Tone (MST) scale and the chromatic cluster-based scale revealed greater variability in both performance and fairness across the skin tone groups, especially in the mid-to-dark range. In contrast, simpler and broader scales such as luminance-based clustering and binary Fitzpatrick showed more stable trends, but they may have hidden important differences between skin tones.\u003c/p\u003e\u003ch2\u003eConclusions\u003c/h2\u003e \u003cp\u003eSkin tone classification plays a key role in how both performance and fairness are evaluated in bruise detection models. Granular skin tone scales such as MST and chromatic clustering may not show the highest performance, but reveal disparities more clearly, whereas broader scales may mask them despite performing well. Addressing these biases requires careful selection of skin tone grouping methods for evaluation.\u003c/p\u003e","manuscriptTitle":"The Effect of Skin Tone Classification on Bias in Bruise Detection","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2025-06-30 08:06:38","doi":"10.21203/rs.3.rs-6875703/v1","editorialEvents":[{"type":"communityComments","content":0},{"type":"decision","content":"Revision requested","date":"2026-04-10T14:29:28+00:00","index":"","fulltext":""},{"type":"editorInvitedReview","content":"","date":"2025-11-25T03:21:30+00:00","index":"hide","fulltext":""},{"type":"editorInvitedReview","content":"","date":"2025-11-23T22:52:23+00:00","index":"hide","fulltext":""},{"type":"editorInvitedReview","content":"","date":"2025-11-23T19:37:41+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"272161512461809360232028068864080684665","date":"2025-11-20T10:57:58+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"60491064608798987489927400261907691836","date":"2025-11-19T00:17:57+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"158144409122768381264429677254373269060","date":"2025-11-13T12:52:25+00:00","index":"hide","fulltext":""},{"type":"editorInvited","content":"","date":"2025-10-17T15:12:01+00:00","index":"","fulltext":""},{"type":"reviewersInvited","content":"","date":"2025-06-23T14:08:55+00:00","index":"","fulltext":""},{"type":"editorAssigned","content":"","date":"2025-06-23T06:24:21+00:00","index":"","fulltext":""},{"type":"checksComplete","content":"","date":"2025-06-20T19:54:42+00:00","index":"","fulltext":""},{"type":"submitted","content":"BMC Medical Informatics and Decision Making","date":"2025-06-20T19:51:30+00:00","index":"","fulltext":""}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"bmc-medical-informatics-and-decision-making","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"midm","sideBox":"Learn more about [BMC Medical Informatics and Decision Making](http://bmcmedinformdecismak.biomedcentral.com/)","snPcode":"","submissionUrl":"https://www.editorialmanager.com/midm/default.aspx","title":"BMC Medical Informatics and Decision Making","twitterHandle":"BMC_series","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"em","reportingPortfolio":"BMC Series","inReviewEnabled":true,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"c07f378c-feb9-4864-8faf-0e7c07a8d61e","owner":[],"postedDate":"June 30th, 2025","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"under-review","subjectAreas":[],"tags":[],"updatedAt":"2026-04-30T08:38:28+00:00","versionOfRecord":[],"versionCreatedAt":"2025-06-30 08:06:38","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-6875703","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-6875703","identity":"rs-6875703","version":["v1"]},"buildId":"8U1c8b4HqxoKbykW_rLl7","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}