Organizations: School of Digital and Physical Sciences at the University of Hull, Cottingham Road, Hull, HU6 7RX, United Kingdom · Faculty of Electrical Engineering at the University of Sarajevo, Zmaja od Bosne bb, Sarajevo, 71000, Bosnia and Herzegovina
Skin cancer, particularly melanoma, remains a major cause of morbidity and mortality, making early detection critical. AI-driven dermatology systems often rely on skin lesion segmentation as a preprocessing step to delineate the lesion from surrounding skin and support downstream analysis. While fairness concerns regarding skin tone have been widely studied for lesion classification, the influence of skin tone on the segmentation stage remains under-quantified and is frequently assessed using coarse, discrete skin tone categories. In this work, we evaluate three strong segmentation architectures (UNet, DeepLabV3 with a ResNet50 backbone, and DINOv2) on two public dermoscopic datasets (HAM10000 and ISIC2017) and introduce a continuous pigment or contrast analysis that treats pixel-wise ITA values as distributions. Using Wasserstein distances between within-image distributions for skin-only, lesion-only, and whole-image regions, we quantify lesion skin contrast and relate it to segmentation performance across multiple metrics. Within the range represented in these datasets, global skin tone metrics (Fitzpatrick grouping or mean ITA) show weak association with segmentation quality. In contrast, low lesion-skin contrast is consistently associated with larger segmentation errors in models, indicating that boundary ambiguity and low contrast are key drivers of failure. These findings suggest that fairness improvements in dermoscopic segmentation should prioritize robust handling of low-contrast lesions, and the distribution-based pigment measures provide a more informative audit signal than discrete skin-tone categories.
Figures & tables
Fig. 1 : Research Analysis Overview: After semantic segmentation, the skin tone is measured by green colored six comparison patterns in (a). From left to right, the original image, the ground truth annotation, and the image in which the annotations have been used to remove the hair and extract only the skin-colored pixels. The fourth image shows annotations used to measure only the lesion’s color, and then the base color image. The image (b) reproduced from [ 36 ] illustrates the ITA that identified pixel-wise color links to skin. (c) is a detailed description of six comparison patterns.
Acronym
Full name
Evaluation Criteria
IoU
Intersection over Union
Higher values indicate better performance
DC
Dice Coefficient
ST
Sensitivity (True Positive Rate)
SP
Specificity
PA
Pixel Accuracy
AUC
Area under the Curve
TABLE I : Segmentation Performance Evaluation Metrics
DataSet
Model
Epoch
IoU
DC
ST
SP
PA
AUC
CK
FNR
FPR
HAM
UNET
24
0.88
0.93
0.96
0.97
0.97
0.96
0.90
0.04
0.03
DeepLab
19
0.88
0.93
0.96
0.97
0.97
0.96
0.90
0.04
0.03
DinoV2
19
0.89
0.94
0.96
0.97
0.97
0.97
0.91
0.04
0.03
ISIC
UNET
43
0.73
0.83
0.83
0.97
0.93
0.90
0.78
0.17
0.03
DeepLab
49
0.73
0.83
0.83
0.97
0.93
0.90
0.77
0.17
0.03
DinoV2
42
0.75
0.84
0.84
0.97
0.94
0.91
0.79
0.16
0.03
TABLE II: General Performance of each Model: Bold values indicate metrics with high performance, defined as mean scores above 95% for metrics where higher is better and below 5% for metrics where lower is better.
Fig. 2 : General Performance of each Class with HAM: The figure on the left shows metrics for evaluating prediction accuracy, and the higher the accuracy, the higher the value. In contrast, the figure on the right shows the error rate, and the lower the score, the better the model’s performance.
Fig. 3 : Skin Color Distribution Skin or Lesion Color Each Class: The eight figures in the upper row show the distribution of each class based on the evaluation of the colors of each image in the HAM. Each figure follows the measure methods for color patterns in Table 1 and shows trends in the distribution of color patterns in the dataset.
Fig. 4 : Correlation Ranges between Skin Color Measures and Segmentation Performance Across Datasets and Models: This figure presents Spearman rank correlations between each segmentation performance metric and skin-color distance measures across different comparison patterns. The light blue band visualizes the minimum–maximum range of correlations across metrics.
UNET
DeepLabV3
DINOv2
Metric
Corr
L
H
Corr
L
H
Corr
L
H
Pattern 4
IoU
0.21
0.16
0.25
0.22
0.18
0.27
0.29
0.24
0.33
DC
0.21
0.17
0.25
0.22
0.18
0.26
0.29
0.25
0.33
ST
-0.27
-0.31
-0.22
-0.19
-0.23
-0.14
-0.06
-0.10
-0.01
SP
-0.52
-0.55
-0.48
-0.54
-0.57
-0.50
-0.58
-0.61
-0.54
PA
-0.57
-0.60
-0.53
-0.59
-0.62
-0.56
-0.59
-0.62
-0.56
TABLE III : Confidence intervals of Spearman correlation coefficients for Patterns 4-6
Fig. 5 : DINOv2 Absolute Spearman correlations between skin tone and segmentation performance metrics across seven disease classes
Fig. 6 : Sample images measured for skin tone nuances using Pattern 6: The difference between skin pixels and lesion areas can be clearly captured and quantified.
UNET
DeepLabV3
DINOv2
Metric
Corr
L
H
Corr
L
H
Corr
L
H
MEL
Pattern 4
SP
-0.59
-0.67
-0.50
-0.64
-0.71
-0.55
-0.66
-0.73
-0.57
PA
-0.52
-0.61
-0.40
-0.58
-0.67
-0.48
-0.58
-0.66
-0.48
AUC
-0.48
-0.59
-0.37
-0.51
-0.61
-0.40
-0.49
-0.58
-0.39
FPR
0.59
0.50
0.68
0.64
0.56
0.71
0.66
0.59
0.73
Pattern 5
ST
0.57
0.47
0.66
0.42
0.29
0.53
0.30
0.16
0.42
TABLE IV : Confidence intervals of moderate-to-strong Spearman correlation coefficients for malignant skin lesions (Pattern 4-6)
Kuniko Paxton is a PhD candidate in Digital and Physical Sciences at the University of Hull, and the funding institution is DAIM. The research interests are Pruning Learning, Fairness in AI, Explainability, and Distributed Learning. Contact her at [email protected]
Table 11
Koorosh Aslansefat is an assistant professor of computer science at the University of Hull, HU6 7RX Hull, U.K., Co-Leader of the Dependable Intelligent Systems research centre. His research interests span artificial intelligence safety, Markov modelling, and real-time dependability analysis. Aslansefat received his PhD in computer science from the University of Hull. He is a Senior Member of IEEE. Contact him at [email protected]
Table 12
Dhavalkumar Thakker is a Professor of Artificial Intelligence (AI) and the Internet of Things (IoT) at the University of Hull, where he leads a group focused on Responsible Artificial Intelligence. His research emphasizes AI Explainability, AI Safety, and Fairness. With nearly two decades of experience, Dhavalkumar has been at the forefront of innovative solutions through funded projects. His interdisciplinary research spans Generative AI and the applications of Edge computing alongside IoT technologies. He has a track record in leveraging AI for Social Good, notably in Smart Cities, Digital Health, and the Circular Economy. Contact him at [email protected]
Table 13
Yiannis Papadopoulos is Professor of Computer Science and Leader of the Dependable Intelligent Systems (DEIS) Research Group at the University of Hull in the U.K. For over 30 years, Papadopoulos and his research group have pioneered cutting-edge model-based, bio-inspired and statistical technologies for the analysis and design of dependable engineering systems, with a recent intense focus and contributions towards achieving trustworthy, safe AI. Many of the software tools that they have crafted, including HiP-HOPS and EAST-ADL, have become commercial and are used in transport and other industries. Contact him at [email protected]
Table 14
Amila Akagic is professor who received the B.Sc. (2006) and M.Sc. (2009) degrees in electrical engineering, with a focus on computer science and informatics, from the University of Sarajevo. She received the Ph.D. degree from Keio University, Japan, in 2013. She is currently a Full Professor in the Department of Computing and Informatics at the University of Sarajevo. Her research interests include artificial intelligence, machine learning, deep learning, computer vision, image processing, and digital signal processing.
Table 15
Medina Kapo received her B.Sc. degree in 2021 and M.Sc. degree in 2023 in Computer Science and Informatics from the Faculty of Electrical Engineering, University of Sarajevo. She is currently employed as a Teaching Assistant at the Faculty of Electrical Engineering, University of Sarajevo, where she is also pursuing her Ph.D. studies in Computer Science and Informatics. Her research interests include computer vision, deep learning, machine learning, the application of artificial intelligence models and methods in medicine, and database systems.
Department of Computer Science, Quaid-e-Awam University of Engineering, Sciences & Technology · Department of Artificial Intelligence, Quaid-e-Awam University of Engineering, Sciences & Technology, Pakistan · Department of Computer Science, Sindh Madressatul Islam University, City Campus, Karachi +1
Department of Computer Science, Indian Institute of Science Education and Research Berhampur, India · Department of Mathematics, Indian Institute of Technology Jodhpur, India · School of Artificial Intelligence and Data Science, Indian Institute of Technology Jodhpur, India