Automated Measurement of Depigmentation Extent with a New AI Tool Applied to the Example of Vitiligo

Abstract

Introduction

We have developed a digital algorithm to assess skin pigmentation, specifically an artificial intelligence-based image analysis tool that segments photographed lesions and then scores them by Facial Vitiligo Area Scoring Index (F-VASI), in place of trained site investigators. Vitiligo, the disease used in this exemplary demonstration of the algorithm, is a chronic, acquired, immune-mediated depigmentation disease characterized by white macules and/or patches of skin. The F-VASI is a clinician-reported outcome that relies on manual assessment of affected body surface area (BSA) and level of depigmentation and is subject to inter- and intra-rater variability. Here, we present automated medical image segmentation of vitiligo lesions and digitization of validated scores, including F-VASI, BSA, and percentage of depigmentation (%Depigmentation).

Methods

Our convolutional neural network (“UNet”) uses encoder-decoder architecture to process photographic images and quantify areas of skin affected by vitiligo.

Results

We trained and validated our model using cross-polarized participant photos from clinical trials, achieving 81% accuracy when predicting vitiligo lesions in new photos. In addition, we created an algorithm to digitize F-VASI assessment using estimates of BSA and %Depigmentation that were calculated using the predicted lesions in the photos. We were able to achieve an interclass correlation coefficient of 0.91 when comparing our digital F-VASI score to the manually estimated F-VASI score.

Conclusion

We found that using a UNet to segment vitiligo lesions can allow us to digitize clinically meaningful measures for vitiligo.

Trial Registration

The phase 2b study: NCT03715829.

Supplementary Information

The online version contains supplementary material available at 10.1007/s13555-026-01736-8.

Plain Language Summary

Vitiligo is a skin condition in which the skin loses its pigment (depigmentation) and white patches appear. Doctors usually measure how much of the skin is affected by examining the extent of the white patches on the patient using a measure called the Vitiligo Area Scoring Index (VASI), and its face-specific sub-score, F-VASI. Although widely used, VASI requires a trained dermatologist to judge the size and degree of depigmentation of the patches, which takes time. Plus, the assessment results may vary from one person to another. We have developed a digital algorithm that uses artificial intelligence (AI) to automatically identify and measure vitiligo patches in patient photographs. We trained the model using photographs collected during clinical trials, which had been marked by an expert during the trial to show exactly where the vitiligo patches were located. Once trained, the AI was able to correctly outline vitiligo patches on new photographs and then calculate a digital version of the F-VASI score. We compared the AI-generated results with assessments made by trained dermatologists. The digital assessments closely matched (with about 90% agreement) the human assessments, showing strong agreement for both the size of the affected area and the level of depigmentation. This work shows that AI can provide fast, objective, and consistent measurements of vitiligo.

Supplementary Information

The online version contains supplementary material available at 10.1007/s13555-026-01736-8.

Key Summary Points

Why carry out this study?

The Facial Vitiligo Area Scoring Index (F-VASI) is a standardized, clinician-reported outcome that relies on manual assessment of affected body surface area and level of depigmentation and is subject to inter- and intra-rater variability.

We have developed an AI (artificial intelligence)-based image analysis algorithm to segment photographed lesions and then score by F-VASI in place of trained site investigators.

Our convolutional neural network (“UNet”) uses encoder-decoder architecture to process photographic images and quantify areas of skin affected by vitiligo.

What was learned from the study?

The correlation coefficient between computationally calculated and physician-assessed extent of facial vitiligo demonstrated great agreement (interclass correlation coefficient 0.81).

We show that using a UNet to segment vitiligo lesions can allow us to digitize clinically meaningful measures for vitiligo and other skin diseases.

Introduction

Nonsegmental vitiligo is a chronic acquired immune-mediated depigmentation disease characterized by well-defined depigmented white macules and/or patches of skin due to melanocyte destruction, with or without immune-mediated comorbidities [1]. Vitiligo can be limited in extent or involve multiple body sites and a large body surface area (BSA) [2]. To evaluate vitiligo extent, various clinician-reported outcomes such as the Vitiligo Area Scoring Index (VASI), and its face-specific sub-score, F-VASI, are used. These clinician-reported outcomes require healthcare experts to manually assess affected body surface area and level of depigmentation and to calculate visual differentiations between endpoints in order to gain an understanding of disease progress. Manual assessment is subjective [3] and time-consuming, and implementation across trials shows heterogeneity in methodology [4], suggesting risk of observer error. Although formal timings of F-VASI scoring are rarely reported, the multiple steps required certainly render manual scoring labor-intensive. The absence of an automated and objective tool specifically designed to calculate F-VASI score is a compelling area for improvement in the practice of dermatology.

Currently, across medicine in general, there are many competing models that perform medical image segmentation [5]. Many of these models are variations of the encoder-decoder UNet-like architecture and fully convolutional networks [5]. Briefly, an encoder compresses an image and identifies important visual features and then a decoder re-expands the representation, applying a meaningful label such as vitiligo versus non-vitiligo skin to each pixel. The model then outputs a map showing which parts of a photograph correspond to different regions. This class of models perform superbly in processing and segmenting images, even with limited data [5]. Diseases with symptoms that can be visually captured are prime candidates for creation and fine-tuning of artificial intelligence (AI) models that can automatically detect and quantify their extent.

Current methods for manually measuring and quantifying the extent of vitiligo are antiquated and time-consuming (Table 1). Patients must be diagnosed visually and have their lesions quantified using measures including body surface area (BSA) and %Depigmentation. VASI comprises BSA and %Depigmentation, which both are based on subjective judgments. BSA is measured by hand-unit, finger-unit, and/or finger-tip unit, and %Depigmentation is measured by visually comparing the percentage of depigmentation in the measured lesions or areas of a participant’s diseased skin and their normal skin. Even when performing these measurements in a scrutinized, regulated clinical trial setting, there is variation between and within clinicians’ measurements [4, 6, 7]. As the F-VASI score is complex to calculate and physicians must be trained to make a good calculation with it [8], it is important to simplify this process. There exists a need to create a digital measurement of these vitiligo endpoints that can be at least as accurate as a clinician’s assessment. The application of AI to this challenge could form a supportive tool to help clinicians’ disease assessment.

Table 1.

Assessment techniques

Technique	Rater	Output
Fully manual techniques
%BSA estimation (Rule of Nines or Palm Rule)	Clinician	Visual estimation of extent, frequently used as basis for other scores
Vitiligo Area Scoring Index (VASI) [19]	Clinician	Estimation of extent and depigmentation (%BSA multiplied by clinician-rated %Depigmentation)
Vitiligo European Task Force assessment (VETF) [20]	Clinician	Extent (by %BSA, above), staging (degree of pigment loss and hair involvement), spreading (activity based on clinician assessment using Wood’s lamp)
Vitiligo Extent Score (VES) [21] and Self-Assessment VES (SA-VES) [22]	Clinician or patient	Uses pictorial templates of body regions with different coverage patterns; clinician (VES) or patient (SA-VES) selects the picture that best matches the condition
Vitiligo Disease Activity Score (VIDA) [23]	Patient	Six-point scale based on patient recall of new lesions or enlargement over time
Vitiligo Signs of Activity Score (VSAS) [24]	Clinician	Visual assessment of clinical signs converted into a score
Digitally supported techniques
Online calculators [25]	Clinician or patient	Web platforms that guide clinician or patient through VES/VASI process; can sum scores, store/export results
VASI from digital photographs	Clinician	Standardized digital photos are taken, on-screen grids and rulers are used to estimate hand units and depigmentation, rather than live exam
Fully AI/deep learning-based tools
Convolutional neural networks [15, 16, 26]	AI	Machine-learning models trained on clinical photographs to automatically identify and quantify severity

AI artificial intelligence, %BSA percentage of body surface area

Dermatologic photography poses technical challenges. Consistent illumination is required, as shadows, reflection, and variable color temperature can change the apparent hue and contrast of lesions [9]. Ensuring sufficient resolution for diagnostic detail while maintaining manageable file sizes and secure storage adds further technical constraints [10]. Reproducible patient positioning and camera geometry are also essential for consistent interpretation across visits, operators, and locations [11]. As such, we used high-quality, clinical-trial photographs and clinical measurements of participants’ vitiligo extent to train our deep learning model (U-Net architecture) to automatically quantify depigmented skin regions. We evaluated its agreement with dermatologist assessments across a range of face/upper body sites and disease severities.

Related Work

Recently, there have been several publications that attempt to use AI to evaluate and delineate vitiligo lesions and to determine their severity [12–16]. One study used two deep convolutional neural networks to first achieve classification/localization and then segmentation of vitiligo lesions, respectively [15]. The study included 2720 images of vitiligo and used its hybrid AI model achieved an accuracy of 87.8% [15] Another recent study compared AI-based digital assessment of hypopigmented dermatoses to human-derived assessment. Using 11,483 images from 4744 patients, the study showed that AI performed as well as three experienced dermatologists [16].

The present study aimed to develop a digital algorithm that uses AI to automatically identify and measure, by F-VASI, facial vitiligo patches in patient photographs. We trained the model using expert-marked clinical trial photographs and evaluated its performance by comparing the model’s estimation of F-VASI with that of human experts on a previously unseen set of patient photographs.

Methods

Study Design and Data Sources

Photographs of patients with vitiligo in a randomized phase 2b study sponsored by Pfizer [17] were used for a proof-of-concept-phase segmentation of facial vitiligo lesions (outlining/tracing their borders from normally pigmented skin) performed manually using AnyLabeling software [18], as guided by dermatologists. We implemented a convolutional neural network (U-Net) to delineate depigmented areas in similar images. Because the trial offered an available set of patients with a confirmed diagnosis and high-standard images, we used these vitiligo clinical trial data for our model training and validation. Participant photographs from the completed phase 2b B7981019 trial (NCT03715829) were captured via Canfield Scientific’s high-resolution total-body photography technology. The images consist of photographs of the upper chest and head/facial areas (forward facing, left and right laterals) to provide a set of facial photos for each participant. Site raters evaluated participants and documented their vitiligo lesions by VASI score.

The participants’ photographs were then manually segmented using AnyLabeling software by one research associate who was trained by dermatologists. The initial annotations were further reviewed, adjusted, and confirmed by dermatologists. The labeled contours were converted to image masks—colored overlays that highlight affected areas—for model training. Images from 80% of participants were randomly selected for training the model, and images from the remaining 20% were reserved for validation. To avoid the possibility of an imbalance in skin tones between the training and validation datasets, the training and validation participants were stratified based on the Fitzpatrick skin phototype at baseline (Table 2).

Table 2.

Fitzpatrick skin phenotypes

N = 106	Train (n1 = 85)	Validation (n2 = 21)
Fitzpatrick skin phototype (I and II)	11 (12.9%)	3 (14.3%)
Fitzpatrick skin phototype (III and IV)	58 (68.2%)	14 (66.7%)
NA (missing)	16 (18.8%)	4 (19.0%)

Ethical Approval

The phase 2b study (NCT03715829) from which patient photographs were used was conducted at 80 sites in Australia, Belgium, Canada, Germany, Italy, Japan, South Korea, Spain, Taiwan, and the USA from November 2018 to February 2021. The final protocol of that study, any amendments, and informed consent documentation were approved by the institutional review board/independent ethics committee at each study center (Supplementary Table 1). The study was conducted in compliance with the Declaration of Helsinki and all International Conference on Harmonization Good Clinical Practice guidelines. All patients provided written informed consent to take part in that trial and consented for their images to be used and published.

Model Architecture

A U-Net model with a pretrained backbone was built and trained to identify vitiligo lesion areas. The convolutional neural network architecture Inception-Resnet-V2 was selected as encoder, and pretrained weights from the ImageNet dataset were used as initial weight. The decoder was custom coded and freshly trained on the training dataset. The original images were sliced into 512 × 512-pixel patch images and rescaled to 128 × 128 pixels for model input. Interclass correlation coefficient, a measure of agreement between the model’s predicted maps and those marked by human experts, was used as the parameter to evaluate overall model performance. The model was trained with a batch size of 30 image patches and 30 epochs (passes through the full data set) on a Nvidia A10g GPU with 24-GB memory.

Creation of Digital Endpoints

Once the U-Net model was fully trained and validated, we used the output in the creation of digital endpoints (F-VASI, percent BSA, and depigmentation level). The percentage BSA (%BSA) impacted by vitiligo was calculated by dividing the total facial lesion area by total facial area in an image. To digitize change in skin color tones, the RGB images were converted to the CIELab scale, a device-independent scale describing lightness and color. The IWA_Newtone was calculated for each pixel in a lesion and were averaged across all pixels in the lesion to represent the lesion level color tone, where the darker skin tone corresponds to a lower IWA_Newtone value. The skin color tone of unaffected skin was calculated by taking the median of the IWA_Newtone from the pixels in neighboring unaffected regions. The depigmentation level was derived by taking the difference in lesion skin tone and normal skin tone (IWA_diff). The final digital VASI score was defined as the product of %BSA and %Depigmentation.

Validation of Digital Endpoints Against Local-Read Endpoints

The model-generated digital %BSA and F-VASI were validated against their corresponding local-read counterparts as the golden standard. We used the Pearson correlation coefficient, interclass correlation coefficient, R-squared, regression slope, regression intercept, and mean-squared error to compare the digital endpoints to the local-read endpoints. A correlation coefficient of 1.00, R-squared of 100%, regression slope of 1.00, regression intercept of 0.00, and low mean-squared error would indicate that the digital endpoints are identical to those assessed by an expert human and could be used as a substitute for local-read endpoints.

Results

Automated Facial Vitiligo Lesion Segmentation with Deep Learning

A total of 7192 image patches, each 512 × 512 pixels, and the corresponding annotation masks were derived from facial images of 85 participants for model training. For model validation, 390 image patches and pairing masks were constructed from the 21 participants who were independent of the training set. Our UNet model achieved nearly 100% interclass correlation coefficient in the training dataset, and the optimal model reached a coefficient of 81% in the validation dataset. Figure 1 shows three patch-level images, annotated masks, and predicted masks by UNet model.

Fig. 1.

Masks

Concordance between Computational and Physician Assessment of Extent of Facial Vitiligo

The segmentation of facial vitiligo lesions computationally allows the precise calculation of the extent of depigmentation. The extent of facial vitiligo was derived by dividing the total number of pixels in UNet-estimated vitiligo lesions by the total number of pixels in the front facial region, a number ranging from 0% (no depigmentation) to 100% (complete depigmentation). This number was further multiplied by a constant of 4% to be converted to percent BSA (fBSA) and compared with physician-assessed extent of facial vitiligo by central reading labs. Among the 21 validation participants, the interclass correlation coefficient between computationally calculated and physician-assessed extent of facial vitiligo was 0.81 (Fig. 2).

Fig. 2.

Correlation between computationally calculated and physician-assessed extent of facial vitiligo. AI artificial intelligence, fBSA extent of facial vitiligo by percentage of body surface area (%BSA), ICC interclass correlation coefficient

Digital VASI Agreement with Physician-Assessed VASI

The difference (IWA_diff) in the median value of IWA_newtone in the vitiligo area and normal skin area was used as the change in skin paleness and served as a surrogate measure of level of depigmentation. The final digital VASI for facial vitiligo was calculated by taking the product of IWA_diff and percent BSA and was subsequently compared with physician-assessed VASI. The digital VASI and physician-assessed F-VASI showed an interclass correlation coefficient of 0.91, supporting that the digital F-VASI can reliably measure facial vitiligo disease extent (Fig. 3).

Fig. 3.

Correlation between computationally calculated and physician-assessed VASI. ICC interclass correlation coefficient, VASI Vitiligo Area Scoring Index

Discussion

Manual assessment of the extent of vitiligo lesions is slow, complex, and carries the risk of inter-rater variability. The successful application of AI-based deep learning model to the automatic segmentation of such lesions in skin photographs would, by contrast, offer rapid, objective, and reproducible measurements of disease extent, reducing diagnostic uncertainty and enhancing the consistency of disease-severity monitoring across clinical settings. Among the assessments from the 21 validation participants in this study, the correlation coefficient between computationally calculated and physician-assessed extent of facial vitiligo demonstrated great agreement (interclass correlation coefficient 0.81), indicating that the computationally calculated extent of facial vitiligo could be used to monitor treatment effects in clinical settings. Furthermore, final digital VASI for facial vitiligo showed an interclass correlation coefficient of 0.91 with physician-assessed F-VASI, indicating that our digital F-VASI algorithm can reliably measure vitiligo disease severity according to F-VASI score.

This algorithm outputs a F-VASI score quickly, compared to the several minutes required by a dermatologist, not including the time required to initially learn how to assess VASI. The automated computation of standardized metrics offers objective, generalizable, and consistent indices for assessing the status and progress of diseases during treatment. This may ensure more efficient and precise analysis of digital imaging data. Automated digital F-VASI could also support clinical-trial efficiency by enabling centralized, reproducible assessments that reduce human variability, accelerate endpoint processing, and facilitate rapid interim analyses.

This proof-of-concept work was conducted using high-quality clinical trial photographs, and the digital VASI algorithm’s performance in real-world imaging conditions (e.g., non-standardized variable lighting, smartphone cameras) remains to be evaluated. In addition, the validation dataset comprised images from only 21 participants; larger datasets will further strengthen robustness. Future external validation of our model would be welcomed to confirm the generalizability of our findings. This present study used only face and upper-body photographs and so was limited to the digital estimation of F-VASI, which represents a limitation compared to the whole-body VASI score.

The application of AI-based image analysis to the classification and segmentation of vitiligo has been the subject of considerable contemporary research, recently well-summarized by Parikh and colleagues [12]. They report classification accuracies ranging from 66% to 97% for deep learning systems described in recent literature, and segmentation accuracies between 93% and 97% [12]. To our knowledge, the use of AI technology to provide F-VASI assessment, as described in our present study, is more novel. Digitized F-VASI scores calculated by our AI-based image analysis algorithm were as accurate and reproducible as those of site investigators and treating dermatologists. Wider employment of such tools could lead to more rapid diagnosis and enhanced quality of life for dermatology patients by lessening the burden of assessment. Such benefits could be further realized through combination with better photography techniques or UV imaging. Although developed and validated here for nonsegmental vitiligo, the underlying principle of AI-based lesion segmentation combined with quantitative colorimetric analysis may be extensible to other dermatologic conditions characterized by visible morphological or pigmentary alterations, including post-inflammatory dyspigmentation, acne scarring, and inflammatory dermatoses.

Conclusions

Overall, this work demonstrates that AI-enabled lesion segmentation and digital F-VASI quantification can serve as reliable, objective tools that complement clinician evaluation and may ultimately contribute to a more standardized and patient-centered approach to vitiligo care.

Supplementary Information

Below is the link to the electronic supplementary material.

Author Contributions

Yalei Chen, Tatjana Lukic, All-Shine Chen, Roni Adiri, Helen Tran, Gregor Schaefer, Pranab Ghosh, Subha Madhavan, Koshika Soma and Margaret Gamalo all contributed to the study conception and design. The first draft of the manuscript was written by Yalei Chen and all authors commented on subsequent versions of the manuscript. All authors read and approved the final manuscript.

Funding

This study was funded by Pfizer Inc. The journal’s Rapid Service Fee was funded by Pfizer Inc.

Data Availability

The datasets generated during and/or analyzed during the current study are not publicly available due to patient privacy concerns, as the photo images include participants’ facial areas.

Declarations

Conflicts of Interest

Yalei Chen, Tatjana Lukic, All-Shine Chen, Roni Adiri, Helen Tran, Gregor Schaefer, Pranab Ghosh, Subha Madhavan, Koshika Soma and Margaret Gamalo are all employees of, and may hold stock or stock options in, Pfizer Inc.

Ethical Approval

The phase 2b study (NCT03715829) from which patient photographs were used was conducted at 80 sites in Australia, Belgium, Canada, Germany, Italy, Japan, South Korea, Spain, Taiwan, and the USA from November 2018 to February 2021. The final protocol of that study, any amendments, and informed consent documentation were approved by the institutional review board/independent ethics committee at each study center (Supplementary Table 1). The study was conducted in compliance with the Declaration of Helsinki and all International Conference on Harmonization Good Clinical Practice guidelines. All patients provided written informed consent to take part in that trial and consented for their images to be used and published.