The Potency of Artificial Intelligence in Diagnosing and Evaluating the Severity of Vitiligo: A Systematic Review and Meta-Analysis

Abstract

Neural networks (NNs), a subset of artificial intelligence (AI), have the ability to assess the severity of vitiligo and improve diagnosis accuracy. In contrast to human physicians’ diagnosis, which is primarily based on subjective assessment, standardised and objective diagnostic methods are needed to provide an accurate diagnosis and grade the severity of vitiligo. The aim of this study is to analyse the sensitivity and specificity of AI in vitiligo diagnosis and severity assessment. The PRISMA 2020 protocol was followed to screen literature in several databases: Proquest, PubMed, Taylor and Francis, SAGE, JSTOR, and ScienceDirect. Human studies that investigated the potential of AI in vitiligo diagnosis were included in vitiligo diagnosis were included. Statistical analysis was carried out using RevMan 5.4 and RStudio, and risk of bias tools were used to assess study bias. Certainty of evidence was analysed using GRADE. The systematic review consisted of 15 studies, and seven studies were statistically analysed. The most studied NNs were Inception V3 (n = 2), visual geometry group (n = 2), and ResNet (n = 2). The overall sensitivity ranged from 72.5% to 99.7%, with a pooled sensitivity of 95.5%. Specificity ranged from 72% to 99.9%, with a pooled specificity of 97%. Among the evaluated models, Inception V3 showed the highest diagnostic performance. The summary receiver operating characteristics curve (SROC) from multiple analyses showed an area under the curve of 0.96. AI showed potential in diagnosing and evaluating the severity of vitiligo.

Introduction

Vitiligo is a chronic acquired depigmenting skin disorder that can be widespread or localised. As the most prevalent acquired skin depigmenting condition, vitiligo significantly lowers quality of life. The prevalence of vitiligo ranges from 0.2% to 1.8% globally.[1,2,3] Vitiligo’s chronicity and the absence of effective therapies can seriously impair patients’ quality of life and lead to significant psychological stress. Observable lesions can cause despair, anxiety, and embarrassment, particularly if they are on areas that are easily seen, such as the hands, neck, and face.[4] Dermatologists’ diagnosis largely depends on subjective visual assessment and clinical experience in this diagnostic procedure. For precise and timely detection, particularly in individuals with an unusual presentation, examination by experienced clinicians using Wood’s lamp is essential. Point-counting techniques and the Vitiligo Area Scoring Index (VASI) are examples of semi-objective approaches. A more objective method to diagnose and score the grade of vitiligo is needed.[5,6,7]

Artificial intelligence (AI) deep learning (DL) has been utilised in clinical settings and medical image analysis in recent years. DL, in contrast to earlier machine learning techniques, makes use of intricate techniques to provide data prediction, classification, large-scale pattern recognition, and decision-making assistance. In this case, multilayer artificial neural networks (ANNs) are utilised.[4,8] In contrast to human physicians’ diagnosis, which is primarily based on subjective assessment and is not always repeatable, standardised, and objective, DL techniques were considered as a viable support system capable of providing accurate skin lesion diagnosis and grading the severity of the disease. One of the most often used DL models is the convolutional neural network. They have shown expert-level performance in the classification of skin conditions on medical photographs in recent years.[5,9]

In recent years, a few studies have been reported on how AI could enhance the diagnosis of vitiligo. A computer-based automatic scoring technique would be more objective, simpler to use, save time, and provide a more accurate diagnosis and grading compared to human analysis.[10] The primary goal of treatments is to first halt the progression and severity of vitiligo and then repigment the afflicted skin area to restore it. By analysing the skin and determining variables such as colour changes and the size of the affected area, the efficacy of the treatment is assessed.[11]

A thorough evaluation and meta-analysis evaluating the potential of AI in diagnosing and grading the severity of vitiligo has not yet been published. This article aimed to analyse the efficacy of AI neural networks (NNs) in vitiligo diagnosis and severity grading. Early intervention can be initiated promptly to reduce disease progression and morbidity in patients with vitiligo.

Methods

Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) guidelines were followed in this study, as displayed in Figure 1. The registry number CRD420251048662 is associated with this systematic review’s PROSPERO registration.

Figure 1.

PRISMA flowchart 2020

Literature search

To determine the potential of AI in identifying and assessing the severity of vitiligo, we performed a thorough literature search across multiple databases: PubMed, Proquest, Taylor and Francis, SAGE, ScienceDirect, and JSTOR. The articles were retrieved with no time limitation (until 20 February 2025). The search terms were as follows: (“machine learning” OR “artificial intelligence” OR “neural network”) AND (“vitiligo”). Furthermore, we searched for other potential studies to provide as references.

Study selection

Literature search was conducted using inclusion and exclusion criteria. The inclusion criteria were as follows: (1) full-text article available in English, (2) investigated the potency of AI in diagnosing and detecting the severity of vitiligo, and (3) human studies. The exclusion criteria were as follows: (1) review articles, (2) letters to the editor and commentary, (3) repeated publications, and (4) lack of data.

Data extraction and synthesis

Four authors independently extracted additional information on the included studies. The following were the extracted data: (1) author and publication year, (2) country of origin, (3) subjects (number of vitiligo, number of non-vitiligo), (4) data sources, (5) methods (devices and NN family), and (6) outcomes (sensitivity, specificity, accuracy, area under the curve (AUC), precision, and F1-scores).

Risk of bias assessment

Using the Quality Assessment of Diagnostic Accuracy Studies-2 (QUADAS-2) risk of bias (RoB) tools, four independent reviewers evaluated the bias risk. The Grading of Recommendations Assessment, Development, and Evaluation (GRADE) was used to assess the degree of certainty in the evidence. The fifth author discussed and settled any issues.

Statistical analysis outcome measurement

The outcomes included an analysis and evaluation of the F1-score, precision, area under the curve (AUC), sensitivity, specificity, and accuracy. Every study was carried out in accordance with the ethical guidelines mentioned. RStudio and RevMan 5.4 were used for statistical analysis. Statistical analysis was performed using RevMan 5.4 and RStudio. A diagnostic meta-analysis was conducted by pooling sensitivity and specificity across eligible studies. A random-effects model was applied to account for between-study variability. Heterogeneity among studies was assessed using the I² statistic, with values >50% indicating substantial heterogeneity. Subgroup analyses were performed based on different neural network (NN) families (e.g., Inception V3, VGG, ResNet, and others). Summary receiver operating characteristic (SROC) curves were generated to evaluate the overall diagnostic performance.

Results

Database literature searches yielded a total of 943 articles. Following screening and the elimination of duplicates (n = 185), 15 studies were included in the systematic review, and seven studies were analysed in the meta-analysis. The majority of studies were determined to have a low risk of bias [Figures 2 and 3].

Figure 2.

Traffic plot of QUADAS-2

Figure 3.

Summary plot of QUADAS-2

The characteristics of the eligible studies are shown in Table 1. All studies included were published between 2009 and 2024. The studies were conducted in various countries: China (n = 5),[4,5,12,13,14] Italy (n = 2),[15,16] India (n = 2),[17,18] Thailand (n = 1),[19] Denmark (n = 1),[20] France (n = 1),[3] USA (n = 1),[21] Taiwan (n = 1),[22] and Ecuador (n = 1).[11] The study designs included quasi-experimental human studies (n = 15).[3,4,5,11,12,13,14,15,16,17,18,19,20,21,22] The sample size varied from 10–20 subjects to 30,000 cases of vitiligo.[3,4,5,11,12,13,14,15,16,17,18,19,20,21,22] Several NN families were observed: Inception V3 (n = 2),[14,17] visual geometry group (VGG) (n = 2),[5,14] ResNet (n = 2),[5,14] ANN (n = 2),[16,19] Swin Transformer network (n = 1),[12] XGBoost (n = 1),[13] Random Forest (RF) (n = 2),[13,17] Support Vector Machine (SVM) (n = 1),[13] Naïve Bayes (n = 1),[17] YOLO v3 architecture (n = 1),[4] UNet++ architecture (n = 1),[4] Aysa AI app (n = 1),[18] DenseNet (n = 1),[5] Xception (n = 1),[14] EfficientNet-B2 combined with ConvNet and ImageNet (n = 1),[20] and others.

Table 1.

Characteristics of eligible studies

Author (Year)	Country	Subjects	Sources	Methods		Outcomes						Severity	Artificial Intelligence
				Devices	Nn Family	Sensitivity	Specificity	Accuracy	AUC	Precision	F1-Score
Zhong et al. (2024)[12]	China	Vitiligo (n=2678)	Images dataset (Department of Dermatology at Xijing Hospital, Fourth Military Medical University)	Dermoscope	Swin Transformer networks	94.02%	93.5%	93.82%	0.94	N/A	N/A	This approach helps with decision-making when diagnosing skin disorders by efficiently highlighting particular sections within diagnostic points that are pertinent to each class.	This AI model could distinguish the lesion border, continuity area, and pigmentation contrast by analysing the image heat map.
		Non-vitiligo (n=1642)
Lu et al. (2024)[22]	Taiwan	Vitiligo (n=10)	Individuals with facial vitiligo who received laser ablation and suction blister epidermal grafting (Chang Gung Memorial Hospital)	CRFF-OCT	Improved AI algorithms with the CADe system	N/A	N/A	N/A	N/A	N/A	N/A	CRFF-OCT with AI was effective in accurately quantifying melanin and detecting melanocyte recovery early.	The absence of papillary rings and the presence of large dendritic melanocytes around hair follicles are used to evaluate vitiligo stability and differentiate it from other hypopigmented skin lesions.
Wang et al. (2024)[13]	China	Vitiligo (n=4352)	Image datasets (Shenzhen People’s Hospital)	N/A	XGBoost	N/A	N/A	98%	0.99	N/A	N/A	The model showed a 73% accuracy and an AUC of 0.79 in the onset of segmental or nonsegmental vitiligo prediction.	This model could identify significant risk factors for vitiligo and predict its subtypes at an early stage.
		Non-vitiligo (n=28,270)			SVM			96%	0.96
					RF			97%	0.97
Hillmer et al. (2024)[3]	France	Vitiligo (n=169)	Images dataset and patients (Department of Dermatology of the University Hospital of Bordeaux, France)	N/A	CNN with FotoFinder body scan system	N/A	N/A	93%	N/A	N/A	N/A	From 0% (no depigmentation) to 100% (full depigmentation), the determined percentage in the vitiligo mask images corresponded to the estimated depigmented area of the face.	This model could score vitiligo severity from 0% to 100%. By conducting this assessment over time, a more objective scoring could be achieved.
Marri et al. (2024)[18]	India	Vitiligo (n=69)	A tertiary-care facility’s outpatient dermatology department (Shri B M Patil Medical College, Hospital and Research Centre, University, Vijayapura, Karnataka, India)	N/A	Aysa AI App	100%	100%	99.7%	N/A	N/A	0.985	Plotting the location on a human model, creating a patient profile, identifying lesion shape, and responding to questions about duration and symptoms were all performed with the application.	Even though this app has been focused on how to make a diagnosis, this app could potentially assist in tracking lesions’ changes, in home and clinical settings.
		Non-vitiligo (n=631)
Jalaboi et al. (2023)[20]	Denmark	Vitiligo (n=86)	Images dataset (DermNetNZ and SD-260)	Dermoscope	Efficient Net-B2, ConvNet, and ImageNet	0.90±0.05	1.00±0.00	N/A	N/A	N/A	0.95±0.02	By identifying diagnosis- explaining features, each with a unique localisation, and training ConvNets to concentrate on the pertinent area of the image for each characteristic, the methods were integrated in this work.	This model could distinguish pigmentation, lesion borders, and distribution. Thus, it could track the lesion evolution of vitiligo.
		Non-vitiligo (n=468)
Sharma et al. (2023)[17]	India	Vitiligo (n=355)	Images dataset (Kaggle)	Dermoscope	Inception V3+naive Bayes	99.5%	N/A	99.5%	0.997	0.995	0.995	The Inception V3 feature extractor can distinguish between vitiligo and healthy skin conditions using a variety of classifiers, such as CNN, random forest, decision trees, and naive Bayes.	This CNN model could detect early vitiligo with convolution layers by extracting and pooling layers for dimension reduction, converting multidimensional layers, and connecting them.
		Non-vitiligo (n=847)			Inception V3+CNN	99.8%		99.8%	1.00	0.998	0.998
					Inception V3+random forest	99.9%		99.9%	1.00	0.999	0.999
Guo et al. (2022)[4]	China	Vitiligo (n=167)	Image database (Department of Cosmetic Laser Surgery in the Hospital for Skin Disease and Institute of Dermatology, Chinese Academy of Medical Sciences, and Peking Union Medical College)	DSLR cameras (EOS 550D/800D, Canon, Tokyo, Japan; FinePix S9500, Fujifilm, Tokyo, Japan)	YOLO v3 architecture	92.91%	N/A	N/A	N/A	N/A	N/A	Size changes were the same for the AI model, trained dermatologists (W=0.812, P<0.05), and Photoshop analysis (P=0.075, P=0.212, respectively).	This model could detect and localise vitiligo lesions. VColor is more sensitive to changes in pigmentation than VASI.
		Non-vitiligo (n=100)			UNet++ architecture	72.41%
Zhang et al. (2021)[5]	China	Vitiligo (n=712)	Images dataset (DermNet NZ, DermNet, DermIS,	Dermoscope	VGG	0.9721±0.002	0.9629±0.012	N/A	0.9946±0.005	N/A	0.9684±0.005	Patients can use their cellphones to upload pictures of skin lesions, and depending on	This convolutional layer could capture high-level features of images. To provide the output as class probabilities associated
		Non-vitiligo (n=629)			ResNet	0.9515±0.010	0.9567±0.016		0.9848±0.007		0.9584±0.004
					DenseNet	0.9623±0.032	0.9609±0.027		0.9819±0.027		0.9618±0.030
			AtlasDerm, SD-260, DanDerm, and Kaggle)									the CNN output and the patient’s metadata, doctors can decide if an outpatient examination is necessary.	with each class (vitiligo or not), the classifier module uses fully connected layers.
Liu et al. (2020)[14]	China	Vitiligo (n=30,000)	Images dataset (ISIC 2018)	Dermoscope	Xception	88%	75.8%	84.5%	0.911	N/A	N/A	Due to their strong contrast under various colour space transformations, lesion regions can be efficiently recognised using the fusion approach, which also lowers the rate of misdiagnosis.	Even though this model showed low specificity, it has the potential to assess contrast in vitiligo lesions. The fusion method lowered the rate of misdiagnosis and efficiently recognised the influence of lesion sites.
		Non-vitiligo (n=8677)			Inception V3	88.4%	75.2%	84.6%	0.912
					VGG16	89.5%	73.2%	84.9%	0.917
					ResNet50	89.6%	77.5%	86.1%	0,922
Neri et al. (2020)[15]	Italy	Vitiligo (n=15)	Patients with vitiligo, ranging in age from 28 to 60 years old	N/A	Semi- automatic tool	N/A	N/A	N/A	N/A	N/A	N/A	By overlaying the identified patches of every patient on a face template using an occurrence map, the tools additionally enable statistical data interpretation.	The majority of the vitiligo patch area was dark red, indicating that the same areas were identified from the majority of global threshold values, which mostly influence the patch’s contour for vitiligo.
Chica et al. (2019)[11]	Ecuador	Vitiligo (n=20)	Patient images (Hospital in Cuenca, Ecuador)	Conventional digital photographic camera	ICA	100%	99.99%	N/A	N/A	N/A	0.9046	This neural network could distinguish vitiligo and healthy skin in areas up to 1 × 1 pixels.	The ANN with an MLP architecture could classify the lesion as either vitiligo-prone or healthy skin based on the pattern of light reflection.
					FCM	100%	99.87%				1
					MLP	100%	100%				1
Kislal and Halasz (2013)[21]	USA	N/A	N/A	Canon SD 850 IS 8 megapixel camera, Canon USA Inc., Lake Success, NY, USA	XML-RPC by a Ruby on Rails web framework	N/A	N/A	59%	N/A	N/A	N/A	Vitiligo was measured by this software using digital photos. Clinicians may track the severity of these conditions and, consequently, measure the impact of treatment thanks to the web-based program.	After assessing how many depigmented pixels there are in a picture, the program was able to convert the pixels to square millimetres.
Jolivot et al. (2013)[19]	Thailand	Vitiligo (n=110)	Clinical images (Department of Dermatology of Hospital Kuala Lumpur and Hospital Serdang, Malaysia)	ASCLEPIOS, a MSI system	ANN	N/A	N/A	N/A	N/A	N/A	N/A	The region of vitiligo lesions is obviously impacted by the melanin map [melanin (STD 0.097), thickness of the epidermis (STD 0.016); dermal thickness (STD 0.041); haemoglobin concentration (STD 0.012)]	The multispectral imaging system provides skin reflectance images in visible and near-infrared domains. The algorithm analyses this cube, retrieving skin parameter maps and enabling quantitative measurement of cutaneous tissue.
		Non-vitiligo (n=22)
Cazzaniga et al. (2009)[16]	Italy	Vitiligo (n=325)	Patients with different varieties of vitiligo (focal vs. generalised)	N/A	ANN (discriminative and regression network)	88.94%	80.95%	85.85%	0.931	N/A	N/A	In this work, a predictive method for the amount of time required to achieve a specific level of depigmentation in vitiligo patients receiving excimer laser treatment using an incremental dose strategy was created.	This method developed a mathematical model that required a preliminary hypothesis of a linear connection between time and repigmentation rate.

CRFF-OCT: high-cellular-resolution full-field optical coherence tomography; CADe: computer-aided detection; AUC: area under the curve; CNN: convolutional neural network; SVM: Support Vector Machine; RF: Random Forest; AI: artificial intelligence; DSLR: digital single-lens reflex; ANN: artificial neural network; ICA: Independent Component Analysis; FCM: Fuzzy C-Means; MLP: Multilayer Perceptron Type; XML-RPC: Extensible Markup Language – Remote Procedure Call; ASCLEPIOS: Analysis of Skin Characteristics by Light Emission and Processing of Images of Spectrum; MSI: multispectral imaging

From the systematic review of 15 studies, it was found that the sensitivity and specificity ranged from 72.41% to 100% and 73.2% to 100%, respectively. Accuracy ranged from 59% to 99.9%. Precision and F1 score ranged from 99.5% to 99.9% and 90.46% to 100%, respectively. NNs such as Inception V3 and ResNet showed high efficacy in sensitivity and specificity. The majority of the studies focused on diagnostic performance rather than explicitly grading the severity of vitiligo.

Seven studies (Liu et al. (2020),[14] Sharma et al. (2023),[17] Zhang et al. (2021),[5] Guo et al. (2022),[4] Jalaboi et al. (2023),[20] Marri et al. (2024),[18] and Zhong et al. (2024)[12]) out of 15 studies were eligible for statistical analysis in the diagnostic meta-analysis [Figure 4]. These included studies reported sensitivity, specificity, number of total subjects, and vitiligo cases explicitly as the key information to be input and pooled in statistical analysis. The 15 studies were previously included in the data synthesis to describe their findings qualitatively. In Figure 4, subgroup analysis was performed and visualised in forest plots, categorised based on the type of NNs (Inception V3, VGG, ResNet, and others). Some studies (Liu et al. (2020),[14] Sharma et al. (2023),[17] Zhang et al. (2021),[5] and Guo et al. (2022)[4]) reported more than one NN and thus appear in multiple subgroups. The use of letters a, b, and c following the publication year (e.g., Sharma et al. 2023a, 2023b)[17] was necessary due to limitations in meta-analysis programmes, which do not allow identical author-year citations during statistical analysis, as they conduct several assessments in the same study.

Figure 4.

Forest plot of artificial intelligence diagnostic performance

Overall, the NNs showed high sensitivity and specificity. In the multiple analysis, each of the NN families was analysed. For Inception V3 (n = 4),[14,17] sensitivity and specificity ranged from 88% to 100% and 75% to 100%, respectively. In VGG (n = 2),[5,14] the sensitivity ranged from 90% to 97%, and its specificity ranged from 73% to 96%. For ResNet (n = 2),[5,14] the sensitivity and specificity ranged from 90% to 95% and 78% to 96%, respectively. Various other NN families (Swin Transformer networks, Aysa AI App, Xception, ANN, YOLO v3 architecture, and UNet++ architecture) showed both sensitivity and specificity from 72% to 100%.[4,5,12,14,18,20]

The plot’s upper left section contained the majority of the machine learning models, suggesting relatively high sensitivity (>0.72) and high specificity (>0.72). Based on the affected region, these models aided in assessing the severity of the disease and were quite successful in differentiating between vitiligo and other conditions. Plotting the data into the summary receiver operating characteristics (SROC) curve revealed that, as Figure 5 illustrates, all seven experiments were over the dotted line.

Figure 5.

Summary receiver operating characteristic curve of the studies

This forest plot in Figures 6 and 7 summarises the sensitivity of diagnostic tools in assessing vitiligo severity across 15 studies. A great diagnostic performance was shown in high pooled sensitivity 0.955 (95% CI: 0.916–0.976) with the random-effects model. The majority of the studies reported a sensitivity of more than 90%. Large heterogeneity between studies was found (I² =95.7%, P < 0.001). The prediction interval was between 0.572 and 0.997, indicating that despite the good diagnostic performance, the results may be varied across settings.

Figure 6.

RStudio forest plot for sensitivity analysis

Figure 7.

RStudio forest plot for specificity analysis

The forest plot in Figure 6 shows the specificity of AI in diagnosing vitiligo. Using a random-effects model, the pooled specificity was 97% (95% CI: 0.905–0.991), suggesting the ability to distinguish non-severe cases. The majority of the studies showed high specificity, nearly 100%. However, due to variability across study designs or tools, a significant heterogeneity was found (I² =98.1%, P < 0.0001). In the future diagnostic performance settings, the results may be different, as shown in a wide prediction interval, ranging from 0.149 to 1.000.

The SROC curve in Figure 8 shows the overall diagnostic performance of AI in diagnosing vitiligo. The steep curve in the upper left corner shows high sensitivity and specificity. The AUC was 0.96, suggesting high diagnostic accuracy, with the red diamonds showing the estimation summary.

Figure 8.

RStudio SROC curve

As indicated in Table 2, the majority of the included studies’ GRADE Quality of Evidence Assessments indicated high-to-low evidence quality because of research design, bias risk, and indirectness restrictions. Small sample sizes and variations in study designs are the main drawbacks. Nevertheless, this study offers important information on vitiligo diagnosis in a clinical environment, despite the little imprecision brought on by the large CI.

Table 2.

GRADE assessment of AI models for vitiligo diagnosis

Outcome	Number of Studies	Study Design	Risk of Bias	Inconsistency	Indirectness	Imprecision	Publication Bias	Certainty of Evidence
Diagnostic accuracy of Inception V3	4 (Liu et al. 2020, Sharma et al. 2023a, Sharma et al. 2023b)	Quasi-experimental	Not downgraded	Not downgraded	Not downgraded	Not downgraded	Not applicable	High ⨁⨁⨁⨁
Diagnostic accuracy of VGG	2 (Liu et al. 2020a, Zhang et al. 2021a)	Quasi-experimental	Not downgraded	Downgraded	Not downgraded	Not downgraded	Not applicable	Moderate ⨁⨁⨁⬤
Diagnostic accuracy of ResNet	2 (Liu et al. 2020b, Zhang et al. 2021)	Quasi-experimental	Not downgraded	Downgraded	Not downgraded	Not downgraded	Not applicable	Moderate ⨁⨁⨁⬤
Diagnostic accuracy of Other AI Models	5 (Guo et al. 2022, Jalalobi et al. 2023, Marri et al. 2024, Zhang et al. 2021b, Zhong et al. 2024)	Quasi-experimental	Downgraded	Downgraded	Not downgraded	Not downgraded	Not applicable	Low ⨁⨁⬤⬤

Discussion

This systematic review and meta-analysis highlighted the direct comparison of several NNs in the diagnosis of vitiligo. We performed a statistical analysis to draw a conclusion in order to improve the quality of the systematic review. The included research demonstrated how AI might be used to define the severity of vitiligo and draw the lesion margin.

Most studies reporting on AI have focused on pigmented skin neoplasms, despite the fact that DL has been increasingly popular in dermatology in recent years. There is little information on depigmented skin disorders such as vitiligo.[4] We analysed the outcomes based on several parameters (specificity, sensitivity, accuracy, AUC, precision, and F1-Score) to have a global vision of their performance. The overall outcome of using NNs in machine learning was great in diagnosing and evaluating the severity of vitiligo in patients. Because the visual perception of various skin tones is subjective and can vary greatly from person to person, pigmentation evaluation is difficult. There is currently no widely used measurement to evaluate the pigmentation of the skin in vitiligo studies. Lesions can be classified based on the proportion of pigment loss. Another validated scoring method that offers more accurate pigmentation assessments is VASI.[23,24,25]

To achieve some degree of repigmentation, vitiligo typically takes a lengthy course of treatment; thus, early detection and diagnosis confirmation are needed. There are clear benefits to managing vitiligo patients when a prediction rule for clinical response is available. To the best of our knowledge, no formal attempt has been made to construct a predictive rule, and there are very few studies that evaluated characteristics linked to clinical response in vitiligo.[16]

A study by Lu et al. (2024)[22] stated that CRFF-OCT with AI could evaluate vitiligo stability and differentiate from other hypopigmented skin lesions, as shown by the absence of papillary rings and the presence of large dendritic melanocytes around hair follicles. Features of repigmentation were recorded and examined. Melanin loss was indicated by a noticeable dark band in the skin of vitiligo lesions, as shown by the CRFF-OCT study.[22]

Early therapeutic intervention is recommended as treatment is more effective for newly established lesions than for older ones.[26] As no quantitative diagnostic tools are already available, more objective methods are essential in differentiating vitiligo from other hypopigmented disorders so an accurate treatment plan can be carried out.[27]

The included studies have limitations; mainly, all the designs were non-randomised controlled trials. Furthermore, there is a heterogeneity in devices used to take the picture of the lesion and the NNs that were used to process the images. To acquire better evidence-based quality, research with a randomised controlled trial (RCT) design is required to validate the use of AI in diagnosing and evaluating the severity of vitiligo.

Conclusion

The AI for diagnosing and evaluating the severity of vitiligo showed high sensitivity, specificity, and accuracy, with the majority of the included studies showing a low risk of bias. However, none of the studies were RCTs. Therefore, to get a higher evidence-based quality, studies with an RCT design on this topic need to be conducted.

Key messages

Artificial intelligence with its subset, neural network, such as Inception V3, visual geometry group, and ResNet, showed potential in diagnosing and evaluating the severity of vitiligo. However, none of the studies were RCTs. Therefore, to get a higher evidence-based quality, studies with an RCT design on this topic need to be conducted.

Conflicts of interest

There are no conflicts of interest.

Funding Statement

Nil.