Quantitative CRFF‐OCT Imaging Features for Characterization of Disease Activity in Non‐Segmental Vitiligo: A Machine Learning–Assisted Study

ABSTRACT

Vitiligo is an autoimmune pigmentary disorder characterized by progressive melanocyte loss and unpredictable activity. Objective of disease activity remains challenging, particularly in lesions with subtle clinical changes Cellular‐resolution full‐field optical coherence tomography (CRFF‐OCT) enables non‐invasive, high‐resolution, histology‐like visualization of skin microstructure. This study evaluated the feasibility of integrating CRFF‐OCT with machine learning‐assisted quantitative analysis for imaging‐based characterization of active and stable vitiligo lesions. Fifty patients with non‐segmental vitiligo were prospectively enrolled (2021–2022). CRFF‐OCT imaging was performed on lesional, perilesional, and normal‐appearing skin within the same anatomical region. Quantitative features describing epidermal structure, dermal–epidermal junction (DEJ) morphology, and pigment‐associated reflectivity were extracted. A machine learning‐assisted computer‐aided detection (CADe) framework incorporating 13 features was developed for image‐level classification of lesion activity. CRFF‐OCT imaging demonstrated distinct microstructural patterns between active and stable lesions. Basal epidermal pigment‐associated reflectivity was significantly lower in stable lesions compared with active lesions (9.94% ± 10.06% vs. 21.84% ± 11.08%), with corresponding differences in lesion‐to‐normal reflectivity ratios (0.21 vs. 0.56). Quantitative analysis revealed significant differences in epidermal thickness, DEJ associated reflectivity, inter‐layer contrast, and reflectivity heterogeneity. Among the 13 extracted features, 9 differed significantly between groups and were incorporated into classification models. The CADe framework achieved a maximum image‐level classification accuracy of 80.6% using a support vector machine model. These findings demonstrate the feasibility of CRFF‐OCT‐based quantitative imaging for objective characterization of vitiligo lesion status and support its potential role in disease activity assessment and longitudinal monitoring.

Abbreviations

AI

Artificial intelligence

AUC

Area under the ROC curve

CAD

Computer‐aided diagnosis

CRFF‐OCT

Cellular‐resolution full‐field full‐field optical coherence tomography

DEJ

Dermal‐epidermal junction

FN

False negative

FOV

Field of view

FP

False positive

k‐NN

k‐nearest neighbours

LDA

Linear discriminant analysis

LR

Logistic regression

NPV

Negative predictive value

OCT

Optical coherence tomography

PCA

Principal component analysis

PPV

Positive predictive value

RCM

Reflectance confocal microscopy

ROC

Receiver operating characteristic

SVM

Support vector machines

TN

True negative

TP

True positive

UV

Ultraviolet

VASI

Vitiligo Area Scoring Index

VIDA

Vitiligo Disease Activity index

1. Introduction

Vitiligo is a common autoimmune pigmentary disorder characterized by progressive loss of functional melanocytes, resulting in well‐demarcated depigmented patches that substantial psychosocial burden comparable to other chronic inflammatory skin diseases [1]. Accurate assessment of disease activity is clinically important, as ongoing immune‐mediated melanocyte injury during the active phase may influence therapeutic response and long‐term outcome. However, distinguishing active from stable vitiligo remains challenging in routine practice, particularly when clinical signs are subtle or evolving [2].

Current assessment of vitiligo activity relies primarily on clinical evaluation, including patient‐reported progression, morphological features such as hypochromic borders or confetti‐like depigmentation, and scoring systems such as the Vitiligo Disease Activity (VIDA) index. Although widely used, these approaches are inherently subjective and may lack sensitivity for detecting early or subclinical changes. Adjunctive tools, including Wood's lamp examination, dermoscopy, and serial photography, improve visualization of pigmentary contrast but provide limited information regarding tissue‐level alterations.

Advances in non‐invasive skin imaging have enabled In Vivo evaluation of pigmentary disorders at the microstructural level [3]. Cellular‐resolution full‐field optical coherence tomography (CRFF‐OCT) represents a recent imaging modality that provides real‐time, histology‐like vertical visualization of the epidermis and dermal–epidermal junction (DEJ), at micrometre‐scale resolution. This imaging orientation allows direct assessment of pigment‐associated reflectivity across epidermal layers within a single frame, and has shown feasibility in evaluating pigmentary disorders and monitoring treatment related changes [4, 5].

Despite these advances, interpretation of high‐resolution OCT images remains time‐consuming and operator‐dependent. Machine learning‐based quantitative analysis offers an opportunity to extract reproducible imaging features and to standardize interpretation. While machine learning approaches have been applied to dermatologic imaging, data on quantitative analysis of cellular‐resolution OCT imaging in vitiligo remain limited.

In this study, we aimed to evaluate the feasibility of integrating machine learning‐assisted computer‐aided detection (CADe) with CRFF‐OCT imaging to characterize microstructural differences between active and stable non‐segmental vitiligo lesions. By quantifying epidermal, DEJ‐associated, and inter‐layer reflectivity features, we sought to establish a quantitative imaging framework that may support objective assessment of disease activity and facilitate biomarker‐driven evaluation of vitiligo lesions.

2. Materials and Methods

2.1. Subject Recruitment

This prospective observational imaging study enrolled fifty individuals with active or stable non‐segmental vitiligo from a tertiary vitiligo clinic between May 2021 and Dec 2022. The study protocol was approved by the Institutional Review Board at Chang Gung Memorial Hospital (approval number 202002210B0A0), and written informed consent was obtained from all participants.

Clinical assessments were performed at monthly visits and included lesional depigmentation assessed using the Vitiligo Area Scoring Index (VASI) depigmentation scale, Vitiligo Disease Activity (VIDA) score, dermoscopy, digital photography, and Wood's lamp examination. Clinical disease activity status served as the reference standard and was independently determined by two dermatologists based on established morphological criteria, including presence of hypochromic borders, confetti‐like depigmentation, type 2B Koebner phenomenon, and recent disease onset (< 6 weeks prior to assessment). VIDA was used as supportive measure.

CRFF‐OCT image interpretation was performed by two dermatologists and one dermatopathologist, all blinded to clinical activity status. Model developers were blinded to clinical information during feature extraction and model development.

2.2. CRFF‐OCT Imaging Acquisition Protocol

CRFF‐OCT imaging was performed using the ApolloVue S100 Imaging System (Apollo Medical Optics, Taipei, Taiwan). Imaging was conducted on vitiligo lesions, perilesional skin, and normal‐appearing skin at baseline (D0) and follow‐up visits (D30, D90, and D180). Although longitudinal imaging was performed, the present analysis focuses on baseline cross‐sectional comparisons between active and stable lesions.

Images were acquired in both B‐scan (vertical) and en‐face (horizontal) modes. Each B‐scan image consisted of 952 × 717 or 944 × 717 pixels with an approximate spatial resolution of 0.5 μm per pixel and 8‐bit grayscale depth. The system achieved axial and lateral resolutions of approximately 1–2 μm, enabling visualization of epidermal layers, dermal–epidermal junction (DEJ), and superficial dermis. The field of view was 500 × 400 μm² for B‐scan images and 500 × 500 μm² for en‐face images.

For each lesion, a minimum of three B‐scan images were obtained. Images with motion artifacts, poor signal quality, or prominent adnexal structures were excluded. The average acquisition time per field of view was approximately 10 min, while feature evaluation and analysis required approximately 30 min per patient before machine learning analysis. Normal skin samples were obtained from perilesional unaffected skin within the same anatomical region, approximately 10 cm from the lesion.

2.3. Machine Learning‐Assisted Image Processing and Feature Extraction

A machine learning computer‐aided detection (CADe) framework was developed to quantify CRFF‐OCT features (Figure S1). The analysis pipeline consisted of (i) skin‐layer segmentation, (ii) definition of anatomically relevant regions of interest centered on the DEJ, and (iii) extraction of quantitative features describing reflectivity and microstructural morphology. To ensure accurate stratification, epidermal boundaries were manually annotated and verified prior to automated segmentation (Figure S2). The epidermis was first separated from the original OCT image, and morphological operations were applied to define an epidermal layer excluding the stratum corneum (L_SC). The mean thickness of L_SC was used to standardize DEJ‐centered compartments, including a “dark band” region (L_DB) and superficial dermis (L_DE), defined as fixed proportions of L_SC thickness. The dark band was further subdivided into upper and lower subregions (L_DB1 and L_DB2) to capture intra‐band heterogeneity. In total, seven partitioned layers were defined for feature extraction.

Based on these segmented layers, thirteen quantitative features were derived (Table S1), including layer‐specific mean reflectivity, reflectivity variability, epidermal thickness, DEJ undulation, and inter‐layer reflectivity contrasts. Inter‐layer contrast features (e.g., DB_b_diff., DB_EP_b_diff., and DB_DE_b_diff.) were calculated as differences between the mean reflectivity of adjacent layers. Reflectivity heterogeneity within the dark band (DB_horizontal_std.) was quantified using the standard deviation of vertically averaged reflectivity values. DEJ undulation was defined as the ratio of DEJ contour length to image width.

2.4. Machine Learning Classification Model

Four classifiers were trained using the extracted OCT features: logistic regression (LR), linear discriminant analysis (LDA) [6], k‐nearest neighbours (k‐NN) [7], and support vector machines (SVM) [8] Principal component analysis (PCA) [9] was applied for dimensionality reduction prior to classification. Model performance was evaluated using 10‐fold cross‐validation at the image level in a curated dataset comprising 144 OCT images (60 stable and 84 active). Images were treated as independent samples, and redundant or closely correlated images from repeated follow‐up of the same lesion were excluded.

In each fold, models were trained on k–1 partitions and tested on the remaining partition; this was repeated until each fold served once as the validation set. Images were classified as stable when the predicted probability was ≥ 0.5 and as active when < 0.5.

2.5. Statistical Analysis

Normality of continuous variables was assessed using the Kolmogorov–Smirnov test [10]. Normally distributed data were analysed using Student's t‐test, while non‐normally distributed data were analysed using the Mann–Whitney U test. A two‐sided p < 0.05 was considered statistically significant [11].

Diagnostic performance metrics, including accuracy, sensitivity, specificity, positive predictive value (PPV), negative predictive value (NPV), receiver operating characteristic (ROC) curves, and area under the ROC curve (AUC), were calculated using clinical activity status as the reference standard. All statistical analyses were performed using GraphPad Prism (version 11, GraphPad Software, San Diego, CA, USA) [12].

3. Results

3.1. Participant Characteristics

Fifty patients with non‐segmental vitiligo were included, comprising 21 men and 29 women, with a mean age of 43.0 ± 16.3 years. Lesions were most frequently located on the face (n = 27), followed by the trunk (n = 15), neck (n = 4), and limbs (n = 4). Based on clinical assessment, 28 patients were classified as having stable vitiligo (VIDA 0 to −1), and 22 as having active vitiligo (VIDA 1–4). The mean duration since the last active episode was longer in the stable group (1.92 ± 2.57 years) compared with the active group (0.12 ± 0.05 years). Lesional depigmentation, assessed using the VASI depigmentation scales, is summarized in Table 1; most lesions demonstrated ≥ 90% depigmentation (96%, 43/50%).

TABLE 1.

Clinical and demographic characteristics of patients with nonsegmental vitiligo.

Characteristics	N = 50
Sex (Male: Female)	21:29
Age (years; Mean ± SD)	43 ± 16.3
Vitiligo Lesion Distribution.
Face	27
Neck	4
Torso	15
Limbs	4
Duration to last active episode (years; Mean ± SD).
Stable	1.92 ± 2.57
Active	0.12 ± 0.05
Prior phototherapy exposure	48 (96%)
VIDA Score*
4	18
3	2
2	0
1	2
0	9
−1	19
VASI Score**
100% (Complete Depigmentation)	40
90% (Specks of Pigment Present)	3
75%	3
50%	3
25%	1

Abbreviations: IDA, Vitiligo Disease Activity Score; VASI, Vitiligo Area Severity Index.

^*VIDA Score: Stable (VIDA 0 to −1), Active (VIDA 1–4).

^**VASI Score: lesional depigmenation based on VASI depigmentation scale, not total VASI.

3.2. CRFF‐OCT Imaging Feature of Active and Stable Vitiligo

A total of 186 B‐scan CRFF‐OCT images were analysed, including 95 images from clinically active lesions and 91 from stable lesions. In CRFF‐OCT, pigment is represented by increased reflectivity within the basal epidermal layer. Stable vitiligo lesions exhibited a well‐defined hyporeflective dark band at the DEJ, consistent with the absence of pigment‐associated reflectivity. In contrast, active vitiligo lesions demonstrated a less well‐defined hyporeflective dark band at the DEJ, indicating residual pigment with variable degrees of depigmentation. While Wood's lamp examination identified hypochromic borders at lesion peripheries, CRFF‐OCT additionally detected residual pigment‐associated signals within lesion centers that appeared clinically indistinguishable from stable depigmented areas (Figure 1).

FIGURE 1.

Clinical, Wood's lamp, and CRFF‐OCT findings in stable and active vitiligo lesions. Stable lesions demonstrate well‐demarcated depigmented patches with homogeneous enhancement under Wood's lamp. On CRFF‐OCT, a well‐defined hyporeflective band is observed at the dermal–epidermal junction (DEJ), consistent with absence of pigment‐associated reflectivity. In contrast, active lesions show hypochromic borders and confetti‐like depigmentation clinically, with corresponding Wood's lamp enhancement. CRFF‐OCT reveals a less well‐defined and heterogeneous hyporeflective band at the DEJ, indicating residual pigment and ongoing melanocyte loss.

Quantitative analysis of basal epidermal reflectivity revealed significantly lower pigment‐associated signal in stable vitiligo lesions compared with active lesions (Figure 2). In stable vitiligo, the average basal layer pigment reflectivity was 9.94%, compared with 48.34% in normal skin (n = 30, p < 0.05). In active vitiligo lesions, the basal layer pigment reflectivity was 21.84% ± 11.08% compared with 40.53% ± 13.84% in normal skin (p < 0.05). We further compare the basal layer pigmentation in vitiligo lesions to normal skin in both stable and active groups, showing statistically significant differences in pigment ratios between the stable and active stages, with values of 0.21 and 0.56 (p < 0.05), respectively.

FIGURE 2.

Quantitative analysis of basal epidermal reflectivity in stable and active vitiligo lesions. (a) Basal epidermal reflectivity (%) in stable and active vitiligo lesions compared with normal skin. Stable lesions demonstrate significantly lower reflectivity compared with active lesions and normal skin, indicating reduced residual pigment. (b) Lesion‐to‐normal reflectivity ratio (L/N). Stable lesions show significantly lower ratios compared with active lesions, reflecting greater pigment loss. Data are presented as mean ± standard deviation. *p < 0.05; ***p < 0.001.

3.3. Predictability of Machine Learning‐Assisted CADe Algorithm for Vitiligo Stability

CRFF‐OCT imaging demonstrated marked differences in reflectivity patterns at and immediately above the dermal–epidermal junction (DEJ) between stable and active vitiligo lesions. Stable lesions consistently exhibited a more prominent hyporeflective band at the DEJ, whereas active lesions showed a less well‐defined hyporeflective region with greater variability in reflectivity (Figure 3). These differences were quantitatively captured by features derived from the DEJ‐centered dark band (L_DB), including mean brightness, horizontal standard deviation, and transverse standard deviation, all of which differed significantly between stable and active lesions (Table S1).

FIGURE 3.

CRFF‐OCT imaging features distinguishing stable and active vitiligo lesions. Representative B‐scan images demonstrate microstructural differences at the dermal–epidermal junction (DEJ) and within the epidermis. (a) Stable lesions show a well‐defined hyporeflective band at the DEJ with relatively uniform reflectivity patterns, consistent with the absence of pigment‐associated signal. (b) Active lesions demonstrate a less well‐defined and irregular DEJ band with increased reflectivity heterogeneity, indicating residual pigment and ongoing melanocyte loss. These imaging differences correspond to quantitative features, including DEJ‐associated reflectivity, inter‐layer contrast, and epidermal thickness, used for machine learning‐assisted classification.

In addition to dark band–specific metrics, inter‐layer reflectivity contrast features were evaluated to characterize differences between the dark band and adjacent epidermal and dermal layers, as well as between the upper and lower portions of the dark band. These contrast‐based features revealed distinct reflectivity gradients in stable versus active lesions, contributing to improved separation between lesion states. Epidermal thickness measurements also differed between groups, with stable lesions demonstrating greater epidermal thickness compared with active lesions (Figure 3). This difference was consistently observed across analysed images and was not influenced by ongoing treatment at the time of imaging.

To further characterize epidermal reflectivity, stratified analyses were performed for the epidermal layer, including the stratum corneum (L_EP) and the epidermal layer excluding the stratum corneum (L_SC). In both stratifications, mean reflectivity values differed significantly between stable and active lesions, whereas no significant differences were observed in reflectivity variability within these layers.

Among the thirteen quantitative CRFF‐OCT features extracted, nine showed statistically significant differences between stable and active vitiligo lesions (Table 2). These features were subsequently incorporated into the machine learning‐assisted CADe models for image‐level classification. Across models, the CADe system achieved a minimum classification accuracy of 77.8%. The highest performance was obtained using the support vector machine model, which yielded an accuracy of 80.6%, sensitivity of 71.7%, specificity of 86.9%, positive predictive value (PPV) of 79.6%, and negative predictive value (NPV) of 81.1% (Table 3). The ROC curve was shown in Figure S3.

TABLE 2.

Quantitative CRFF‐OCT features in active and stable vitiligo lesions.

Feature	Active lesions	Stable lesions	p
	Mean ± SD	Mean ± SD
EP_b_avg	49.952 ± 4.665	47.805 ± 4.961	0.009*
EP_b_std.	57.410 ± 4.133	57.828 ± 3.234	0.498
EP_thickness	60.678 ± 12.156	67.788 ± 18.295	0.010*
EP_undulation	1.365 ± 1.015	1.238 ± 1.534	0.068
SC_b_avg.	42.841 ± 5.722	42.417 ± 18.494	0.662
SC_b_std.	16.685 ± 2.225	2.571 ± 5.747	< 0.001*
DB_b_avg.	34.097 ± 6.127	29.722 ± 5.38	< 0.001*
DB_b_std.	10.160 ± 2.487	8.445 ± 2.132	< 0.001*
DB_horizontal_std.	6.744 ± 1.944	5.335 ± 1.861	< 0.001*
DE_b_avg.	34.507 ± 7.685	33.912 ± 11.464	0.710
DB_b_diff.	2.364 ± 2.558	3.345 ± 2.492	0.023*
DB_EP_b_diff.	5.596 ± 3.345	9.977 ± 4.755	< 0.001*
DB_DE_b_diff	−0.410 ± 5.650	−4.190 ± 10.361	0.012*

Note: b_avg, mean reflectivity; b_std, reflectivity variability.

Abbreviations: DB, dark band; DE, dermis; EP, epidermis; SC, stratum corneum.

^* p < 0.05 indicates a statistical significance.

TABLE 3.

Performance of machine learning–assisted CADe models for classification of vitiligo lesion activity.

Classifier	Accuracy (%)	Sensitivity (%)	Specificity (%)	PP (%)	NPV (%)	AUC
Logistic regression (LR)	77.78 (112/144)	70.00 (42/60)	83.33 (70/84)	75.00 (42/56)	79.55 (70/88)	0.84
Linear discriminant analysis (LDA)	77.78 (112/144)	66.67 (40/60)	85.71 (72/84)	76.92 (40/52)	78.26 (72/92)	0.83
k‐nearest neighbours (kNN)	79.86 (115/144)	71.67 (43/60)	85.71 (72/84)	78.18 (43/55)	80.90 (72/89)	0.81
Support vector machine (SVM)	80.56 (116/144)	71.67 (43/60)	86.90 (73/84)	79.63 (43/54)	81.11 (73/90)	0.84

Abbreviations: AUC, area under the receiver operating characteristic curve; kNN, k‐nearest neighbours; LDA, linear discriminant analysis; LR, logistic regression; NPV, negative predictive value; PPV, positive predictive value; SVM, support vector machine.

4. Discussion

Accurate assessment of disease activity in vitiligo remains a clinical challenge, particularly in lesions lacking overt inflammatory signs. In this study, we evaluated the feasibility of integrating cellular‐resolution full‐field optical coherence tomography (CRFF‐OCT) with machine learning‐assisted quantitative analysis to characterize microstructural differences between active and stable non‐segmental vitiligo lesions at the image level. By systematically quantifying epidermal, dermal–epidermal junction (DEJ)–associated, and inter‐layer reflectivity features, we demonstrated that CRFF‐OCT captures reproducible structural patterns that can be objectively summarized using a quantitative imaging framework.

Among the evaluated parameters, DEJ‐centered alterations were the most discriminative features. Stable lesions consistently exhibited a well‐defined hyporeflective band at the DEJ, whereas active lesions demonstrated a less sharply demarcated band with increased reflectivity heterogeneity. These differences were quantitatively captured through features describing dark‐band intensity, intra‐band variability, and inter‐layer contrast. An alternative explanation for the increased reflectivity heterogeneity observed within the DEJ‐associated hyporeflective band in active lesions is the presence of ongoing inflammatory activity. Active vitiligo is characterized by immune‐mediated melanocyte injury and disruption of the epidermal pigmentary unit. Because OCT signal intensity reflects tissue optical scattering properties, heterogeneity within the DEJ region may not solely represent pigment‐associated reflectivity but may also reflect inflammatory and microstructural alterations associated with active disease. Future studies integrating CRFF‐OCT with histopathological and molecular biomarkers may help further define the biological significance of these imaging signatures.

Stable lesions showed greater epidermal thickness than active lesions. This finding should be interpreted cautiously, as epidermal thickness may be influenced by anatomical site and prior treatment exposure. In particular, stable lesions may have received longer or more cumulative phototherapy before stabilization, whereas active lesions may have only initiated treatment. Therefore, the relative contribution of disease status, treatment duration, and anatomical variations cannot be determined in the present study. Future studies using lesion‐to‐normal thickness ratios and detailed cumulative phototherapy data may help clarify this observation. While the underlying biological mechanisms cannot be fully resolved by imaging alone, these reproducible structural differences support the role of CRFF‐OCT‐derived features as candidate imaging biomarkers for lesion status.

Objective assessment of vitiligo activity is clinically relevant, as conventional evaluation may underestimate subclinical or evolving disease. In this pilot study, the machine learning‐assisted CADe framework, based on interpretable hand‐crafted features, achieved encouraging but exploratory classification performance, with relatively high specificity (86.9%) and moderate sensitivity (71.7%), along with positive and negative predictive values of 79.6% and 81.1%, respectively. These findings should be interpreted as proof of concept, demonstrating the feasibility of feature‐based differentiation rather than establishing diagnostic thresholds. Importantly, the proposed framework significantly reduced analysis time compared with manual expert interpretation (reduced from 20–30 min to 3 min per case), suggesting potential utility in standardizing and streamlining evaluation of high‐resolution OCT datasets. Interestingly, some misclassified lesions may themselves be biologically informative. Lesions that were clinically classified as stable but predicted as active by the quantitative imaging model may represent subclinical structural alterations not yet appreciable on routine clinical examination or Wood's lamp assessment. Such cases suggest that CRFF‐OCT may detect early microstructural changes that precede overt clinical progression. Although longitudinal validation is required, this observation suggests a potential role for quantitative CRFF‐OCT as an adjunctive tool for identifying subclinical disease activity.

Non‐invasive imaging modalities have increasingly been recognized as adjunctive tools for evaluating vitiligo. Recent studies have shown that dermoscopic and reflectance confocal microscopy (RCM) features may predict treatment response to topical ruxolitinib in patients with nonsegmental vitiligo [13]. This finding highlights the potential of tissue‐level imaging to detect early biological changes preceding visible repigmentation. While RCM provides high‐resolution en‐face visualization, CRFF‐OCT offers complementary vertical, histology‐like visualization of the epidermis and DEJ, with greater penetration depth and suitability for quantitative analysis. The ability to assess pigment‐associated reflectivity across epidermal layers within a single vertical B‐scan frame may facilitate more intuitive interpretation and enable standardized feature extraction.

Our findings extend prior work on CRFF‐OCT in vitiligo. In earlier studies, this modality was applied to monitor vitiligo post‐interventional changes following melanocyte grafting, focusing on inflammatory responses and graft outcomes [5]. In contrast, the present study addresses an earlier stage of disease evaluation by focusing on pre‐interventional lesion characterization. This shift toward quantitative, imaging‐based assessment of lesion status highlights the potential role of CRFF‐OCT in guiding clinical decision‐making, particularly in determining disease stability prior to therapeutic intervention.

Several limitations should be acknowledged. First, although the study was conducted prospectively, external validation in an independent cohort was not performed, and generalizability across different clinical settings and imaging platforms remains to be established. Second, analyses were performed at the image level, and multiple images from the same patient may introduce intra‐subject correlation, potentially inflating performance estimates. Third, feature selection was conducted prior to cross‐validation, which may introduce optimistic bias. Given the relatively small sample size, patient‐level validation and nested cross‐validation were not feasible. Although longitudinal imaging was performed, the present analysis did not incorporate temporal features, and images were treated as independent samples. Future studies with larger datasets and patient‐level validation strategies will be required to address these limitations.

Despite these limitations, this study highlights the potential of CRFF‐OCT as a non‐invasive modality for quantitative characterization of vitiligo lesions. A key strength of this approach lies in the use of interpretable, biologically meaningful imaging features, which may facilitate clinical translation compared with black‐box deep learning models. By enabling objective assessment of microstructural changes, CRFF‐OCT based analysis may support more standardized evaluation of disease activity and longitudinal monitoring.

Further studies incorporating larger, multicenter cohorts, patient‐level validation strategies, and standardized imaging protocols will be essential to confirm these findings. Intergration with advanced machine learning or deep learning approaches may further enhance auromated feature extraction and classification, ultimately defining the role of CRFF‐OCT based imaging biomarkers in precision management of vitiligo.

5. Conclusion

To our knowledge, this study represents one of the first and largest investigations of machine learning‐assisted quantitative analysis of cellular‐resolution full‐field optical coherence tomography (CRFF‐OCT) for imaging‐based assessment of vitiligo disease activity. CRFF‐OCT enables in vivo, non‐invasive, real‐time, histology‐like visualization of epidermal and dermal–epidermal junction microstructure, allowing detailed characterization of pigment‐associated changes. Quantitative imaging features derived from CRFF‐OCT demonstrate feasibility in differentiating active and stable lesions and support a more objective and standardized approach to lesion assessment. Although these findings remain exploratory and external validation, they underscore the potential of CRFF‐OCT based quantitative imaging as an adjunctive tool for evaluating disease activity and monitoring vitiligo in clinical practice.

Author Contributions

Conceptualization: Chau Yee Ng. Methodology: Chau Yee Ng, I‐Ling Chen and Yi‐Ting Chen. Software: I‐Ling Chen, Yi‐Ting Chen, Feng‐Ling Tsai and Hsing Cheng. Validation: Chau Yee Ng, Yi‐Ting Chen, Yen‐Lin Huang and Tsung‐Fu Tsai. Formal analysis: Chau Yee Ng, I‐Ling Chen and Yi‐Ting Chen. Investigation: Chau Yee Ng, Yi‐Ting Chen, Yi‐Ching Lai and Sheng‐Ni Chen. Resources: Chau Yee Ng. Data curation: Chau Yee Ng, Yen‐Lin Huang, Tsung‐Fu Tsai, Yi‐Ting Chen and Yi‐Ching Lai. Writing – original draft preparation: Chau Yee Ng, Sheng‐Ni Chen, I‐Ling Chen and Yi‐Ting Chen. Writing – review and editing: Chau Yee Ng, I‐Ling Chen, Yi‐Ting Chen, Feng‐Ling Tsai and Sheng‐Ni Chen. Visualization: Chau Yee Ng, Yi‐Ting Chen, I‐Ling Chen and Hsing Cheng. Supervision: Chau Yee Ng. Project administration: Chau Yee Ng. Funding acquisition: Chau Yee Ng.

Funding

This work was supported by grants from Apollo Medical Optics Ltd. (XMRPG3L0841) for patient participation, image acquisition, interpretation, and device equipment.

Ethics Statement

BLINDED IRB approved this study (approval number 202002210B0A0).

Conflicts of Interest

I‐Ling Chen, Yi‐Ting Chen, Feng‐Ling Tsai, Hsing Cheng, Ling Huang are employees of Apollo Medical Optics Ltd. and participated for technical support; Chau Yee Ng, Yen‐Lin Huang, Tsung‐Fu Tsai, Sheng‐Ni Chen, Yi‐Jing Lai have no conflicts of interest and are in charge of study design, data analysis, patient participation recruitment, and literature preparation.

Supporting information

Figure S1: Workflow of the machine learning‐assisted computer‐aided detection (CADe) framework for quantitative analysis of CRFF‐OCT images. The pipeline includes image preprocessing, epidermal and dermal–epidermal junction (DEJ) segmentation, region‐of‐interest definition, quantitative feature extraction, and subsequent classification using machine learning models.

Figure S2: Flowchart of CRFF‐OCT image segmentation and skin layer partitioning for quantitative analysis. The workflow illustrates manual annotation of epidermal boundaries followed by automated segmentation of the epidermis and dermal–epidermal junction (DEJ). The epidermal layer excluding the stratum corneum (L_SC) is used to standardize depth, enabling definition of DEJ‐centered regions, including the dark band (L_DB) and superficial dermis (L_DE). The dark band is further subdivided into upper and lower compartments to capture reflectivity heterogeneity, resulting in seven partitioned layers for subsequent feature extraction.

Figure S3: Receiver operating characteristic (ROC) curves of machine learning–assisted CADe models for classification of vitiligo lesion activity.

ROC curves are shown for logistic regression (LR), linear discriminant analysis (LDA), k‐nearest neighbours (kNN), and support vector machine (SVM) models. The corresponding area under the curve (AUC) values were 0.84, 0.83, 0.81, and 0.84, respectively.

Table S1: Quantitative features used for lesion classification.

Data Availability Statement

The data that support the findings of this study are available on request from the corresponding author. The data are not publicly available due to privacy or ethical restrictions.