This observational validation study enrolled N = 265 female adults from age 20 to 59 representing Fitzpatrick phototypes I - VI who met the inclusion criteria of healthy skin, no history of cosmetic procedure and self-declaration of ethnicities, either Caucasians, East Asians, South-East Asians, Middle Asians, Hispanics, or African American. Participants were analyzed in three strata: Fitzpatrick I - II (n = 58), III - IV (n = 61), V - VI (n = 146), following the standard Fitzpatrick scale and segmentation. Mean ages (±SD) were: I - II: 35.4 ± 14.6, III - IV: 37.3 ± 16.2 and V - VI: 36.0 ± 14.1, respectively.

Figure 1. Overview of study design and analytical pipelines (CTK Derma AI vs baseline), evaluation against expert grading, phototype-stratified analysis, and repeatability testing.

The study was conducted in Korea, in accordance with the Declaration of Helsinki and approved by CTK IRB committee, and written informed consent was obtained from all participants.

All facial skin images were captured using a handheld diagnostic device based on the CTK ChoiceDx platform (e.g., Dx-Pro^® portable system; Figure 2), under a standardized and fixed capture protocol, including controlled distance and positioning, closed optical geometry at 6500 K for pore and wrinkle imaging, and spot imaging with cross-polarized filtering, and the consistent CMOS camera settings, to minimize operational variability.

(a) (b)

Figure 2. Handheld CTK ChoiceDx device (e.g., Dx‑Pro^® portable system) and representative standardized capture setup.

When feasible, a reference target (neutral gray card and/or color calibration chart) was included within the imaging field to enable post hoc normalization of brightness and color characteristics across captures.

Two primary regions of interest (ROI) were defined for quantitative analysis, consistent with established facial aging assessment frameworks. The periorbital region was selected for wrinkle evaluation, reflecting its sensitivity to fine lines and aging-related structural changes, whereas the central cheek region was selected for pore conspicuity, pigmentation (spots) and skin tone (L), as it represents a clinically relevant and relatively homogeneous area for assessing skin texture and discoloration [1][8]. For each ROI, image-based features were extracted using the CTK Derma AI pipeline or conventional image processing (Control). Specifically: Pore analysis was performed by detecting circular or elliptical microstructures corresponding to follicular openings, followed by quantification of their density, size distribution, and contrast relative to surrounding skin regions. Spot analysis targeted regions of hyperpigmentation, identified through local intensity deviation and chromatic irregularity, with features such as area coverage, edge definition, and color variance contributing to the score. Wrinkle analysis was based on the detection of linear and curvilinear ridge-like structures, incorporating measures of length, depth-associated contrast, orientation consistency, and spatial frequency within the periorbital region. L-value was analyzed at the cheek region for skin tone. Extracted features were integrated into composite scores using algorithm-specific weighting schemes optimized against expert visual grading.

All image analyses were performed in a fully automated manner without manual intervention to eliminate operator bias. For three attributes (pores, spots, wrinkles), analysis outputs were normalized to a continuous 0 - 100 scale, where higher scores represent better skin condition (i.e., lower presence or severity of the corresponding aging signs), and lower scores indicate increased prominence or deterioration. Skin tone analysis output was shown as L-value (higher is lighter skin tone, lower is darker skin tone) This normalization enables cross-subject and cross-attribute comparability and facilitates statistical correlation with expert grading.

These quantitative scores were subsequently compared with expert visual grading using correlation-based metrics and dependent correlation analyses, as described below.

A panel of dermatologists and/or trained expert graders (N = [3]) scored the following attributes on a color-calibrated monitor:

1) Pores conspicuity.

2) Pigmented spots/uneven pigmentation (“spots”).

3) Wrinkles.

4) Skin Tone.

Clinical scale selection and rubric design were informed by validated photoaging and skin aging scoring literature with 0 - 4 grading scale (0: least severe, 1: less severe, 2: moderate, 3: severe, 4: very severe) (photonumeric scales; SCINEXA; related reviews) [5]-[8][10]. Skin tone expert grading was carried out with 0 - 4 grading scale (0: very light, 1: light, 2: moderate, 3: dark, 4: very dark).

Inter-rater reliability was assessed using intraclass correlation coefficients (ICC) under standard formulations [20][21].

Objective color metrics were computed in CIELAB (L*a*b*), a device-independent color space standardized in international colorimetry practice and widely used in industry for color specification and difference calculations [22][23]. This provides a consistent and interpretable foundation for skin tone quantification across devices and illumination when combined with calibration.

2.5.1. CTK Derma AI Analysis

CTK Derma AI Analysis comprised:

1) Face/ROI localization.

2) Automated capture-condition calibration (color constancy/correction using reference target when available; learning-based normalization otherwise), motivated by evidence that AI-based color constancy can improve dermatologic image consistency [16].

3) Objective feature extraction for pores, spots, wrinkles, and objective skin tone metrics in calibrated CIELAB space [22][23].

4) Supervised mapping from calibrated features to predicted attribute scores aligned to expert grading. For the model development, the dataset was randomly divided into a training dataset (n = 132) and an independent validation subset (n = 133). All reported performance metrics by CTK Derma AI were derived from the validation dataset.

5) All image capture calibration and analysis were then carried out by the validated CTK Derma AI framework.

2.5.2. Control: Conventional Feature Detection without Calibration

In short, baseline means conventional non-AI image processing-based feature detection without capture calibration. The control represents conventional rule-based image processing methods widely used prior to AI-based dermatological image analysis. Therefore, the control used the same ROI definition but analyzed images as captured, without calibration; pores/spots/wrinkles were quantified via conventional feature detection and texture/contrast descriptors, and skin tone was estimated using non-calibrated color statistics.

To evaluate measurement repeatability across all skin tones, each participant underwent three repeated captures under the same protocol. Repeatability was quantified using test-retest ICC and coefficient of variation (CV). ICC methodology follows established guidance for rater/repeatability reliability [20][21].

The primary endpoint was Pearson’s correlation coefficient (r) between algorithm-derived scores and expert visual grading for pores, spots, and wrinkles. Because CTK Derma AI and Control were evaluated on the same participants, differences between correlations were treated as dependent correlations. The pre-specified inferential approach was based on Fisher’s r-to-z transformation within a Steiger/Meng-style framework, implemented in R, as recommended for overlapping or dependent correlation comparisons [24]-[26]. Repeatability across the three repeated captures was assessed using the coefficient of variation (CV) and the intraclass correlation coefficient (ICC) based on a two-way random-effects, absolute-agreement model [ICC (2, 1)] [20][21]. Complementary interpretation followed the principle that correlation strength alone does not equate to full method agreement, as noted by Bland and Altman [27].

A total of N = 265 participants were analyzed across phototype strata: Fitzpatrick I - II (n = 58; mean age 35.4 ± 14.6), Fitzpatrick III - IV (n = 61; mean age 37.3 ± 16.2), and Fitzpatrick V - VI (n = 146; mean age 36.0 ± 14.1) (See Table 1).

Table 1. Participant characteristics by phototype stratum.

Across the full cohort and all skin tones, repeatability was high. Each subject underwent three repeated captures under the standardized protocol, and the overall coefficient of variation (CV) was 4.5% across all phototypes, indicating strong measurement stability (see Table 2). Repeatability was additionally summarized with test-retest ICC following established reliability framework after final computation.

Table 2. Repeatability across three repeated captures (template).

Across all participants (N = 265), CTK Derma AI showed significantly higher correlations with expert grading than control for all evaluated attributes (pores, spots, wrinkles). For each attribute, Pearson’s r was computed between algorithm outputs and expert scores, and the difference between dependent correlations (same participants) was tested using Fisher r-to-z transformation within Steiger-type dependent-correlation tests, implemented in R.

Overall results are summarized in Table 3 (Overall: N = 265) and visualized in Figure 3.

Table 3. Correlations vs expert grading and dependent-correlation comparison: Overall (N = 265).

*p(Δr) computed using Fisher r-to-z dependent-correlation tests (Steiger-type) in R. **95% CI(Δr) computed using dependent-correlation CI procedures consistent with Steiger/Meng frameworks (available via R implementations such as cocor).

Phototype-stratified analyses showed that CTK Derma AI maintained robust correlations with expert grading across all strata and outperformed baseline for each attribute within each stratum: Fitzpatrick I - II (n = 58), III - IV (n = 61), and V - VI (n = 146). Differences in r between Pipeline A and B were evaluated using the same dependent-correlation Fisher r-to-z/Steiger-type tests in R, with confidence intervals reported via the same framework.

(a) (b)

(c) (d)

(e) (f)

(g) (h)

Figure 3. Correlations vs expert grading and dependent-correlation comparison (overall).

Phototype-stratified results are presented in Table 4 (Stratified) and visualized in Figure 4. Example skin images of pore, spot and wrinkle at the original, control analysis (without calibration), and CTK Derma AI analysis were shown in Figure 5.

Table 4. Correlations vs expert grading and dependent-correlation comparison: stratified by phototype across Fitzpatrick I - II, III - IV, and V - VI.

*, ** same as Table 3.

(a)

(b)

(c)

Figure 4. Correlations vs expert grading and dependent-correlation comparison (by Phototype).

Figure 5. Example skin images of pore, spot and wrinkle at the Original (left), Control analysis (without calibration, middle), and CTK Derma AI analysis (right) for pore, wrinkle and spot.

Across all skin tones and all evaluated attributes, CTK Derma AI achieved consistently superior agreement with expert grading versus the baseline workflow with statistically significant difference, and demonstrated highly repeatability with a provisional ICC (2, 1) of 0.975 (95% CI, 0.970 - 0.979), that was consistent with the low overall coefficient of variation (4.5%) across three repeated captures (Table 2). These findings support CTK Derma AI as a calibrated, phototype-robust image-based measurement framework.