Epidermal renewal during the treatment of atopic dermatitis lesions: A study coupling line‐field confocal optical coherence tomography with artificial intelligence quantifications

Abstract

Objective

This study explores the application of Line‐field Confocal Optical Coherence Tomography (LC‐OCT) imaging coupled with artificial intelligence (AI)‐based algorithms to investigate atopic dermatitis (AD), a common inflammatory dermatosis.

Materials and methods

AD acute and chronic lesions (ADL) were compared to clinically healthy‐looking skin (ADNL). LC‐OCT was used noninvasively and in real‐time to image the skin of AD patients during flare‐ups and monitor remissions under topical steroid treatment for 2 weeks. Quantitative parameters were extracted from the images, including morphological and cellular‐level markers of epidermal architecture. A novel cellular‐level parameter, nuclei “atypia,” which quantifies the orderliness of epidermal renewal, was used to highlight abnormal maturation processes.

Results

Compared to healthy skin, AD lesions exhibited significant increases in both epidermal and stratum corneum (SC) thickness, along with a more undulated dermo‐epidermal junction (DEJ). Additionally, keratinocyte nuclei (KN) were larger, less compact, and less organized in lesional areas, as indicated by the atypia parameter. A higher degree of atypia was observed in chronic lesions compared to acute ones. Following treatment, all the parameters normalized to levels observed in healthy skin within 2 weeks, mirroring clinical improvements.

Conclusion

This study provides insights into the quantification of epidermal renewal using a noninvasive imaging technique, highlighting differences between ADL/ADNL and acute/chronic lesions. It also presents the AD treatment mechanism, paving the way for future investigations on AD and other skin barrier function‐related conditions.

1. INTRODUCTION

The epidermis, the outermost layer of the skin, plays a multifunctional role in protecting the body from external aggressors, such as mechanical, chemical, and microbial factors, as well as light exposure and pollution. Additionally, it contributes to maintaining skin homeostasis, regulating water flow exiting the body and body temperature. This essential role, often denoted by the term “skin barrier function,” relies significantly on the continuous renewal of the epidermal layer. Comprising stacked layers of cells, primarily keratinocytes, the epidermis varies in thickness from 40 μm to 1 mm, depending on anatomical location. ¹ Keratinocytes, originating from the basal layer—the deeper layer of the viable epidermis—migrate over approximately 28 days to the surface layer, known as the stratum corneum (SC). At this point, they transform into dead, keratin‐filled cells called corneocytes. The epidermis lies atop the skin's dermis, separated by the dermal–epidermal junction (DEJ).

The skin barrier function undergoes modification to varying degrees in different situations: physiological (e.g., age), dermatological pathologies (e.g., atopic dermatitis [AD]), or exposome‐related exposition (such as season, pollution, lifestyle, or stress). Significant alterations in the barrier function often result in visible and/or perceived manifestations, including redness, itchiness, burning, or stinging sensations, indicating sensitivity. Connecting these macroscopic‐level signs to the physiological mechanisms of epidermal renewal at the tissue, cellular, and molecular levels is crucial for understanding these modifications.

AD, the most common type of eczema, affects 223 million people worldwide, with up to 20% of children and 10% of adults. ² AD is an inflammatory skin disorder characterized by dry, itchy skin and eczema lesions, which come up chronically in successive flare‐ups and remissions. Lesions can appear on multiple body sites and can be divided into two types: acute lesions, which are recent, and chronic lesions, more ancient, with thicker, lichenified skin, and frequent marks of scratching (excoriation). Flare‐ups typically appear more frequently in cold weather and seem to be sensitive to pollution ³ , ⁴ as well as psychological stresses. ⁵ The diversity in body size, severity, triggers, and timeline, is one of the most challenging aspects of AD research as it adds a large variability to patient cohorts.

AD is relatively well known at a molecular level. It has been described that AD lesional skin shows an increased expression of Th2 inflammation‐related genes and a decreased expression of epidermal differentiation and barrier‐related genes. ⁶ , ⁷ , ⁸ It has also been studied at the tissular level using histological analysis of skin biopsies: in AD lesions, the epidermis is thicker, in part as a result of spongiosis, intercellular edema, and inflammatory infiltrates of lymphocytes. ⁹ , ¹⁰ The stratum corneum is also known to thicken in AD lesions. ¹¹

Despite being well studied, multiple questions remain with regards to AD: for example, what is the difference between chronic and acute lesions, what are the local mechanisms that precede the apparition of a lesion or precede its resolution, why do lesions appear on certain body sites and not others, why are they sensitive to external and internal stresses. Addressing these questions is crucial to improve patient care through better diagnosis and treatment. Future studies would greatly benefit from noninvasive in vivo methods, aiming to identify novel biomarkers, particularly at the less explored cellular and tissular levels. Using methods that can be used at the same location as clinical and molecular assessments holds promise for achieving a comprehensive multiscale outlook. The development of one such method is the primary goal of the present study.

Skin's morphology is typically assessed with histological cross‐sections on skin biopsies, which is invasive and suffers several limitations: possible modifications of thicknesses during the histological preparation process, lack of 3D, lack of quantifications, as well as difficulty in following longitudinally the flare‐up and treatments Thus, the development of noninvasive in vivo imaging techniques for the monitoring of dermatological pathologies and assessment of treatment efficacy has been a long‐time challenge for imaging research groups. Reflectance Confocal Microscopy (RCM), ¹² Multiphoton Microscopy (MMP), ¹³ and Optical Coherence Tomography (OCT) ¹⁴ are the three most common imaging techniques currently used in skin research to study internal skin structures and cellular properties in real‐time and in vivo. RCM and MMP typically present good lateral resolution but come short in penetration depth. RCM also has limited axial penetration. ¹⁵ OCT meets the conditions of good penetration depth and high axial resolution yet has limited lateral resolution.

Therefore, a new imaging technique has been developed based on the fusion of OCT and RCM to combine their advantages in terms of resolution and penetration depth. ¹⁶ This technique, called Line‐field Confocal Optical Coherence Tomography (LC‐OCT), provides a 3D imaging modality and can achieve axial and lateral resolution of around 1 μm and an imaging depth of 0.5 mm. It also presents a huge advantage in terms of acquisition time, that is, less than a minute for a full 3D field‐of‐view (1.2 mm × 0.5 mm).

High‐performance image processing algorithms harnessing the power of artificial intelligence (AI), and in particular machine learning, have recently been developed and applied to LC‐OCT images. A 2D U‐Net model based on Convolutional Neural Networks ¹⁷ , ¹⁸ was used to segment the skin layers and quantify histological metrics, such as layer thicknesses (viable epidermis, SC) and DEJ undulation. Taking advantage of the LC‐OCT cellular resolution, keratinocyte nuclei (KN) were also segmented from 3D LC‐OCT stacks to develop specific metrics, such as cell density, nuclei size and shape ¹⁹ through the training of a Deep Learning algorithm, called 3DStarDist. The cell network homogeneity has been quantified using a novel multiparametric parameter, atypia, comparing cells to their neighbors on morphological parameters. ²⁰

In this study, we apply LC‐OCT imaging coupled with AI‐based morphological quantifications to AD lesions, utilizing the morphological and cellular‐level metrics. The aim is to compare lesional areas to healthy‐looking control skin, as well as acute versus chronic lesions, and monitor over time the evolution of the flare‐ups under topical steroids treatment.

2. MATERIALS AND METHODS

2.1. Study population

The study was conducted in the spirit of the French and European Guidelines for Good Clinical Practice, the recommendations of the ICH (International Conference on Harmonization), EMA/CHMP/ICH/135/1995 adopted in November 2016 and according to the Helsinki Declaration in its latest version (Seoul 2008) and the laws and regulations in force. The study was not concerned with Jardé law (March 5, 2012) relative to research on human persons (decree n°2016‐1537, 16 November 2016 and decree n°2017‐884, 9 May 2017). All volunteers gave their written informed consent. The study took place at the Bioclinical Research Center (BRC) of L'Oréal Advanced Research located in Saint‐Louis Hospital (Paris, France). The subjects were recruited from the Department of Dermatology at Saint‐Louis Hospital and from the internal volunteer base of BRC.

In total, 22 lesions AD acute and chronic lesions (ADL) from six adult subjects (three men and three women, aged from 20 to 60 years old) with AD have included: three subjects who presented mild AD, one moderate, and one severe. Volunteers were required to avoid any treatment, for at least 10 days (topical) or 3 months (systemic), prior to the first visit.

As expected, due to the nature of the pathology, lesions had to be included on diverse body sites: internal and external forearm, thigh, elbow, clavicle, shoulder, calf ankle, wrist, chest, and knee. To maintain some homogeneity, a few areas were excluded (face, palms, soles, folds) as they are known to present morphological specificities. A healthy‐looking area (ADNL) either adjacent to the lesion or on the same location on the other side of the body was included as a control. A control area can thus serve as a control for multiple lesions (adjacent on both sides or adjacent and contralateral). Some lesions (6) have no respective control due to a lack of healthy‐looking adjacent skin. A total of 13 control areas were included at baseline.

The lesions (acute/chronic) and control areas were imaged at the start of the study (baseline visit D00). Then, volunteers applied once a day a topical steroid treatment adapted to lesions and recommendations, prescribed by the patient's dermatologist. Lesions were imaged again after 7 and 14 days of treatment (visits D07 and D14). Due to noncompliance with the treatment in one out of six volunteers (not included in the results at D14) and to the infection of one ADL area, 20 (16 acute/4 chronic) out of 22 ADL areas (16 acute/6 chronic) were included in D07, with the number reducing to 18 (15 acute/3 chronic) at D14. For ADNL, 13 areas were included at D00 compared to 9 at D07 and 8 at D14 also attributable to the challenges in spotting an ADNL area in a healed region, where flare‐ups disappear. The characteristics of each lesion included are summarized in Figure S1 in Supplementary data.

2.2. Clinical scoring

The global severity of the pathology was assessed by the dermatologist investigator with three standard clinical scores for assessing AD. The Eczema Area and Severity Index (EASI) considers the overall extent of the lesions and the severity according to specific clinical signs (erythema, edema, excoriation, and lichenification) on each lesional area. ²¹ The EASI ranges from 0 to 72 and classifies AD severity as 0 = none, 1–7 = mild, 7–21 = moderate, and> 21 = severe. The validated Investigator Global Assessment (vIGA) score was assessed for each lesion using a 5‐point scale (0 = clear, 1 = almost clear, 2 = mild, 3 = moderate, 4 = severe) to describe the overall appearance of the area at a given point. ²² Lesions were also classified as acute or chronic by the dermatologist according to visual observations of the presence or absence of lichenification and medical interrogation. Lesions included in the study did not concern crusty or excoriated lesions.

2.3. LC‐OCT

LC‐OCT images (3D volume blocks, vertical and horizontal cross‐sections) were acquired on the lesions using a DeepLive system (DAMAE MEDICAL, France). ¹⁶ The experimental setup is reviewed in detail elsewhere. ²³ Paraffin oil (with optical index n ∼ 1.4) was used as an immersion medium. The DeepLive device uses a two‐beam interference microscope with a supercontinuum laser as the light source at the central wavelength of 800 nm and a line‐scan camera as the detector. It can acquire images up to a depth of around 500 μm with a lateral resolution of 1.1 μm and an axial resolution of 1.2 μm. The X‐Y field of view of 3D stacks acquired by the DeepLive device is 1.2 mm × 0.5 mm. The DeepLive system also acquires dermatoscopic pictures of the imaged areas simultaneously with LC‐OCT images.

2.4. Quantification of epidermal morphology using AI‐based algorithms

Deep learning AI‐based algorithms based on 3D convolutions ¹⁹ and the 3D StarDist model ²⁴ described in detail elsewhere ²⁰ were used to quantify the epidermal morphology at different scales. The algorithm was trained using manually annotated 2D vertical images labeled by trained experts from an independent dataset. The segmentation outcome was also validated by the trained experts through visual inspection of processed images to ensure accuracy and reliability.

Tissular (histological) metrics were obtained from the skin layers segmentation: ¹⁹ stratum corneum thickness (SC thickness, in μm), viable epidermis thickness (VE thickness, in μm), and dermal–epidermal junction (DEJ undulation, in %). The DEJ undulation is expressed in % as the area of the DEJ layer divided by the area of the horizontal surface.

Cellular metrics were obtained by segmenting the keratinocytes nuclei in the viable epidermis, which appear as black (non‐echogenic) areas, and analyzing their size (KN volume, in μm³) and shape (KN compactness, no unit, ranging from 0 to 1, where 1 is a perfect sphere). ¹⁹ Figure 1 shows a typical example of the tissular, and cellular quantifications obtained.

FIGURE 1.

(A) Macroscopic pictures of ADL and ADNL areas at D00 (ankle, subject 1). (B) Corresponding vertical LC‐OCT cross‐sections (C—D) Corresponding 3D LC‐OCT stacks. Using an AI‐based algorithm, the skin layers have been segmented (green) and in (C) the KN is sorted by size (red, blue, yellow), and in (D) the atypia parameter is calculated. ADL, AD acute and chronic lesions; AI, artificial intelligence; KN, keratinocyte nuclei; LC‐OCT, line‐field confocal optical coherence tomography.

Finally, metrics were obtained to characterize the cellular network. The cell density (KN density, in cell number/mm²) was calculated by dividing the total number of nuclei in the epidermis by the skin surface area. A multiparametric parameter named atypia, highlighting the disorganization of the KN network inside the epidermis, was created by comparing cells to their neighbors in terms of morphology (KN atypia, no unit). A nucleus with low atypia (close to 0) resembles very closely its neighbors, leading overall to a relatively homogeneous epidermis, while a nucleus with high atypia (close to 1) is very different from neighboring cells, resulting in a more chaotic overall organization. Additional details are provided elsewhere. ²⁰

To reflect the maturation process that keratinocytes undergo in the epidermis from the basal layer to the stratum corneum, a local analysis is needed. For this purpose, cellular layers were defined. On each X‐Y coordinate, going downwards from the top of the viable epidermis to the DEJ, a new layer is defined each time a keratinocyte cell is crossed on the vertical axis. This results in non‐flat cellular layers, as expected from a biological point of view. A schematic representation is available in Supplementary data (Figure S2). KN volume, KN compactness, and KN atypia were computed layer by layer, averaging the metric over all nuclei in the layer. KN density is only available on the overall epidermis as a result of the definition of a layer.

2.5. Statistical analysis

In this paper, boxplots are presented, where each point represents one lesion. The bold line represents the median, the diamond the average, and the box limits the first and third quartiles. Where scatter plots are presented, the error bars represent the standard deviation of the mean. For statistical tests, the unpaired Wilcoxon test with a significance level of 5% was used, appropriate for the number of samples considered. Effect sizes were calculated according to Cohen's d formula to quantify the size of the difference between the two groups. Significance is shown in the figures as ***, **, and * for a p‐value < 0.05 with very strong ]2–Inf [, strong ]1.5–2] (effect size was between 1.5 (not included) and 2 (included) based on Cohen's d formula in statistics), and moderate ]0.8–1.5] effect sizes, respectively.

3. RESULTS

The present study was designed to compare AD lesional areas (ADL) with healthy‐looking control areas (ADNL) as well as acute/chronic lesional areas through histological and cellular‐level metrics. Furthermore, the effect of a topically applied steroid treatment over the course of 2 weeks, from baseline (D00) to visits at day 7 (D07) and day 14 (D14) was also investigated.

Clinically, the lesions healed under treatment as expected, with a vIGA score reaching 0 (clear) for 14 lesions and 1 (almost clear) for 4 lesions by day 14 (Figure S3).

All metrics are presented below merging all areas included in the study, regardless of patient, anatomical zone, severity, etc. Notably, all conclusions remain valid when considering each patient alone, albeit with less statistical power due to the reduced number of data points.

Figure 1 presents a comparative analysis between a lesional area (ADL—left) and an adjacent healthy‐looking area (ADNL—right) at baseline. It encompasses three distinct perspectives: macroscopic images (Figure 1A), vertical LC‐OCT cross‐sections (Figure 1B), and 3D LC‐OCT stacks featuring AI‐generated visualizations (Figure 1C,1D). In Figure 1C, the layer segmentation and cell nuclei sorting by volume are illustrated. The green segments, arranged from top to bottom, represent the skin surface, the delineation of the living epidermis with the stratum corneum (SC), and the DEJ. KN are color‐coded according to their volume: red for small nuclei (30% smallest), blue for intermediate (40% intermediate), and yellow for large nuclei (30% largest). Figure 1D depicts nuclei atypia using a color scale ranging from 0, indicating a nucleus closely resembling its neighbors, to 1, denoting a nucleus significantly different from its surrounding counterparts.

FIGURE 3.

Layer‐by‐layer analysis of KN volume, compactness, and atypia for ADL and ADNL areas, from the top of the viable epidermis (layer 0) to deeper levels. The three parameters are represented at baseline (left) and after one (middle) and 2 weeks (right) of treatment. The stars indicate the effect size (*** very strong, ** strong, *moderate,—weak) when the p‐value is significant (< 5%). ADL, AD acute and chronic lesions; KN, keratinocyte nuclei.

3.1. Histological metrics: viable epidermis thickness, SC thickness, DEJ undulation, and cell density

An AI‐based algorithm was used to extract classical morphological metrics on the 3D LC‐OCT images: viable epidermis thickness, SC thickness, and DEJ undulation. Figure 2 shows the quantification of these histological metrics for ADL and ADNL areas at D00, D07, and D14.

FIGURE 2.

Histological metrics (viable epidermis thickness, SC thickness, and DEJ undulation) at baseline and after 7 and 14 days of topical steroids treatment for lesional skin (ADL) and control healthy‐looking areas (ADNL). The stars indicate the effect size (*** very strong, ** strong, *moderate) whenever the p‐value is significant (p‐value < 5%). ADL, AD acute and chronic lesions; DEJ, dermal–epidermal junction; SC, stratum corneum.

Despite the diversity of the areas (in terms of patient, anatomical area, and severity) and the corresponding variability, clear and statistically significant differences were observed in all layer metrics. At D00, the epidermis was significantly thicker in ADL areas compared to ADNL areas: the mean thickness was about twice as high (140 ± 27 μm for ADL areas, 63 ± 11 μm for ADNL areas). The stratum corneum was also thicker in lesional skin (24 ± 6 μm for ADL areas, 16 ± 4 μm for ADNL areas). Of note, the DEJ was significantly more undulated at D00 in the ADL area with a mean value of 85 ± 43 % compared to 13 ± 9 % in ADNL areas. A higher heterogeneity was seen in all metrics in lesional skin, as illustrated by the higher standard deviations.

The lesions were imaged again seven (D07) and fourteen (D14) days after the start of treatment with topical steroids. Figure S4 shows the typical evolution of an ADL area upon treatment (3D stacks). Lesional skin returned to a normal clinical appearance, and accordingly, the histological metrics evolved in the direction of non‐lesional skin values, as shown in Figure 2. The SC became more homogenous and thinner, as did the viable epidermis, while the DEJ flattened. The viable epidermis thinned down mostly in the first week of treatment (‐32% at D07 vs. D00) and stayed almost stable thereafter (‐11% at D14 vs. D07), while the SC thickness did not evolve much in the beginning (‐6% at D07 vs. D00) and then thinned down (‐22% at D14 vs. D07). The DEJ flattened on a timeline similar to the viable epidermis thinning: ‐43% in the first week, and ‐9% between D14 and D07. Interestingly, at D14, the differences between ADL and ADNL areas were still statistically significant on the viable epidermis thickness and DEJ undulation, despite the clinical score returning to normal.

Figure S5 shows the evolution of the histological metrics on acute and chronic lesions. No statistical differences were observed in the viable epidermis thickness and DEJ undulation. Only the SC was found to be significantly thicker in chronic lesions when compared to acute lesions at day 14.

3.2. Cellular‐level metrics: keratinocytes nuclei volume, compactness, cellular network atypia

Epidermal cells are recognizable in LC‐OCT images through their nuclei, which appear as small black, non‐echogenic dots. AI‐based algorithms were used to segment the KN and quantify their shape, in terms of size (KN volume) and sphericity (KN compactness). Additionally, a parameter called atypia (unitless, ranging from 0 to 1) was created by comparing the shape and size of nuclei with neighboring cells to reflect the heterogeneity of the epidermal layer.

A layer‐by‐layer analysis was performed, considering each cellular layer individually (as opposed to a slice of epidermis of arbitrary thickness), as shown in Figure 3. Due to the thicker epidermis in ADL areas, the corresponding layers in ADL and ADNL do not necessarily represent the same time point in the process of keratinocyte maturation. Comparisons are more accurate for the layers closest to the surface, where the maturation of the keratinocytes is known to be almost completed. Additionally, the results on the deepest layers (10 and downwards) tend to be less interpretable due to missing data from either the absence of material (more layers were detected in the thicker ADL areas epidermis than in the thinner ADNL areas epidermis) or lack of signal (laser scattering from the thicker epidermis).

In healthy‐looking ADNL areas, the keratinocyte's nuclei size and shape were observed to change as they underwent their maturation from the stratum basale to the stratum granulosum. KN start small and spherical in shape in the basal layer and flatten and increase in volume as they ascend into the viable epidermis. This was already described in the literature for healthy skin, ¹⁹ but has never been investigated this finely using a layer‐by‐layer approach. In healthy‐looking skin, we report that the maturation‐related parameters, that is, the KN shape and size, seem to change drastically in the very last few (3 to 4) layers of cells before the stratum corneum, while they plateau at similar values for least seven layers underneath (below this, the signal becomes too weak to segment the nuclei).

In atopic lesions, a more “chaotic” differentiation can be seen: the distribution of KN size is very heterogeneous in the 3D stack, as shown with colors in Figure 1, with large nuclei observable near the bottom layer of the epidermis and small ones close to the surface. Additionally, the plateau effect observed in ADNL areas on cellular layers 3 to 9 is less pronounced in ADL areas, where the gradient of maturation of the KN seems to be more continuous from layers 0 onwards.

At D00, the keratinocytes nuclei were consistently larger in ADL compared to ADNL areas on each layer, until layer 9 (Figure 3,3A). KN were less spherical for shallow layers 1 to 4 in ADL compared to ADNL areas at D00 while for deeper layers 8 and 9, KN in ADL areas were observed to be more spherical than in ADNL areas (Figure 3,3B). The atypia of KN (resemblance to close neighbors) was much larger at baseline in ADL zones compared to ADNL areas in all layers, with a peak value for the middle layer 4.

Upon treatment, a normalization was observed on all parameter's depth distribution. At D07, some significant differences still appear in the shallowest layers, while the deepest layers were statistically equivalent to healthy‐looking adjacent areas. At 14 days, when the lesions were clinically consistent with healthy‐looking skin, all cellular parameters had normalized on all layers including the shallowest.

Finally, despite the small number of chronic lesions included (6 at baseline), statistically significant differences were found in the KN atypia when comparing acute and chronic lesions, as shown in Figure 4. At D00, chronic ADL areas displayed a much higher atypia in shallow layers than acute ADL areas. At D07, all lesions improved, as evidenced by the KN atypia moving in the direction of ADNL values on all layers, but while the acute ADL areas were no longer significantly different from ADNL areas, chronic lesions still had higher atypia, in particular in the deepest layers. At D14, the number of chronic lesions no longer allowed for a robust statistical analysis.

FIGURE 4.

KN atypia for ADL and ADNL at different layers of the epidermis from top (0) to bottom (12). Atypia is represented at D00 (acute N = 16/chronic N = 6), D07 (acute N = 16/chronic N = 4), and D14 (acute N = 15/chronic N = 3), after the beginning of topical steroid treatment. Statistics compared acute versus chronic lesions at D00 and D07. At D14, the number of chronic lesions no longer allowed for a robust conclusion. The stars indicate the effect size (*** very strong, ** strong, *moderate) whenever the p‐value is significant (p‐value < 5%). ADL, AD acute and chronic lesions; KN, keratinocyte nuclei.

For KN density, the variation between acute/chronic remained relatively small (∼ 8000 cell/mm²) compared for instance to variations observed between anatomical locations in previous works ¹⁹ (e.g., ∼ 15 000 cell/mm² between the cheek and the forehead). Hence, the variations are considered to be of little biological relevance.

4. DISCUSSION

In the present study, LC‐OCT imaging was used to observe in vivo and noninvasively the skin of AD patients during eczema flare‐ups and follow the remissions of the flare‐ups under topical steroid treatment after 1 and 2 weeks. AI‐based algorithms were used to extract quantitative parameters from the images, including morphological markers of overall architecture (layers thickness and undulation) and novel cellular‐level markers, such as cell nuclei volume, shape, and overall homogeneity.

Compared to adjacent healthy‐looking skin, AD lesions had a thicker epidermis, thicker SC, and more undulated DEJ, while KN was bigger and less compact. The epidermis maturation overall was more chaotic and less homogeneous, as shown through the new atypia parameter. Regarding acute/chronic lesions, the SC thinned down faster for acute lesions than for chronic lesions. This is correlated with vIGA scores (Figure S3), which never reached 0 for chronic lesions, suggesting a slower recovery for chronic lesions.

Under treatment, all parameters normalized closer to the level of healthy‐looking skin control areas under 2 weeks, as the clinical scores of the lesions normalized as well.

The results on morphological markers are in accordance with what is known in the literature on AD lesions when performing histological cuts on skin biopsies. ⁹ , ²⁵ The LC‐OCT combined with AI‐based algorithms add a new dimension to these quantifications by resolving most of the issues encountered with histology: the measurements are noninvasive, painless, leave no scar, can be performed at the same location several times, including during flare‐up remission, and yield more precise quantifications. Some markers, that is, the epidermis thickness and DEJ undulation, do not reach normal levels by the 2‐week follow‐up visit, despite the clinical scores normalizing, making it a subclinical level marker of lesional skin. It remains to be investigated whether they would finally reach healthy‐looking skin levels after a longer time, or whether they would stabilize at this slightly (but statistically significantly) elevated level. The latter could mean a “risk factor” for flare‐up re‐occurrence, consistent with the relapse of flare‐ups at the same locations observed clinically in AD.

In addition to tissular “histologic” metrics, the LC‐OCT technique gives access to novel cellular‐level parameters: the KN volume, compactness, and resemblance to close neighbors (through the “atypia” parameter), which can be calculated very finely in a layer‐by‐layer approach, and overall epidermal cell density.

In healthy skin, KN typically becomes larger and less spherical as they mature, as shown in previous studies ¹⁹ and observed here in the ADNL areas. However, this is the first time that the maturation process is followed this finely through a layer‐by‐layer approach, using the KN volume and compactness as markers of maturation. In the healthy‐looking skin from AD patients, the quantification of the KN volume and compactness revealed that the cells maturation is not a continuous gradient in the viable epidermis, but rather than the cells undergo a plateau of maturation for a few layers in the epidermis, while in the last, shallowest, layers of the viable epidermis, they mature very quickly. We believe that this layer‐by‐layer analysis might highlight the difference between the stratum granulosum and the stratum spinosum at the nucleus level, as it is already well‐known that keratinocyte cells are flatter and wider in the stratum graulosum. ²⁶ This result will need to be validated for healthy skin in volunteers not affected by AD.

The KN was found to be modified in AD lesional areas compared to the control healthy‐looking skin. They tended to be larger in size in all layers quantified, and less spherical, in particular in the very last few layers of the living epidermis. While the KN is enlarged in lesional areas, the mean compactness does not evolve accordingly, translating into an abnormal maturation process.

This is even more clearly evidenced when considering the atypia parameter, which expresses the local variability rather than the mean value. The atypia was very significantly increased in lesional areas, which presents a chaotic keratinocyte maturation process. This is the cellular‐level result of the altered expression of keratinocyte proliferation and differentiation‐related genes previously observed in AD lesional areas. ²⁷

Atypia is a multi‐parametric parameter developed using Deep Learning in a successful attempt to discriminate automatically between healthy and pathological skin in actinic keratosis, a type of precancer skin. ²⁰ Although the underlying molecular mechanisms are completely different, we can see that in both cases, atypia translates into anomalies in the epidermal differentiation process. Recent findings indicate a slight increase in atypia during the aging process, highlighting a modified maturation process. ²⁸ This confirms atypia as a very interesting novel parameter to appreciate the epidermal maturation process efficiency. Several other pathological conditions could be explored using atypia as a marker of epidermal maturation uniformity, as they are known to have at least some element of abnormal cellular differentiation.

The effect of treatments, whether dermatological, cosmetic, or surgical, should also be measurable using this parameter, as shown here for the first time with the example of topical steroids in AD lesions. Our study shows again the advantages of using noninvasive techniques, as the same areas can be investigated at three different time points, which gives a very precise idea of the timeline of lesional areas' healing. The epidermal thickness, DEJ undulation, and cellular parameters in the deeper layers (layer 5 and onwards, below the stratum granulosum) were modified mostly in the first week of treatment, getting closer (but not equal) to normal healthy‐looking skin levels. In contrast, the SC thickness and cellular markers in the very last layers changed mostly between weeks 1 and 2 of treatment, at which point the clinical scores reached normal levels. This is consistent with the propagation of the effect of the topical steroid treatment from the deeper layers, where it is active, to the shallower levels, where it is finally visible, along with the epidermal differentiation process. The timeline could be refined in future studies by imaging the lesions at more frequent intervals.

Both acute and chronic lesions were included in the study. Chronic lesions are clinically different from acute lesions, but in the present study, they were found to be equivalent on all quantitative parameters at baseline, except for the atypia which was markedly larger in chronic lesions. Over time, chronic lesions healed slower than acute lesions, which again was not shown on any parameters but on the atypia in earlier stages (D07) and the SC thickness in the latter stages (D14). Furthermore, when looking at the KN volume and KN compactness for acute/chronic ADL area (Figure S6), no statistical difference was found, supporting the relevance of atypia, and highlighting the heterogeneity of the KN network (atypia depending on KN volume, compactness and resemblance to neighbors). This suggests that atypia could be useful as a marker for AD lesion type. This will have to be validated on a larger cohort.

In the present study, treatment was maintained until the disappearance of the eczema lesions (vIGA = 0), in accordance with AD treatment recommendations. It is important to note that the results obtained are valid for all patients and all areas concerned, which is an important aspect of a real‐life protocol designed to take account of the diversity of lesions, topography, and management of AD.

Our study underscores the robust efficacy of utilizing LC‐OCT in conjunction with AI‐based algorithms for quantifications in the investigation of AD. Despite the inherent limitations of our study—such as a restricted number of participants, absence of healthy subjects, absence of treatment standardization, and variability in anatomical areas—our findings reveal remarkably significant distinctions at both tissular and cellular levels between lesional and non‐lesional areas. This not only provides valuable insights into the current understanding of AD but also paves the way for discerning more nuanced changes in forthcoming, more extensive studies. In the area of AD, these advancements could potentially address some of the most pressing questions in the field: elucidating the mechanisms preceding a flare‐up in conjunction with external and internal stress factors, aiming to reduce treatment duration or even prevent crises altogether, comprehending the distinctions between pediatric and adult dermatitis, acute and chronic lesions, and lesions in various body areas to characterize the phenotypic signature of AD lesion types. These findings could lead to the personalization of treatment according to patients' characteristics (age, exposome, severity of AD) but also adapted to the lesion type (acute or chronic) and possibly localization. The multiscale approach facilitated by the novel tissular, and cellular levels opens the door to bridging the gap between molecular‐level knowledge and clinical observations, as well as perceived sensations.

5. CONCLUSION

In our study, we utilized LC‐OCT imaging alongside AI‐based algorithms to noninvasively observe AD patients' skin during flare‐ups and monitor remissions under topical steroid treatment. Compared to healthy‐looking skin, AD lesions showed marked differences in epidermal thickness, stratum corneum thickness, and dermal‐epidermal junction undulation. KN in lesional areas were larger and less compact. Under treatment, all parameters normalized toward healthy skin levels within 2 weeks, aligning with clinical improvements. This work highlights the efficacy of a novel parameter, “atypia”, which compares neighboring cell nucleus’ morphology, to quantify the quality of the epidermal renewal process. This parameter shows clear differences between acute/chronic lesions at baseline. Atypia holds promise in the study of other pathologies affecting epidermal renewal and subsequent evaluation of treatment efficacy. Our work paves the way for more extensive investigations to address fundamental questions around AD, offering a multiscale approach to bridge molecular insights with clinical observations.

CONFLICT OF INTEREST STATEMENT

H.L.B., E.R., S.B., G.R., and B.L. are employees at L'Oréal. E.R. and A.S. are dermatologist consultants for L'Oréal Advanced Research in BRC, St Louis Hospital. J‐D.B. received research grants for L'Oréal. M.J. does not possess any conflict of interest.

Supporting information

Supporting Information

Le Blay H, Raynaud E, Bouayadi S, et al. Epidermal renewal during the treatment of atopic dermatitis lesions: A study coupling line‐field confocal optical coherence tomography with artificial intelligence quantifications. Skin Res Technol. 2024;30:e13891. 10.1111/srt.13891

DATA AVAILABILITY STATEMENT

The authors declare that materials, data, and associated protocols are available upon request to H. Le Blay: heiva.leblay@loreal.com.