Age‐dependent changes in epidermal architecture explored using an automated image analysis algorithm on in vivo reflectance confocal microscopy images

Abstract

Background

Reflectance confocal microscopy (RCM) allows for real‐time in vivo visualization of the epidermis at the cellular level noninvasively. Parameters relating to tissue architecture can be extracted from RCM images, however, analysis of such images requires manual identification of cells to derive these parameters, which can be time‐consuming and subject to human error, highlighting the need for an automated cell identification method.

Methods

First, the region‐of‐interest (ROI) containing cells needs to be identified, followed by the identification of individual cells within the ROI. To perform this task, we use successive applications of Sato and Gabor filters. The final step is post‐processing improvement of cell detection and removal of size outliers. The proposed algorithm is evaluated on manually annotated real data. It is then applied to 5345 images to study the evolution of epidermal architecture in children and adults. The images were acquired on the volar forearm of healthy children (3 months to 10 years) and women (25–80 years), and on the volar forearm and cheek of women (40–80 years). Following the identification of cell locations, parameters such as cell area, cell perimeter, and cell density are calculated, as well as the probability distribution of the number of nearest neighbors per cell. The thicknesses of the Stratum Corneum and supra‐papillary epidermis are also calculated using a hybrid deep‐learning method.

Results

Epidermal keratinocytes are significantly larger (area and perimeter) in the granular layer than in the spinous layer and they get progressively larger with a child's age. Skin continues to mature dynamically during adulthood, as keratinocyte size continues to increase with age on both the cheeks and volar forearm, but the topology and cell aspect ratio remain unchanged across different epidermal layers, body sites, and age. Stratum Corneum and supra‐papillary epidermis thicknesses increase with age, at a faster rate in children than in adults.

Conclusions

The proposed methodology can be applied to large datasets to automate image analysis and the calculation of parameters relevant to skin physiology. These data validate the dynamic nature of skin maturation during childhood and skin aging in adulthood.

1. INTRODUCTION

Reflectance confocal microscopy ¹ (RCM) is an in vivo imaging technology that enables the visualization of the epidermis and upper layers of the papillary dermis at the cellular level. It provides real‐time optical sections of the epidermis at sequential depth, thus giving information about the 3D structure of the observed area. The energy of the incident light does not initiate any photobiological process in the skin, hence, allowing the observation of cells in situ without a change in their structure or function. RCM is a great alternative to invasive methods, such as biopsies, since it allows repeated sampling of the same area without damage to the skin. It can, therefore, be used to study continuous dynamic changes in skin structure, for example, during epidermal maturation and aging.

RCM images are currently analyzed manually, providing a qualitative description of epidermal structure, but a quantitative analysis of these images would provide more information on skin structure and its impact on skin physiology. ² An important first step in the quantitative investigation of epidermal cell organization in space is keratinocyte identification. Unfortunately, in RCM images, this is done manually and is tedious, time‐consuming, and subject to expert interpretation. Using an automated approach to obtain keratinocyte position would streamline RCM image analysis, thus, helping in unlocking insights faster, reducing study lead time while decreasing error risk and bias in cell identification and image analysis. This would also help to investigate the correlation between changes in epidermal structure and physiological and functional changes in both healthy ³ , ⁴ and diseased skin. ⁵ , ⁶ , ⁷

We recently proposed a method for automating cell identification on RCM images ⁸ based on keratinocyte morphological features. The proposed approach is a three‐step pipeline to detect keratinocytes on RCM images of the Stratum Granulosum (SG) and Stratum Spinosum (SS). Quantitative parameters can then be computed and used to study epidermal architecture at the cellular level.

To date, attempts at automating cell identification on RCM images have been reported only on a limited number of images. ⁹ , ¹⁰ In the current work, we attempt to apply the proposed automated algorithm at large scale to analyze RCM images of healthy skin and examine the effects of age (0–80 years), body site location (cheek and volar forearm), and epidermal layer (SG and SS) on the calculated cell geometry and topology parameters.

2. MATERIALS AND METHODS

2.1. Clinical studies

In vivo RCM images were acquired on the volar forearm of 40 healthy children (3 months–10 years), and of 20 adults (25–40 years), and on the volar forearm and cheek of 80 women (40–80 years). All participants have minimally pigmented skin, with Fitzpatrick types between I and IV. Inclusion criteria required that the participants be in good health, have no history of skin disease, and have not applied any products on the observed area on the day of the study. The study was initiated following approval from an independent institutional review board (studies 19.0198 and 20.0022) and in accordance with the Declaration of Helsinki. Subjects or their guardians gave written informed consent prior to participation to the study. Study participants were divided into 11 age groups: 0–1, 1–2, 2–4, 4–6, 6–8, 8–10, 20–40, 40–50, 50–60, 60–70, and 70–80 years of age (Table 1). Age group 1–2 years old was removed from the subsequent analyses due to insufficient data size (only four participants with usable data).

TABLE 1.

Study participants repartition per age group.

Age group (yrs.)	0–1	2–4	4–6	6–8	8–10	20–40	40—50	50–60	60–70	70–80
Number of participants	8	10	9	10	10	20	15	15	15	15

Images were captured using a Vivascope 1500 (Lucid Inc., Rochester, New York) reflectance confocal microscope with a z‐optical resolution of 5 μm and xy‐optical resolution of 1 μm. Images started at the Stratum Corneum (SC) and progressed down towards the Stratum Basale (SB) and the first layers of the papillary dermis. The image size was 1000 by 1000 pixels, with a resolution of 1 μm² per pixel.

2.2. Image analysis workflow

First, RCM images in each acquired image stack were classified into one of six categories: outside of skin, SC, SG, SS, SB, and dermis. This was done using a hybrid deep learning algorithm ¹¹ trained on 1500 images, which resulted in a test accuracy of 82%. Image classification was corrected manually. Subsequent analysis was focused only on images of the SG and SS, in which individual viable keratinocytes could be observed and were characterized by a grainy cytoplasm surrounded by bright, grainy membranes, forming a honeycomb pattern. ¹² , ¹³ Further image analysis was conducted using a three‐step approach ⁸ based on the intensity and morphological features (membrane thickness, length, and keratinocyte size) of keratinocytes visible in RCM images (Figure 1). The first step of the image analysis workflow was to identify and separate the regions of interest (ROI) of tissue containing keratinocytes from the dark background (microrelief lines). This is achieved using the Morphological Geodesic Active Contour method. ¹⁴ We further refine the ROI area by removing noninformative bright areas sometimes present in RCM images by applying a sequence of morphological closings and openings to the binarized image. Other spurious areas, due to low contrast and a drop in signal‐to‐noise ratio, were removed using a Support Vector Machine algorithm trained on Grey Level co‐occurrence matrix features: homogeneity, contrast, dissimilarity, and energy. ¹⁵ The second step was to segment individual cells in the identified tissue area. To perform this task, we used successive application of Sato ¹⁶ and Gabor filters, which highlight tube‐like structures in the image such as cell membranes. The outputs of these filters are then locally normalized, binarized, and skeletonized. The obtained skeletons (representing cell membranes) are pruned to remove spurious branches. The third step was image postprocessing to improve cell detection and remove size outliers, such as detected areas that were too small or too large to represent cells.

FIGURE 1.

Diagram of the image analysis workflow used for keratinocytes detection. The sections are color coded as follows: blue, ROI identification; pink, individual cells identification within the ROI; teal, post‐processing steps. ROI: region of interest.

2.3. Statistical analysis

The population median value was used to express characteristics in each age group, epidermal layer, and body site. Quantitative variables were compared using ANOVA on a fitted linear model. ¹⁷ A linear regression with respect to age was fitted to all parameters per body site and epidermal layer to determine the direction and strength of the relationship between age and each parameter. Statistical significance was considered for a p‐value < 0.05. All statistical analyses were conducted using Python 3.6.7.

3. RESULTS

The image analysis workflow was applied to all images. Cell contours and centers were obtained for each image, and used to calculate cell area, perimeter, density, and number of Delaunay nearest neighbors. ¹⁸

3.1. Cell geometry

All geometrical parameters gradually change with age and differ between epidermal layers and body sites. The differences between children and adults are significant for the SG of the volar forearm, but not of the cheeks, and not significant for the SS for either of the two body sites tested. Cells are significantly larger in the volar forearm compared to the cheeks (Figure 2). On the volar forearm, the gradual change in cell area with respect to age is significant for the SG (R ² = 0.465), but not for the SS (Figure 3A). On the cheek, there is no significant correlation for either layer (Figure 3B).

FIGURE 2.

Median cell area ± standard error of mean per age group, epidermal layer, and body site reflect the dynamic maturation and ageing of the epidermis. * indicates that the median cell area is significantly different between the SG and the SS for a body site and age group; # indicates that the median cell area is significantly different between the cheek and the volar forearm per age group and epidermal layer. SG, stratum granulosum; SS, stratum spinosum.

FIGURE 3.

Median cell area per participant on (A) the volar forearm (SG R ² = 0.465; SS R ² = 0.00693) and (B) the cheeks (SG R ² = 0.0324; SS R ² = 0.0708) colored by age group. We fit a linear regression for each epidermal layer and body site but plot it only when significant. SG, stratum granulosum; SS, stratum spinosum.

SC thickness and suprapapillary epidermis (SPE) thickness on the volar forearm were measured by applying a hybrid‐deep learning method ¹¹ to each stack of RCM images and calculating the depth difference of the uppermost and the lowest optical sections that contain the desired structures, and averaging them per participant (Figure 4). Both SC and SPE thickness increase significantly during childhood (R ² = 0.187 for SC and R ² = 0.279 for SPE) but not in adults.

FIGURE 4.

(A) Median SC thickness per participant (children R ² = 0.187), (B) Median SPE thickness per participant (children R ² = 0.279). A linear correlation with age was not significant in the adult group for either the SC or the SPE thickness. SC, stratum corneum; SPE, suprapapillary epidermis.

These findings suggest that cell turnover is faster on the face than on the arm and is faster in children than in adults ⁴ , ¹⁹ and is associated with a higher keratinocytes proliferation rate, ²⁰ leading to a thickening of epidermal layers and an increase in cell size. These quantitative geometrical measures are in agreement with current physiological knowledge about epidermal maturation. Indeed, even though epidermal structure and barrier function is competent at birth and during childhood, the epidermis is more susceptible to outside‐in (penetration of noxious substances ⁴ , ²¹ , ²² , ²³ due to lower glyph density) and inside‐out aggressors (water evaporation leading to tissue desiccation ²⁴ due to higher transepidermal water loss, higher conductance, and lower natural moisturizing factors in infants ³ , ²³ ).

Interestingly, while cells are generated in the basal layer and grow larger as they climb toward the surface through the spinous and then the granular layer, their overall shape and relative dimensions do not vary. Indeed, the cell shape aspect ratio ($\frac{Majoraxislength}{Minoraxislength}$) remains the same through all epidermal layers, body sites, and age groups (Figure 5).

FIGURE 5.

Median cell aspect ratio per participant. SG, stratum granulosum; SS, stratum spinosum.

3.2. Cell topology

Using the detected cell centers, a Delaunay triangulation was built for each image and used to calculate the average probability distribution of the Delaunay nearest neighbors per age group, epidermal layer, and body site (Figure 6 and Supporting Information Figure S1). While geometrical parameters change with age, body site, and epidermal layer, structural organization does not and seems to be preserved through epidermal maturation and ageing in healthy skin.

FIGURE 6.

Average probability distribution of the number of the Delaunay nearest neighbors per cell for each age group, in the stratum granulosum on the volar forearm.

The Fisher–Pearson coefficient of skewness ²⁵ of the probability distribution was calculated for each age group, epidermal layer, and body site, as a measure of their asymmetry, and the obtained values were compared to those of two previously published models based on game theory ²⁶ : a cooperators model representing healthy epithelium and a defectors model representing cancerous tissue (Figure 7 and Supporting Information Figure S2). The obtained skewness values (Supporting Information Table S1) are closer to that of the cooperators model (skewness = 0.44, the data is symmetrical) and much lower than that of the defectors model (skewness = 0.98, greatly skewed data). There are no statistically significant differences between skewness values of the SG and SS distributions on both the cheeks and the volar forearm.

FIGURE 7.

Distribution of skewness per age group for (A) SG on the volar forearm, (B) SS on the volar forearm, (C) SG on the cheeks, and (D) SS on the cheek. In red, the skewness of the probability distribution of the Delaunay nearest neighbors for the defectors model. In green, the skewness of the probability distribution of the Delaunay nearest neighbors for the cooperators model. SG, stratum granulosum; SS, stratum spinosum.

4. DISCUSSION

Results from this study demonstrate that the epidermis undergoes constant changes through its maturation and ageing. This work is in accordance with prior published work ²⁷ and extends it by studying a broader age range across thousands of images. Using a novel approach for automated keratinocyte detection on RCM images ⁸ allowed us to quantitatively study epidermal cells spatial organization in different settings.

Automating a step that was until now done manually and required long and tedious work, allowed us to analyze more RCM images than previously possible, therefore, cementing our prior knowledge on the epidermis structural maturation. It also enabled us to extend the study across a larger age range, from birth to 80 years, with a broad sampling across ages, when previous studies focused on differences between infants and adults. ³ , ⁴ , ¹⁹ , ²¹ , ²² , ²³

The results of this study are consistent with prior work on epidermal maturation. Likely due to the slowing down of cell turnover with age, cells are larger in adults than in children, and bigger in the granular layer than in the spinous layer. Reduced cell turnover allows for a longer residence time of cells in the epidermis, resulting in a longer time to mature and grow larger before they get desquamated at the top of the SC. Furthermore, we observed that cells on the volar forearm are larger than on the cheeks, which also reflects higher cell turnover rates on the face.

Additionally, while healthy skin cells organization in space dynamically changes, their spatial arrangement, that is, their topology, remains the same. Skewness of the probability distribution of the Delaunay nearest neighbors is higher in the spinous layer than in the granular layer, and closer to the skewness observed in defectors model. Although these differences are not statistically significant, we could hypothesize that the spinous layer is closer to the basal layer where keratinocytes appear, and thus, the probability distribution of the Delaunay nearest neighbors would reflect this proliferation process and be closer to the defectors model where cancerous cells adopt a cheating competitive (in game theory) strategy and enter an abnormal proliferative state. However, additional analysis of both the spinous and basal layers is required to validate this hypothesis.

Although granular and spinous keratinocyte size vary, their shape is preserved throughout epidermal maturation and ageing. Similar analysis on proliferating basal cells would be of interest, as cellular aspect ratio has only been studied in proliferative or cancer tissues, ²⁸ , ²⁹ , ³⁰ where it was described as a key parameter in normal cell division and function, and as a potential crucial factor in the preservation of cell geometry in proliferative tissues and in the determination of the spatial patterns of daughter cells during cell movement and differentiation.

This study only included Caucasians with Fitzpatrick skin types from I to IV. It would be valuable to extend this work to include participants with different genetic background including more pigmented skin.

In summary, we have shown that an automated method for keratinocytes detection on RCM images can be used to extract geometrical and topological properties which can then be used to compare between age groups, epidermal layers, and body sites. This allowed us to bypass a frequent issue in biomedical image analysis: the tediousness of such a task and its proneness to human interpretation. Using the proposed method on RCM images of diseased skin could provide new knowledge about different skin conditions.

CONFLICT OF INTEREST STATEMENT

The authors declare no conflict of interest.

DISCLOSURES

IL and GNS are employees of Johnson & Johnson Santé Beauté France, a manufacturer of skin cosmetic products.

Supporting information

Supporting Information

Lboukili I, Stamatas GN, Descombes X. Age‐dependent changes in epidermal architecture explored using an automated image analysis algorithm on in vivo reflectance confocal microscopy images. Skin Res Technol. 2023;29:e13343. 10.1111/srt.13343

DATA AVAILABILITY STATEMENT

The code, data, and materials presented in this paper are not publicly available. Requests for the code and data will be considered if received by the authors.; CA : Imane Lboukili Email: ilbouki1@its.jnj.com