Novel image analysis technique decodes the physiological information engraved on the stratum corneum
203
Presented by: Takeshi Tohgasaki
Introduction
Stratum corneum (SC) cells are differentiated from basal keratinocytes with congenital and acquired influence. Therefore, a large amount of physiological information is engraved on SC cells, and the morphological characteristics and biomarker proteins of SC cells are used to evaluate skin conditions. However, these methods can be improved by comprehensively analysing the structure of SC cells in more detail. In this study, we established a novel evaluation system for individual skin characteristics and conditions by analysing only one SC image using an artificial intelligence (AI)-based image recognition technology and applying multiple regression analysis to the SC cell image. This system could be used to predict current as well as future skin conditions and cosmetic effects.
Methods
AI-based recognition technology
SC cells were collected from 1512 females using the tape stripping method. The bright-field image of each SC cell was captured using a digital microscope, and each image was annotated with SC cell regions. Raw and annotated images were used for machine learning using instance segmentation. In addition, a numerify program was developed to quantify 27 morphological parameters (individual SC cell area, roundness of the SC cells, individual cell intensity, etc.) from the AI-recognised regions. Seven biomarker protein expression levels were measured in each SC sample through enzyme-linked immunosorbent assay (ELISA). The raw images and measured biomarker protein levels were machine learned using a convolutional neural network.
Verification of the relationship between SC parameters and physiological skin conditions
We interviewed 507 females about their skin conditions. In addition, we obtained SC cells from the females and measured their physiological skin conditions (transepidermal water loss (TEWL), SC water content, facial image analysis, etc.). The bright-field images of the SC cells were captured, and 27 morphological parameters and biomarker protein levels were predicted using the AI system. These parameters and biomarker protein levels were used as explanatory variables in multiple regression analysis to predict physiological skin conditions and interview responses. The study protocol conformed to the ethical guidelines of the Ethics Committee of FANCL Corporation, and it was conducted in accordance with the principles embodied in the Declaration of Helsinki.
Results
First, we established AI system that recognize the SC cell regions and predicts biomarker protein levels from bright-field SC images by machine learning. The matching rate between the AI-recognised and annotated regions of the SC cells was 74.1%–94.5% and that between the number of individual SC cells and nucleated cells was 43.8%–44.4%. In addition, there was a significant correlation between the biomarker protein levels predicted using AI and those measured via ELISA.
Next, skin physiological indicators were estimated using multiple regression with the morphological parameters and biomarker protein levels. Significant correlations were confirmed between the estimated and measured values of the TEWL, SC water content, and facial image analysis parameters (brown spot content, pore count, porphyrin count, etc.). Moreover, there were significant correlations between the predicted and measured TEWL and SC water content after 1 month in the presence or absence of cosmetic material. Furthermore, the AI system predicted the correct answers in 70% of the instances regarding the presence or absence of rough skin caused by cosmetics.
Discussion and Conclusion
We developed an AI system to recognise SC cell regions and quantify the morphological parameters of SC cells. Correlations were observed between the AI-predicted and measured biomarker protein levels. Multiple regression analysis was performed using these numerical values as explanatory variables, which were correlated with multiple skin physiological indicators. Our method has the potential to predict skin conditions and cosmetic effects. It would be useful for cosmetic selection, prevention of skin problems, and optimisation of treatment for individuals. The structural recognition accuracy and estimation technology can be further improved by adding more data. Furthermore, additional studies on the explanatory variables that strongly contribute to the physiological state of the skin may be useful for elucidating the metabolic mechanism of the skin. Our findings will contribute to the development of dermatological and cosmetic skin care.
Stratum corneum (SC) cells are differentiated from basal keratinocytes with congenital and acquired influence. Therefore, a large amount of physiological information is engraved on SC cells, and the morphological characteristics and biomarker proteins of SC cells are used to evaluate skin conditions. However, these methods can be improved by comprehensively analysing the structure of SC cells in more detail. In this study, we established a novel evaluation system for individual skin characteristics and conditions by analysing only one SC image using an artificial intelligence (AI)-based image recognition technology and applying multiple regression analysis to the SC cell image. This system could be used to predict current as well as future skin conditions and cosmetic effects.
Methods
AI-based recognition technology
SC cells were collected from 1512 females using the tape stripping method. The bright-field image of each SC cell was captured using a digital microscope, and each image was annotated with SC cell regions. Raw and annotated images were used for machine learning using instance segmentation. In addition, a numerify program was developed to quantify 27 morphological parameters (individual SC cell area, roundness of the SC cells, individual cell intensity, etc.) from the AI-recognised regions. Seven biomarker protein expression levels were measured in each SC sample through enzyme-linked immunosorbent assay (ELISA). The raw images and measured biomarker protein levels were machine learned using a convolutional neural network.
Verification of the relationship between SC parameters and physiological skin conditions
We interviewed 507 females about their skin conditions. In addition, we obtained SC cells from the females and measured their physiological skin conditions (transepidermal water loss (TEWL), SC water content, facial image analysis, etc.). The bright-field images of the SC cells were captured, and 27 morphological parameters and biomarker protein levels were predicted using the AI system. These parameters and biomarker protein levels were used as explanatory variables in multiple regression analysis to predict physiological skin conditions and interview responses. The study protocol conformed to the ethical guidelines of the Ethics Committee of FANCL Corporation, and it was conducted in accordance with the principles embodied in the Declaration of Helsinki.
Results
First, we established AI system that recognize the SC cell regions and predicts biomarker protein levels from bright-field SC images by machine learning. The matching rate between the AI-recognised and annotated regions of the SC cells was 74.1%–94.5% and that between the number of individual SC cells and nucleated cells was 43.8%–44.4%. In addition, there was a significant correlation between the biomarker protein levels predicted using AI and those measured via ELISA.
Next, skin physiological indicators were estimated using multiple regression with the morphological parameters and biomarker protein levels. Significant correlations were confirmed between the estimated and measured values of the TEWL, SC water content, and facial image analysis parameters (brown spot content, pore count, porphyrin count, etc.). Moreover, there were significant correlations between the predicted and measured TEWL and SC water content after 1 month in the presence or absence of cosmetic material. Furthermore, the AI system predicted the correct answers in 70% of the instances regarding the presence or absence of rough skin caused by cosmetics.
Discussion and Conclusion
We developed an AI system to recognise SC cell regions and quantify the morphological parameters of SC cells. Correlations were observed between the AI-predicted and measured biomarker protein levels. Multiple regression analysis was performed using these numerical values as explanatory variables, which were correlated with multiple skin physiological indicators. Our method has the potential to predict skin conditions and cosmetic effects. It would be useful for cosmetic selection, prevention of skin problems, and optimisation of treatment for individuals. The structural recognition accuracy and estimation technology can be further improved by adding more data. Furthermore, additional studies on the explanatory variables that strongly contribute to the physiological state of the skin may be useful for elucidating the metabolic mechanism of the skin. Our findings will contribute to the development of dermatological and cosmetic skin care.