Artificial Intelligence in Cosmetic Dermatology A

This article has been accepted for publication in IEEE Access.
This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2023.3295001
Date of publication xxxx 00, 0000, date of current version xxxx 00, 0000.
Digital Object Identifier 10.1109/ACCESS.2017.DOI
Artificial Intelligence in Cosmetic

Dermatology: A Systematic Literature
Review
PAT VATIWUTIPONG1 , SIRAWICH VACHMANUS1 , THANAPON NORASET1 , AND
SUPPAWONG TUAROB.1
1
Faculty of Information and Communication Technology, Mahidol University, Nakhon Pathom 73170 Thailand
Corresponding author: Suppawong Tuarob (e-mail: suppawong.tua@mahidol.edu).
This research project is supported by Mahidol University (Fundamental Fund: fiscal year 2023 by National Science Research and
Innovation Fund (NSRF)).
ABSTRACT
Over the last ten years, the field of dermatology has experienced significant advancements through the
utilization of artificial intelligence (AI) technologies. The adoption of such technologies is multifaceted,
encompassing tasks such as screening, diagnosis, treatment, and prediction of treatment outcomes. The
majority of prior systematic reviews in this domain were centered on medical dermatology, with the aim
of detecting and managing serious skin diseases such as skin cancer. However, the adoption of AI in
cosmetic dermatology, which focuses on improving skin conditions for cosmetic purposes, has not been
comprehensively reviewed. Therefore, the objective of this systematic review article is to analyze the
existing and recent research revolving around applications of AI in the field of cosmetic dermatology.
The study encompasses articles published between 2018 and 2023, where a total of 63 publications are
deemed relevant based on the established inclusion criteria, divided into five categories based on utilization
domains, namely cosmetic product development, skin assessment, skin condition diagnosis, treatment
recommendation, and treatment outcome prediction. This systematic review article provides not only
valuable insights for researchers interested in exploring new research areas related to aesthetic medicine
but also applicable guidance for practitioners seeking to implement AI technologies to address real-world
challenges in cosmetic services.
INDEX TERMS Artificial Intelligence, Machine Learning, Deep Learning, Computer Vision, Cosmetic
Dermatology, Sensitization Testing, Skin Condition Diagnosis, Skin Assessment, Treatment Recommenda-
tion
I. INTRODUCTION revolution of the 19th century and has continued to undergo

further development in the present day [2].
Dermatology is a medical subspecialty that focuses on the
scientific investigation, diagnosis, treatment, and prevention Artificial intelligence (AI) pertains to the emulation of
of disorders affecting the integumentary system, including human intelligence in machines that are programmed to
skin, hair, and nails. Dermatological conditions are varied simulate human thought processes and behaviors [3]. AI
in terms of causes, severity, and symptoms [1]. Despite the technologies have been adopted in medicine to assist repet-
fact that dermatological disorders have been a longstand- itive tasks that rely on human experts, such as screening,
ing concern for humans, it was only in the 18th and 19th diagnosis, treatment, and analyses in epidemiology [4]. Ma-
centuries that skin disorders were investigated through a chine learning (ML) is a subfield of AI that involves the
broader medical lens. The progress of dermatology during development of algorithms and statistical models that en-
that period was concurrent with the advancements made in able machines to learn from data without being explicitly
the field of science. The field of dermatology experienced programmed. In essence, it emulates the cognitive processes
a significant surge in growth subsequent to the scientific of human learning by utilizing experiential data to inform
VOLUME 4, 2016 1
This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License. For more information, see https://creativecommons.org/licenses/by-nc-nd/4
This article has been accepted for publication in IEEE Access. This is the author's version which has not been fully edited and
P. Vatiwutipong et al.: Preparation of Papers for IEEE TRANSACTIONS and JOURNALS
decision-making. The task can be executed under the super- motely and online. The proliferation of teledermatology and
vision of an expert, in a semi-supervised manner, or without self-assessment via smartphone applications can be attributed
any supervision (i.e., unsupervised learning). Recently, the to the restricted availability of dermatologists and advanced
progress in computational hardware technologies has played healthcare services [17]. Furthermore, the exigencies of the
a significant role in the emergence of deep learning (DL) COVID-19 pandemic served as a catalyst for the expedited
as a subfield of machine learning (ML) [5]. DL utilizes implementation of teledermatology, where the utilization of
deep neural network architectures to automatically extract online dermatological consulting was observed to emerge as
features from input data, thereby foregoing the traditional a viable solution during the period of social distancing [18].
domain-expert-dependent feature engineering processes [6]. Considerable research has been carried out regarding the
Numerous studies have indicated that DL exhibits superior utilization of AI within the realm of dermatological disorders.
performance in the field of medicine [7, 8], specifically The majority of prior studies have utilized AI in medical
in dermatology [9, 10] when compared to traditional ML dermatology, specifically in the context of diagnosing and
methods. However, compensating for the absence of guided treating dermatological diseases that, if left untreated, may
feature engineering processes, the superior accuracy of deep have detrimental effects on patients’ well-being or even result
learning is contingent upon the extensive scale of the under- in mortality [19]. Therefore, several review articles have
lying training datasets [11]. Consequently, it is crucial that focused on the utilization of AI and ML methodologies to ad-
deep learning algorithms possess the ability to comprehend dress challenges in medical dermatology. For instance, Wells
patterns in disparate data derived from diverse sources and et al. [20] examined the utilization of AI in dermatopathol-
formats. This is essential not only to guarantee the availabil- ogy. Furthermore, Zhang et al. [21] and Mosquera-Zamudio
ity of adequate training data but also to enable the algorithms et al. [22] conducted a systematic review of scholarly articles
to capture the wide variety of diseases that afflict patients that explore the use of DL to analyze melanoma images.
from different geographical regions and backgrounds [12]. Recently, Jeong et al. [23] analyzed the research trends, find-
According to Borade and Kalbande [13], a significant ings, and constraints pertaining to applying DL in medical
number of dermatologists have relied on conventional diag- dermatology.
nostic techniques in the past, which can be laborious and Cosmetic dermatology is distinct from medical dermatol-
time-consuming. In addition, the field of skincare demands ogy in that it focuses on addressing skin conditions that are
a high level of precision and expertise from many profes- not attributable to illness, including but not limited to wrin-
sions, necessitating specialized knowledge and abilities. For kles, age spots, acne, freckles, and melasma [24]. Although
instance, certain dermatological conditions may present with non-fatal and not posing a direct threat to a patient’s physical
a similar appearance, posing a challenge for even experts in health, these beauty-related dermatological conditions may
their classification. The aforementioned issues necessitate the have psychological implications for individuals, including
utilization of automated procedures that possess the ability to diminished self-esteem and confidence, as well as enduring
furnish dermatologists and relevant healthcare professionals adverse long-term mental effects [25]. In recent years, re-
with the requisite information necessary for their decision- search in cosmetic dermatology has also integrated AI and
making processes. The current trend indicates a significant ML techniques to aid dermatologists in the diagnosis [26, 27]
rise in the adoption of AI and ML methodologies within and prescription of treatment [28], as well as enhancing
the dermatology domain, owing to the vast accumulation of cosmetic product development [29] and engaging with poten-
medical data. These technologies have been utilized as an tial consumer bases [30]. However, to our knowledge, there
assistant to dermatologists for various tasks such as disease has been no systematic review, synthesis, or consolidation
diagnosis, evaluation of the severity of conditions, and devel- of research on the applications of AI and ML in cosmetic
opment of treatment recommendations [14]. Certain studies dermatology. Hence, the objective of this article is to employ
even discovered that AI algorithms exhibit a high level of ac- the standard methodology of systematic literature review
curacy when functioning as clinical assistants, and in certain [31] to collect and compile recent and relevant research that
cases, their accuracy surpassed that of human dermatologists pertains to utilizing AI technologies in tackling challenging
[15]. issues in cosmetic dermatology, with the aim of bridging the
Common obstacles encountered in the field of dermatol- aforementioned contribution gap.
ogy often involve decision-making processes that entail skin The contribution of this systematic review is to examine
or hair pictures, which are frequently presented as computer the utilization of AI and ML techniques in cosmetic derma-
vision tasks that can be addressed by ML techniques. ML tology research, encompassing the entire spectrum of der-
approaches utilized in dermatology have the ability to ac- matological procedures, including the upstream development
quire knowledge from various types of image data, including of cosmetic products, middle-stream diagnosis and treatment
clinical, dermoscopic, histopathological, and self-captured activities carried out by dermatologists, and downstream as-
images. Such ability to intelligently process and extract use- pects focusing on ensuring customer satisfaction. The subse-
ful information from patient or specimen images has proved quent sections of this review paper are structured as follows.
useful not only for clinical dermatology but also teleder- Section II explains the review methodology, which entails
matology [16], where consultant sessions are performed re- the specification of the search terms used to query scholarly
2 VOLUME 4, 2016
documents, research questions, and inclusion and exclusion based on their abstract content, whereby only those papers
criteria. Section III provides a comprehensive representation that specifically address the applications of AI in cosmetic
of the chosen papers from various demographic angles. Sec- dermatology are selected for further review. The articles that
tion IV comparatively discusses the chosen papers, which successfully pass the initial abstract screening procedure are
are divided into categories based on the different stages in subjected to an in-depth review and comparative analysis.
cosmetic dermatological services. Section V sheds light on The scholarly articles reviewed in this study are limited
the relevant potential future challenges and research topics. to journal articles and conference proceedings. This review
Finally, Section VI concludes this systematic review article. excluded posters, abstracts, extended abstracts, review pa-
pers, letters, and preprints. Furthermore, this review en-
II. REVIEW METHODOLOGY compasses AI utilization in cosmetic dermatology, including
This systematic literature review uses a methodology that fol- not only the diagnosis and treatment procedures but also
lows the Preferred Reporting Items for Systematic Reviews all aspects pertaining to cosmetic businesses, ranging from
and Meta-Analyses (PRISMA) method [32]. Specifically, the the development of cosmetic products and measurement of
research questions are first established. Then, a search strat- skin sensitization to the evaluation of customer satisfaction.
egy is determined, including the selection of databases and The scholarly articles being reviewed must necessitate the
search keywords and filtering keyword-matched papers using utilization of either AI or ML in a certain portion of their
inclusion and exclusion criteria. Finally, all selected articles research addressing challenges in cosmetic dermatology. In
are analyzed on their objective, techniques, strengths, and this study, AI is not typically attributed to rule-based decision
limitation. Pertaining to recent literature on AI applications algorithms, classical mathematical models, formulas, or mere
in cosmetic dermatology, three research questions are raised: statistical techniques. Furthermore, the investigation pertain-
(Q1) What are the specific tasks in cosmetic dermatology ing to the implementation of AI in the field of clinical or
where AI approaches are employed? medical dermatology is also deemed ineligible for inclusion.
(Q2) What AI methodologies are employed within the cos- Recall that the differentiation between cosmetic dermatology
metic dermatology domain? and medical dermatology is based on the evaluation of the
(Q3) How well does AI demonstrate proficiency in tasks outcome of the symptom. Medical dermatology encompasses
related to cosmetic dermatology? Could AI effectively conditions that are classified as illnesses or diseases and those
aid dermatologists in their diagnosis and treatment that have the potential to cause mortality, such as malignant
procedures, as well as reduce human involvement in tumors or cancer. If a medical condition pertains to aesthetics
non-clinical cosmetic tasks? or lacks a direct association with well-being or mortality,
such as conditions like acne, wrinkles, melasma, and hair
A. SEARCH STRATEGY loss, it falls under the category of cosmetic dermatology. To
We used the SCOPUS, IEEE Xplore, and PubMed databases summarize, the exclusion criteria utilized are as follows:
as our resources. The following query is used to retrieve the (EC1) not being a full research paper,
initial set of papers: (EC2) not related to dermatology,
(EC3) mainly focused on medical dermatology,
(dermatology OR skin) AND (cosmetic)
(EC3) not using AI approaches in any part of the studies.
AND (artificial intelligence OR machine learning
OR deep learning)
III. SELECTION RESULTS
Duplicated and non-English papers are removed. Further- The number of articles in each screening stage is shown in
more, papers published before 2018 are removed to retain the diagram in Figure 1. This diagram was generated by [33].
only recent papers. The search string was queried on 30 December 2023.
Initially, the SCOPUS database returned 937 articles. The
B. SELECTION METHOD 253 papers that were published before 2018 were removed.
Keyword-matched papers from the preceding stage must un- In the titles and abstracts screening stage, 546 papers that did
dergo a manual check for inclusion in further reviewing pro- not meet our inclusion and exclusion criteria were excluded.
cesses. First, the papers go through a screening process based The full text of all remaining 174 papers was found. After
on their titles, whereby only those papers that are deemed the full texts were screened, we finally obtained 63 articles,
relevant to the fields of AI and dermatology are retained, which were 49 journal articles and 14 conference papers.
while those that are not are excluded. The primary objective The papers were classified into five categories based on
of the title screening process is to eliminate unrelated papers the target tasks in cosmetic dermatology, which AI was ap-
while maintaining the recall. Nevertheless, the possibility plied to address, including the cosmetic product development
of false positives persists, wherein papers may contain AI process, skin assessment, skin condition diagnosis, treatment
and dermatology components but do not specifically pertain recommendation, and treatment outcome prediction. The
to the utilization of computational intelligent technologies number of papers in each category is shown in Table 1.
in addressing challenges in cosmetic dermatology. Conse- The number of selected articles in each category by year of
quently, the remaining papers undergo additional screening publication is shown in Figure 2.
VOLUME 4, 2016 3
FIGURE 1: The PRISMA flow diagram of the study.
TABLE 1: Number of papers of each category

dedicated to expounding upon ML techniques for specificity
Category Number of papers rather than referring to them as AI in a broader sense. We di-
A Cosmetic Product Development 12 vided ML into two main classes, including conventional ML
B Skin Assessment 6 and DL approaches. Deep learning (DL) is a ML technique
C Skin Condition Diagnosis 32 that autonomously extracts and evaluates valuable features
D Treatment Recommendation 4 from raw data. Conversely, traditional ML necessitates expert
E Treatment Outcome Prediction 9 proficiency in feature selection and engineering tasks. In
addition, research has shown that, with sufficiently large data,
DL performance was shown to surpass that of traditional
This review article focused on examining and comparing ML in a variety of predictive tasks [34]. However, if the
the performance of AI techniques employed in the selected training dataset is small, traditional ML with guided feature
literature. It is noteworthy that the predominant AI method- engineering was reported to outperform DL [35]. Popular
ologies utilized in the field of cosmetic dermatology are those conventional ML methods utilized in cosmetic dermatology
falling under the umbrella of machine learning (ML). As research include Support Vector Machine (SVM), Discrim-
such, the subsequent sections of this review article will be inative Analysis (DA), Naive Bayes (NB), Decision Trees
4 VOLUME 4, 2016
FIGURE 4: Number of papers by country
IV. ANALYSIS OF REVIEWED PAPERS

This review study involved the comparative analysis of 63
FIGURE 2: Number of papers of each category by year articles, with a focus on the utilization of ML in the field
of cosmetic dermatology. The investigation aimed to assess
the current state-of-the-art contribution and novelty of such
(DT), k-Nearest Neighbors (kNN), k-Means, Principal Com- computational intelligent technologies.
ponent Analysis (PCA), and Neural Networks (NN). The
widely used DL methods are based on Convolutional Neural A. COSMETIC PRODUCT DEVELOPMENT
Networks (CNN) and Recurrent Neural Networks (RNN). In Developing cosmetic products is a crucial element in cos-
addition, while DL technologies have only recently emerged metic dermatology [37]. This process involves three distinct
[36], it is noteworthy that the utilization of deep learning subtasks, namely generating, making, and testing products
methodologies in cosmetic dermatology has gained signif- [38]. During the generation stage, a recipe is formulated
icant traction in recent years, as evident from the steep based on specific requirements. Then, the developed formu-
upward trajectory of research publications incorporating DL lae are created in a laboratory setting. Finally, it is imperative
techniques, as depicted in Figure 3. to conduct testing procedures to guarantee the safety and
the absence of any adverse consequences. Indeed, recent re-
search has discovered the use of ML to automate and improve
the efficiency and efficacy of cosmetic product development
processes, as detailed in the following subsections.
1) Cosmetics Development
Cosmetics development involves both generating formulae
and creating actual products for testing. In the generation
stage, experts must create new recipes according to the re-
quirements. Normally, this process is carried out manually.
In order to streamline this laborious task, Sunkle et al. [38]
proposed an integrated automated recommendation system
that utilizes a knowledge graph incorporating previous for-
mulation recipes and contextual information. Their method
produces a cosmetic formulation template predicated on the
input specifications. This approach may be employed to
suggest a template to the specialist as an initial reference.
FIGURE 3: Distribution of the numbers of papers utilizing Furthermore, Zhang et al. [39] transformed the cosmetic
conventional ML, DL, and both by year. formulation into an optimization problem. All conditions
were mathematically expressed as a variable and defined as
Among the 63 articles that were selected, a significant Mixed-Integer Nonlinear Programming (MINLP) problems.
number of them had first authors affiliated with institutions The objective function was defined to be the overall sensorial
located in Asia. India is the country from which the majority rating. The optimization problem was numerically solved us-
of papers originated, with a total of 11 papers. China, South ing generalized disjunctive programming reformulation and
Korea, and Taiwan follow closely behind with 10, 9, and 6 model substitution. Linear Regression, Artificial Neural Net-
papers, respectively. Figure 4 illustrates the presentation of works (ANN), and Support Vector Regression (SVR) were
papers categorized by country. employed to predict the sensorial rating. In a recent study,
VOLUME 4, 2016 5
Yeh et al. [40] employed a deep neural network to forecast The descriptor packages, Mold2 and alvaDesc, modeled with
drug-target interactions based on established relationships. Random Forest yielded satisfactory results with an accuracy
Their proposed method involved narrowing down potential of 75.86% and 79.31%.
candidates for multi-molecule drugs aimed at mitigating the Instead of using QSAR, Sharma et al. [56] developed
effects of skin aging in humans. The achieved accuracy of the a skin sensitization method predicting the allergenicity of
test was 93.077%. chemicals. They employed many classification algorithms,
including Logistic Regression (LR), k-Nearest Neighbors
2) Sensitization Testing (kNNs), Decision Tree (DT), Gaussian Naive Bayes (GNB),
In the cosmetics sector, sensitization tests are employed to XGBoost (XGB), Support Vector Classifier (SVC), and Ran-
evaluate the sensitizing potential of chemicals and medical dom Forest (RF). The chemical compounds were obtained
devices on the skin. The aforementioned tests are designed to from the IEDB database. Features were selected by removing
evaluate the capacity of a given material or product to induce the low variance and highly correlated, redundant features.
a delayed hypersensitivity response. Traditionally, there are Finally, a Support Vector Machine classifier with a linear
generally four ways to conduct sensitization experiments: in kernel was trained with these dimension-reduced features.
vivo, in vitro, in chemico, and in silico [41]. In vivo refers to The obtained feature set consists of fourteen 2D features, six
studies carried out inside a living thing, typically an animal, 3D features, and 22 fingerprint features. For the experiment,
whereas in vitro refers to experiments carried out in a lab en- Random Forest based model using all features performed the
vironment employing cells, tissues, or biological substances best (accuracy of 83.39%). Wilm et al. [57] highlighted the
outside of a living thing. In chemico refers to tests carried out lack of interpretability associated with utilizing non-intuitive
in a lab setting apart from a biological environment. Lastly, descriptors as features in a model. They introduced a novel
in silico is testing without any laboratory, the test was done compatible model, Skin Doctor CP:Bio, which utilizes a con-
completely by computation, and this is where ML was mostly cise set of ten highly interpretable features. Their proposed
employed. It is worth noting that the conventional method method, using Random Forest, achieved an efficacy of 82%
for testing sensitization through In vivo experiments often with a significance level of 0.20. Recently, Jeon et al. [58]
involves animal testing. This approach can be both financially used a graph convolutional network in their study to evaluate
burdensome [42] and ethically controversial [43]. The ethical skin sensitization. This study evaluated the potency and cat-
implications surrounding cosmetic testing on animals have egorized it into three distinct classes based on its strength:
been a topic of controversy, particularly following the 2013 strong, weak, and non-sensitizer. The model for assessing
ban by the European Union on the use of animals for testing hazards, which utilized GCN, KeratinoSens, and h-CLAT
cosmetic products and ingredients [44]. Therefore, the ethi- features, yielded the most optimal outcomes, exhibiting an
cal concerns and financial pressures associated with animal accuracy of 88%. However, they found that the model based
testing may have a substantial impact on the adoption of ML on potency alone yielded an accuracy of only 64%.
techniques to optimize skin sensitization testing procedures In addition to conducting in silico testing, researchers
[45, 46, 47, 48, 49, 50, 51, 52]. have employed ML algorithms to analyze laboratory re-
Several ML models have been devised to evaluate skin sen- sults obtained from in vitro testing. The Genomic Allergen
sitization through well-defined methodologies. However, the Rapid Detection (GARD) method was employed as a means
primary factor contributing to inaccuracies was the presence of conducting cell-based testing for skin sensitization. The
of imbalanced data – the number of sensitizers is typically study employed an in vitro model that consistently demon-
greater than that of non-sensitizers. Li et al. [53] aimed to strated a predictive accuracy of approximately 90% for the
tackle this problem by applying a data-rebalancing approach classification of the test data. Forreryd et al. [59] provided
before training an SVM model. The best-proposed model additional information to the preceding study regarding the
for hazard prediction, namely hazard-DA, reached 90.63% classification efficacy by examining a sizable external test
accuracy on the test set. For potency prediction, the potency- dataset comprising 70 observations with Support Vector Ma-
DA model yielded 68.75% accuracy on the test set. The chine (SVM). The results indicated an accuracy of 79%.
utilization of Quantitative Structure-Activity Relationship Furthermore, they presented a conformal prediction frame-
(QSAR) has been widely employed as a means of predicting work that enables the regulation of the error rate by adjusting
toxicity through in silico methods. The study established a the confidence threshold. The findings indicated that their
correlation between the chemical composition of a substance proposed approach achieved an accuracy of 88% using a
and its level of toxicity. Recent studies have reported the use confidence level of 0.85.
of diverse machine-learning classification algorithms in the Skin toxicity refers to the capacity of a substance to induce
QSAR modeling processes [54]. Akturk et al. [55] employed a localized response and/or systemic toxicity upon dermal
the QSAR model to predict the comedogenic compounds exposure [60]. The reduction in skin thickness has been iden-
present in cosmetic commodities, which underlie the prob- tified as a potential indicator of skin toxicity. However, the
lem of acne cosmetica. The study examined KStar, Ran- conventional approach of assessing epidermal layers through
dom Forest, and NNge as classification algorithms, and the manual examination by a pathologist poses challenges in
Random Forest model yielded the most favorable outcomes. terms of efficiency and scalability. To facilitate such delicate
6 VOLUME 4, 2016
processes, Hu et al. [61] utilized DL and image processing VGG16, ResNet-50, InceptionV3, MobileNetV2, DenseNet
methods to measure the epidermal thickness. The estima- 201, SqueezeNet, InceptionResNetV2, and Xception. Re-
tion was significantly correlated with a pathologist’s semi- gardless of the training time, DenseNet201 exhibited the
quantitative evaluation and mildly agreed with one performed highest level of accuracy. Two novel feature-based tech-
by other pathologists. For toxicity prediction, the method niques were introduced for obtaining high-resolution ul-
yielded 0.8 sensitivity. Furxhi et al. [62] predicted Nano- trasound images. First, an analysis was conducted on the
Particles in vitro toxicity using ML techniques on the Safe color-based features, including the mean, standard devia-
and Sustainable Nanotechnology datasets. Eight classifiers in tion, median, and luminosity value of the RGB channels.
different categories of algorithms were selected. The result This analysis used various ML algorithms, namely Logistic
showed that Random Forest and Neural Networks performed Regression, K-Nearest Neighbor (KNN), Neural Networks
best among the classifiers chosen for the evaluation. Jun and (NNs), and Random Forest. Second, a texture-oriented ap-
Shin [63] utilized convolutional neural networks and con- proach was employed, utilizing a set of five conventional
volutional long short-term memory (ConvLSTM) to predict textural characteristics in conjunction with an additional five
the artificial skin images for testing. The evaluation was outputs derived from single-layer pre-trained convolutional
conducted by comparing the projected images with the real networks, which were projected onto a lower dimensional
ones in a 3D culture setting. space using principal component analysis (PCA).
The reviewed papers in this section emphasize ML algo- Zegour et al. [65] proposed a method for assessing skin
rithms used to predict skin sensitivity and toxicity of cos- hydration levels using High-resolution Magnetic Resonance
metic chemicals, especially in the absence of animal testing. Imaging (MRI). The T2 sequence data extracted from MRI
The article discusses different approaches, focused mostly on were utilized as the model features. The segmentation algo-
in silico type testing. Since ML algorithms require sufficient rithms employed in the study were DenseNet and U-net. The
and representative data to achieve high accuracy, the lack of utilization of the Hausdorff distance metric facilitated the
these elements could be a huge challenge. Specifically, the comparison between an algorithmic output and a particular
lack of datasets and data imbalance were shown to hinder variant of human-derived segmentation. The findings indi-
the ML algorithms’ ability to learn. Various classification cated that ML techniques yielded hydration assessment that
algorithms, both DL and traditional ML ones, were used was notably closer to manual expert assessment.
to validate the hypotheses in the studies. Additionally, ML The outermost layer of the skin system, known as the
techniques were also employed to increase the efficacy and skin barrier, serves as a protective wall that shields the body
efficiency of in vitro methods, such as the genomics allergen from external hazards and maintains the body’s homeostasis
rapid detection method. These articles reviewed in this sec- by regulating water loss [66]. A traditional approach to as-
tion also highlight the requirement for additional studies in sessing the efficacy of the skin barrier function is estimating
order to create precise models that can be used as economical transepidermal water loss, which can be time-consuming. To
and humane alternatives to animal testing. speed up this process, Koseki et al. [67] introduced an ML
algorithm that utilizes topological data analysis to forecast
B. SKIN ASSESSMENT skin barrier function by analyzing skin images. Microscopic
The skin assessment task involves assessing primitive skin skin images were utilized to identify the structural charac-
properties, such as color, oiliness, and hydration, as well as teristics of the skin through the application of topological
the compatibility between the patient’s skin and treatments or data analysis. They found a strong correlation between the
products. Individuals undergoing the skin assessment process topological characteristics and the transepidermal water loss.
do not necessarily have unwanted skin conditions but rather Borade et al. [68] conducted a study to examine the skin’s
seek to better comprehend their skin properties for selecting sebum production and determined whether it exhibits char-
fitting cosmetic products or treatments. While the traditional acteristics of oily or dry skin. The experiment involved an
skin assessment procedure is carried out by dermatology examination of SVM, VGG-16, and ResNets in the classifi-
experts, recent literature has shown that such redundant tasks cation of preprocessed skin images. The findings indicated
could be assisted with AI technologies. that ResNets yielded a 98% accuracy, surpassing the per-
Skin hydration is one of the essential characteristics for formance of the other models. Furthermore, Kothari et al.
adjusting a recommended cosmetic treatment or a product [69] categorized skin images into four distinct types, namely
suggestion. Chirikhina et al. [64] employed contact capac- normal, dry, oily, and combination. A face detection tech-
itive imaging and high-resolution ultrasound imaging technique based on Multi-Task Cascaded Convolutional Neural
niques to estimate the water content in the skin. The study Network (MTCNN) was applied to each facial image, which
involved conducting experiments on multiple facial regions, partitioned it into four distinct regions, namely the forehead,
such as the volar forearm, cheek, chin, eye corner, forehead, left cheek, right cheek, and nose. Subsequently, the Convo-
lips, neck, and nose. The skin Epsilon value was utilized as a lutional Neural Network (CNN) was trained to estimate the
reference standard for measuring water content. Several deep oiliness level for each image and proceeded to classify them
learning algorithms were evaluated for the classification of based on their respective types.
contact capacitive images, including AlexNet, GoogLeNet, Skin thickness is another important integral attribute de-
VOLUME 4, 2016 7
termining skin composition and function. Measuring the a particular condition or not; (2) condition classification, a
thickness and density of skin layers in individuals poses multi-class classification problem of identifying which con-
a significant challenge due to the considerable variability dition a patient has among the predefined classes of cosmetic
observed across different sites, genders, ages, and regions conditions; (3) localization, to identify the location and type
[70]. The conventional approach to determining thickness of the condition; and (4) severity estimation which may
involves the extraction of a skin sample through a biopsy, involve a regression problem or a multi-class classification
followed by microscopic examination, which is considered problem to grade the level of severity of a specific condition.
to be an invasive procedure [71]. To mitigate this issue,
Vyas et al. [72] introduced a technique for estimating skin 1) Single Condition Diagnosis
thickness that is non-invasive. According to the authors, the In this subsection, articles that aimed to diagnose specific
lack of available true skin thickness data in the past prevented cosmetic conditions were reviewed. By focusing specifically
an accurate determination of the estimation method’s true on the diagnosis of cosmetic conditions, we aimed to ex-
accuracy. Nonetheless, the authors were able to obtain the plore the current state of research and highlight some of the
gold standard established by dermatologists and use them potential benefits associated with using ML in this context.
to calculate the prediction error directly. Furthermore, the Previously, computer vision techniques were wildly used
problem of estimating skin thickness using a Lytro camera to extract properties from lesion images for dermatological
was addressed by Ko et al. [73], where they proposed a novel disease diagnosis. This problem can be framed as an im-
approach that incorporates texture information and employs age classification task where conventional ML models are
Conditional Generative Adversarial Networks (CGANs) to trained to spot images containing the target skin conditions.
produce a skin depth map with enhanced precision. Nowadays, convolutional neural networks (CNN) and other
Understanding skin properties such as hydration, thick- DL approaches have become more utilized because of the
ness, and oily/dry classification is crucial for developing fact that they generally provide better classification efficacy
effective cosmetic treatments and product recommendations. compared to traditional ML approaches in medical image
The reviewed articles in this section have shown that skin classification tasks [78].
moisture and thickness can be precisely measured by non- Huang et al. [79] employed a CNN model based on
invasive imaging methods such as touch capacitive imag- ResNet-50 to distinguish photos of subjects with rosacea, a
ing, high-resolution ultrasound imaging, and high-resolution chronic inflammatory disease, from other skin conditions.
magnetic resonance imaging. These techniques can also be The model yielded 89.8% in accuracy. They also classified
used to precisely classify skin as oily or dry using DL rosacea lesions into three subtypes: Erythematotelangiectatic
algorithms. Such advancements in AI technologies provided Rosacea (ETR), Phymatous Rosacea (PhR), and Papulo-
a strong basis for creating treatment plans and the opportunity pustular Rosacea (PPR). Sameera et al. [80] used CNN
to create more specialized goods and treatments that cater to to evaluate the probability of the existence of three facial
the particular demands of different skin types. spots, wrinkles, dark spots, and puffy eyes. Their model
can simultaneously differentiate these characteristics and ex-
C. SKIN CONDITION DIAGNOSIS hibit various potential applications, owing to its utilization
In contrast with the skin assessment discussed in the previous of specialized convolution and pooling operations, as well
section, skin condition diagnosis mainly involves assessing as parameter shifting. Consequently, an overall accuracy of
and identifying the types and severity of unwanted cosmetic 94.11% was achieved. Liu et al. [81] used four DL archi-
conditions. Such a process is crucial for predicting effec- tectures, namely DenseNet, ResNet, Swin Transformer, and
tive treatment and management. However, even for expert MobileNet, to diagnose images of subjects with and without
dermatologists, it could be challenging to diagnose certain melasma. The research investigated the effect of different
conditions based on symptom appearances alone, as numer- photo-taking modes used by VISIA, a device for measuring
ous skin conditions can present similar features. Addressing a patient’s dyschromia from images [82]. Each subject was
this problem, ML approaches can be particularly useful, taken five shots, including Normal, UV Spots, Porphyrins,
as they can learn to capture patterns in large datasets and Brown Spots, and Red Areas modes. The experimental re-
identify relationships that may not be immediately apparent sults showed that DenseNet121 performed the best. They
to human dermatologists. In the field of medical dermatology, also discovered that the Brown Spots mode gave the best
automated ML-powered disease diagnosis has become appli- performance among all five modes (Accuracy 0.9442), and
cable, especially in skin cancer detection and classification the best combination was Brown Spots together with Normal
[74, 75, 76, 77]. and UV Spots modes (accuracy 0.974). Aditya et al. [83]
Likewise, in cosmetic dermatology, researchers have diagnosed patients who had symptoms that indicate Alopecia
demonstrated that ML approaches could be used to improve Areata, an autoimmune disorder that causes hair loss in
the accuracy and speed of diagnosis, leading to better pa- patches on the scalp, by taking images of patients’ back heads
tient outcomes. Articles in this category were divided into to create training data for five ML models, including SVM,
five subcategories: (1) single condition diagnosis, a binary CNN, KNN, Random Forest, and Gaussian Naive Bayes.
classification problem of determining whether the patient has CNN was reported to perform best.
8 VOLUME 4, 2016
In addition to skin images, other various health and bear a resemblance to one another. The majority of research
histopathological data were utilized as inputs for the classifi- presents such issues as multi-class classification tasks, which
cation tasks. Alagić et al. [84] employed an artificial neural permit the direct application of traditional and advanced ML-
network (ANN) to analyze skin health data, which included based image classification techniques. The task of classifying
parameters such as pH value, sebum, and transepidermal skin images into respective skin conditions has been pre-
water loss. These features were utilized to distinguish indi- viously accomplished through the utilization of traditional
viduals who are in good health from those who have skin ML techniques, including Support Vector Machines and K-
conditions. The dataset comprises various dermatological Nearest Neighbor algorithms. For example, Abas et al. [88]
conditions such as acne, dry skin, decreased elasticity, and employed a methodology wherein RGB facial images fea-
wrinkles. Dubey et al. [85] diagnosed the scalp by per- turing acne were transformed into Grayscale, followed by
forming an optical coherence tomography. This non-invasive applying an entropy-based filter to isolate the regions of the
imaging technique uses light waves to create high-resolution image that exhibit acne. The segmented image was analyzed,
images of internal tissues and structures. The A-line and B- and various features were categorized into six distinct skin
scan features were extracted from the OCT. Seven models, conditions, one of which was acne. The experimental inves-
including classical ML models and neural networks, were tigation involved using Binary Tree, Discriminant Analysis,
used in the pilot experiment. The multilevel ensemble model k-NN, and Naive Bayesian techniques, where an accuracy of
gave the highest performance using eight features. Jansen 85.5% was attained.
et al. [86] diagnosed Seborrheic Keratosis, a type of benign Recently, DL models, especially CNN-based ones, such as
skin condition commonly found in senior patients. The data AlexNet, GoogleNet, and DenseNet, have been popularized
comprised images of tissue slides from three dermatologi- in the medical areas. For example, Yang et al. [89] used
cal research centers. ResNet34 was used as a classification DenseNet-96 and ResNet-152 to classify the 12,816 cropped
model. Wang et al. [87] investigated the skin disease from benign and pigmented facial skin lesion photos collected at
metagenomic sequencing data of acne, which is a kind of the Hospital for Skin Diseases of the Chinese Academy of
lipid in the face skin. The data were analyzed using Principal Medical Sciences from 2004 to 2016 into six classes based
Component Analysis (PCA), Kernel Principal Component on their skin conditions, including acquired nevi of Ota,
Analysis (KPCA), and Multiset Canonical Correlation Anal- melasma, café-au-lait spots, freckles, seborrheic keratoses,
ysis (MCCA). Each method provided information pertaining and nevi of Ota. The automated classification result was
to lipids that influence the diagnosis results. compared with the three expert dermatologists. The results
To summarize the literature reviewed in this section, the showed that ResNet-152 outperformed the other methods.
majority of research utilizing AI to identify a single cosmetic While these DL algorithms often are accompanied by pre-
skin condition relied on computer vision and ML technolo- trained models that can be directly fine-tuned and applied
gies to train predictive models using annotated historical to image classification tasks, recent literature in cosmetic
data. While traditional computer vision methods relied on dermatology has found that revising the architectures of
hand-crafted color and texture features, the advent of DL these DL algorithms could improve the predictive efficacy.
applications enabled the elimination of the feature engineer- For instance, Huong et al. [90] focused on the problem of
ing process while achieving cutting-edge performance. The limited training data by proposing to ensemble pre-trained
single-condition diagnosis research classified skin images CNN with SVM and KNN. The performance of classify-
into two classes, with a condition or without a condition. ing four skin disease classes of transfer-learned AlexNet,
One major limitation of this approach is that it needs proper AlexNet-SVM, and AlexNet-KNN was evaluated and com-
prior information on which condition this patient may have. pared. The modified model outperformed the non-modified
In fact, some condition is very similar and hard to distinguish one and also reduced the computational time. López-Leyva
by physical observation. Such issues pose challenges for ML et al. [91] addressed this problem through the development
models trained specifically for diagnosing particular condi- of a method aimed at classifying ten distinct categories of
tions in that they may be unable to tell apart closely-resemble skin lesions. Their method relied on the Fourier spectral
skin conditions. Furthermore, cosmetic patients may have information of images within a color model. The Edinburgh
varied conditions. Therefore, the ability to detect only one Dermofit Library utilized a 26 × 1 vector to represent each
condition may not suffice in practice. As a result, these issues image, consisting of Fourier spectral indicators that pertain
behoove the ability to automatically tell apart different types to both the original size image and the cropped version.
of cosmetic conditions, especially those that appear similar Subsequently, the vector that was represented was inputted
to each other, which will be elaborated on in the next section. into a Two-Layer Feed-Forward Neural Network (TLFN)
with the purpose of accurately categorizing the lesion ac-
2) Conditions Classification cording to its respective type. Overall, the proposed method
Identifying various skin conditions can be difficult due to exhibited a 99.33% accuracy, 94.16% precision, 92.9% sen-
their similar visual presentations. The articles within this sitivity, and 99.63% specificity. Jain et al. [92] proposed
particular subcategory examine the utilization of AI in dis- the Optimal Probability-based Deep Neural Network (OP-
criminating between similar beauty-related conditions that DNN) for the purpose of classifying skin images into four
VOLUME 4, 2016 9
distinct categories, namely Basal Cell Carcinoma, Seborrheic acne, in order to evaluate the corresponding severity levels
Keratosis, Melanoma, and Squamous Cell Carcinoma. The [97]. It is imperative for automated systems to possess the
study involved extracting seven distinct color and texture ability to identify and quantify any and all instances of
features from each image, namely mean, standard devia- inflammatory lesions present on a patient’s facial image. This
tion, skewness, contrast, correlation, energy, and extropy. issue necessitates a redefinition of the diagnosis task as a
The OP-DNN methodology was designed to expedite the lesion detection task that is cognizant of location, type, and
training process of conventional DNNs by leveraging the size information to provide requisite information at a more
WOA optimization algorithm instead of refreshing weight granular level.
values at each cycle. The results indicated that the OP-DNN
approach achieved a marginally superior level of accuracy 3) Conditions Localization and Detection
and precision compared to the baseline method while also The task of diagnosing cosmetic conditions through image
exhibiting a notable reduction in training time. Ito et al. classification can pose a challenge due to the presence of
[93] employed the Google Cloud AutoML technology to extraneous information in the background of the images. In
classify scar images into four distinct categories, namely addition, specific therapies for skin conditions that present as
immature scar, mature scar, hypertrophic scar, and keloid. isolated, non-adjacent spots necessitate knowledge regarding
The outcomes of the classification were compared with the the precise locations and dimensions of the lesions, which
expert medical judgment. In a recent study conducted by do not accompany the image classification task. The ability
Borade et al. [94], the authors expanded their analysis beyond to localize and confine the lesions in the input image before
images in the RGB color space to include three additional performing further analyses could mitigate such issues. In-
color spaces: YUV, YCbCr, and HSV. The study employed deed, the process of localizing a condition is a crucial aspect
five traditional ML techniques, namely Support Vector Ma- of both diagnosis and the estimation of its severity, as it
chine (SVM), k-Nearest Neighbor (kNN), Naive Bayes (NB), eliminates extraneous information from the images. The task
Multilayer Perceptron (MLP), and Random Forest (RF) to can be accomplished through the application of conventional
categorize images of four distinct dermatological conditions, image processing methodologies, including edge detection,
namely acanthosis nigricans, melasma, alopecia areata, and thresholding, and clustering. Object detection is a more spe-
acne. cific task that combines both localization and classification
Kim and Song [95] identified several limitations associated together. This section examines scholarly articles that en-
with the utilization of CNN-based models in the classification deavor to identify cosmetic conditions, beginning with the
of facial skin conditions. These limitations include the chal- pure localization approach and progressing to the utilization
lenge of accurately identifying minor skin issues, the need to of object detection methodologies.
classify over 20 distinct conditions, the presence of variations The detection of facial wrinkles through automated tech-
within the same condition, the potential for confusion be- niques poses a significant challenge to cosmetic condition
tween similar conditions, and the possibility of false segmen- detection tasks. Yap et al. [98] conducted a survey on
tation on non-facial regions. The authors proposed effective the topic of automated facial wrinkles detection, including
strategies for overcoming each constraint, and the empirical various types of research, including handcraft-based image
findings demonstrated a 32.58% higher diagnostic efficacy in processing techniques, stochastic models, and mathematical
comparison to the traditional Convolutional Neural Network filters. The conclusion highlighted that while there has been
(CNN). a notable surge in the application of DL methods for image
Apart from cosmetic skin conditions, Jeong et al. [96] inpainting, there is a scarcity of research that has specifically
employed the EfficientNet to categorize ten distinct scalp explored its potential for addressing facial wrinkles.
symptoms, namely normal, drying, oily, sensitivity, atopy, Rew et al. [99] utilized the Deeplab-v3+ and Inception-
seborrheic, trouble, dry dandruff, oily dandruff, and hair loss. ResNet-v2 models to perform pixel segmentation on skin
Their classification technique formed an integral component wrinkles. Then, LightGBM and MP algorithms were em-
of AI-ScalpGrader, which comprised a handheld scalp imag- ployed to enhance the segmentation procedure’s efficacy.
ing device, a smartphone application, and a cloud-based ad- Their proposed segmentation scheme yielded a mean accu-
ministration platform. The system provided accuracy values racy of 85.4%, a mean intersection over union of 74.9%,
ranging from 87.3% to 91.3%. and a mean boundary F1 score of 85.2%, improving over
The capacity to categorize skin images into various skin the panoptic-based semantic segmentation method by 1.1%,
conditions could potentially assist in initial self-diagnosis 6.7%, and 14.8%, respectively. Ismail and Sung [100] intro-
and could be integrated into a decision support system for duced a deep-learning framework designed to identify the
dermatologists. However, the ability to precisely identify locations and types of acne lesions and wrinkles in facial and
the location and dimensions of each lesion could offer vital half-body images. Various deep-learning networks, namely
supplementary insights, especially in cases where the lesions the Faster RCNN and the Residual Network, were explored.
are diminutive or consist of diverse skin conditions. For The convolutional feature map was generated by utilizing
example, some grading systems necessitate the quantification 50 layers of a residual neural network to extract the char-
of papules and pustules, which are indicative of inflamed acteristics of the image. The mAP score of the detection
10 VOLUME 4, 2016
model was found to be 47.96%. Shih et al. [101] employed a formed slightly better than using only the YOLO model. The
weakly supervised algorithm to localize the area of Vitiligo highest obtained accuracy was 60.38% by using Mask R-
to evaluate the treatment system. Specifically, Wood’s lamp CNN and YOLO-v4 with 500 training images. Ding et al.
was utilized to capture and photograph both the impacted [108] conducted a comparable experiment utilizing YOLO-
and unaffected regions in order to establish a repository of v4, YOLO-v5, Single-Shot Multi-box Detection (SSD), and
training images. Then, a CNN model was used to perform Faster R-CNN. As anticipated, YOLO-v5 demonstrated su-
an initial segmentation of the affected region from large- perior performance in all skin conditions, with the exception
scale images, such as those of the head and face. Next, of melasma, when compared to other methods.
physician-validated and authorized images depicting the im- The literature examined in this section presents the di-
pacted region were utilized alongside a substantial quantity agnosis of cosmetic conditions as a task of localization or
of images captured by Wood’s lamp to enable self-learning object detection. This involves obtaining not only detailed
and classification. Finally, facial recognition technology was locations of the lesions but also identifying the specific types
employed to rectify the camera’s shooting angle, thereby of conditions. The utilization of lesion localization and object
mitigating image distortion arising from disparate shooting detection techniques enables the advancement of automated
angles. skin condition diagnosis to a finer granularity. Automated
In addition to detecting skin lesions, Gallucci et al. [102] lesion localization methods have significantly reduced the
employed U-Net, an image segmentation technique, to seg- necessity for a manual cropping process, facilitating stream-
ment and detect hair from images for quantifying hair num- lined and informed diagnosis.
bers. Furthermore, they utilized the said method in con-
junction with the detection of skin lesions. The experiment 4) Severity Estimation
involved a comparison between U-net and several other mod- In addition to ascertaining a patient’s skin condition, it is
els, namely Lenet-5, VGG-16, ResNet-50, and DenseNet- also crucial to consider its degree of severity. Severity es-
121. The employment of U-Net yielded the highest correla- timation is a technique employed to assess the degree of
tion with the manual count by experts. criticality of a given condition or its level. The ability to
In addition to identifying the location of the lesions, ob- acquire this knowledge automatically diminishes the duration
ject detection methodologies were also used to classify the of dermatologists’ involvement in evaluating the severity
localized lesions into respective condition types. Phan et al. of a condition and facilitates the selection of more fitting
[103] designed an LED therapy device that incorporates an treatments [109]. The degree of severity can be assessed
automated algorithm for diagnosing acne vulgaris. The pro- either through continuous scoring, which poses a regression
posed model was derived from a modified version of ResNet- problem, or through discrete categorization of the condition,
50 architecture, which was integrated with YOLO-v2. Once ranging from normal to extremely severe [110, 111].
the location of acne was automatically identified on the There has been a growing interest in utilizing smartphone-
input self-image, the information will be transmitted to the generated selfie images to automatically assess the severity
intelligent LED therapy device for further processing. Wen of various skin conditions, owing to their ease of acquisi-
et al. [104] conducted a comparative analysis of the efficacy tion. However, it should be noted that the accuracy of the
of utilizing CNN as the underlying framework for detecting estimation may be influenced by factors such as lighting
acne vulgaris. A number of object detection algorithms were conditions, facial expressions, and individual variations in
evaluated, namely MobileNetv1, Yolov4, and Inception-v2. skin type. To address these particular problems, Jiang et al.
They also implemented an automated severity evaluation [112] conducted an investigation wherein they modified the
tool that was made publicly available through the WeChat Convolutional Neural Network (CNN) classifier to a CNN
application for self-monitoring of acne. In addition to acne regressor. This allowed them to obtain a score for various skin
detection, Maknuna et al. [105] suggested a scar lesions facial conditions, such as wrinkles, folds, lines, and pores.
detection model for the WSI of HE-stained tissue. Mask The dataset utilized in the study comprised a diverse range
R-CNN was used to detect scar lesions. Then, ResNet-101 of age groups spanning from 18 to 80 years old, distinct
was used as a backbone of the region proposal network. The cohorts including Asian, Caucasian, and African American,
detected region of interest was fed to the image clustering and various lighting conditions such as outdoor natural day-
model, k-Means, to partition the structure and character of light, indoor natural daylight, indoor artificial diffused light,
the scar. and indoor artificial direct light. Additionally, the dataset
You Only Look Once (YOLO) [106] has emerged as an contained diverse facial expressions characterized by slight
efficient object detection algorithm and has also been used for smiles, slight pouts, or disapproval. The obtained scores
skin condition recognition by framing the problems as object were compared with the evaluation provided by the proficient
detection tasks. Liao et al. [107] experimented with distin- evaluator. The findings indicate a lack of complete concur-
guishing acne, freckle, and wrinkle images with YOLO-v3, rence between the automated approach and the evaluation
YOLO-v4, and with and without Mask R-CNN. The results of the specialist. However, they do suggest that the outcome
showed that using Mask R-CNN as a face segmentation was marginally affected by the lighting situation and facial
algorithm before using YOLO to detect the symptoms per- expression.
VOLUME 4, 2016 11
Subsequently, the aforementioned research team pro- ditions. Wang et al. [118] examined Microtia, a congeni-
ceeded with the advancement of an automated grading sys- tal ear malformation, through the utilization of nine CNN-
tem for facial conditions utilizing self-portrait photographs, based models, namely AlexNet, Inception-v3, DenseNet-
where Flament et al. [113] yielded a superior correlation 121, ResNet-18, ResNet-50, ResNet-101, ShuffleNet-v2,
between the automated outcome and the evaluation provided MobileNet-v2, and MnasNet. The objective was to determine
by the expert. In addition, Flament et al. [114] employed the efficacy of these models in accurately classifying the
an algorithm derived from [112, 113] to examine images degree of Microtia based on ear images. The images were
of selfies. A total of 465,587 images of European women assessed and categorized into four distinct levels, namely
and 79,016 images of Chinese women were utilized in the normal ears, grade I microtia, grade II microtia, and grade
study, where the researchers assessed the severity of nine III microtia. Man et al. [119] introduced a novel approach,
skin conditions and examined their correlation with the age SACN-Net, for assessing the extent of hair damage based
of the subjects. In a recent study, Flament et al. [115] eval- on SEM images. The study’s findings indicated that SACN-
uated the accuracy of grading the severity of various skin Net outperformed other established CNN-based models, as
conditions using images captured through selfies. The study evidenced by an accuracy rate of 98.38%. Chang et al. [120]
utilized a sample of 1,041 self-portrait images captured by proposed the ScalpEye system, a handy scalp hair imaging
women in the United States, featuring diverse age ranges, microscope with a mobile device application that connects
Fitzpatrick skin types, geographical locations, and ancestral to the AI training server. The scalp hair imaging microscope
backgrounds. The severity of seven facial skin conditions was took images, then proceeded to the DL model and reported
estimated utilizing algorithms as described in [112, 113]. The severity scores for the four common scalp hair symptoms:
results generated through automation were compared with Dandruff, Folloculitis, Hair Loss, and Oily Har. The reported
the severity levels assessed by proficient dermatologists from severity levels are: minor, normal, middle, and high. For the
the United States. The findings of the study indicated a robust DL model, the authors tried Faster R-CNN Inception-v2,
correlation between the automated and dermatologist results SSD Inception-v2, and Faster R-CNN Inception-ResNet-v2-
for five out of seven conditions, namely Forehead wrinkles, Atrous. The experimental result showed that Faster R-CNN
Periorbital wrinkles, Nasolabial fold, Ptosis of the lower part InceptionResNetV2Atrous was the best algorithm for all four
of the face, and Diffused redness. On the contrary, there exists symptoms, with average precision ranging from 97.41% to
a moderate and weak correlation between the pores present 99.09%. Still, its training time was significantly higher than
on the skin of the cheeks and the darkest skin tones. The the others.
findings indicated that neither age nor ancestry exerted any The reviewed papers presented in this section show that
influence on the observed correlations. deep convolutional neural networks and their variants have
In addition to easily obtainable selfie images, ML solu- emerged as commonly employed algorithms in severity as-
tions have been devised to aid in the assessment of severity sessment tasks. The experiment on estimating a specific con-
grading on high-quality images captured by professional dition’s severity involves the implementation and comparison
cameras. The detection and evaluation of facial wrinkles, of numerous ML algorithms trained on image datasets whose
pores, and acne were conducted by Seck et al. [116] using samples were annotated with appropriate severity levels. The
the high-resolution 3D surface texture obtained from the evaluation of their performance was primarily conducted
light stage. Furthermore, the study conducted by Wang et al. through a comparative analysis of their graded outcome and
[117] sought to develop a tool for evaluating the severity that of an expert. Later on, after the condition type and sever-
of acne vulgaris. To achieve this objective, the researchers ity grade were indicated, the next natural research question
introduced a convolutional neural network (CNN) model, would be whether such predictive diagnosis methods could
which they named lightweight Acne-RegNet. This model be used to infer dermatologists’ treatment options. Indeed,
is capable of accurately categorizing lesions and providing this process is framed as a treatment recommendation task
a corresponding severity score. The comparative analysis which will be discussed in the next section.
involved the proposed models and other lightweight deep-
learning models such as MobileNet-V3, SENet, EfficientNet-
B0, and ghostNet. The Acne-RegNet exhibited superior per- D. TREATMENT RECOMMENDATION
formance compared to other models, achieving an accuracy Patient-centered medicine is an approach that aims to con-
of 94.11% on the test dataset. Furthermore, they continued sider the treatment effectiveness and patient satisfaction by
to examine a visual condition that impacts precision. The tailoring it to the specific disease and patient, considering
findings indicated that the utilization of a front-facing camera individual variability in clinical presentation, medical history,
had a negative impact on the algorithm’s efficacy. The study genes, environment, and lifestyle. Especially in cosmetic
also found that the accuracy was not significantly affected by dermatology, which relates to personal preferences, the treat-
the device brand or the light conditions, including outdoor, ment paradigm has shifted from disease-centered to patient-
indoor, and flash. centered health care [121]. To accomplish this, the develop-
In addition to facial skin conditions, ML techniques were ment of intelligent technologies is needed to overcome these
employed to assess the severity of various cosmetic con- challenges and fulfill this goal [122].
12 VOLUME 4, 2016
Huang et al. [123] invented Alluring, a cloud-based sys- E. TREATMENT OUTCOME PREDICTION
tem for dermatological analysis of skin and scalp, which Following the process of diagnosing the condition, the subse-
utilized skin images to provide treatment recommendations. quent step involves selecting an appropriate treatment plan.
The system comprises a handheld device equipped with a In the context of clinical dermatology, the responsibility of
camera for capturing dermatological images. A skin image selecting treatments primarily falls upon the dermatologist.
is processed by a comprehensive analysis for various factors, Cosmetic cases are different as they do not pose any harm to
including moisture, oil, sensitivity, color, pore size, and pore the patient, and there exists a plethora of treatment options
distribution, utilizing YOLO-v2. Subsequently, the outcome that patients may opt for. Dermatologists are presented with
of the analysis was utilized to suggest a dermatological a wide range of treatment options, and patients may play a
product and facilitate customers to make a purchase within role in selecting a treatment modality based on their per-
the application. sonal preferences and financial considerations. The treatment
In addition to utilizing skin images, genetic information outcome prediction task involves forecasting a patient’s re-
was also employed. Liu et al. [124] developed a method for sponses after receiving a particular treatment. The prediction
recommending cosmetic products by integrating genetic data of treatment outcomes holds significant importance at this
related to consumer skin health, product data, demand factors stage as it could guide dermatologists and patients toward
based on phenotype, and data on the relationship between narrowing down appropriate treatment plans. This section
ingredients and their functions. The data pertaining to skin- will delve into the articles examining AI techniques for
health products were transformed into numerical data and predicting treatment outcomes.
subsequently classified using three ML algorithms, namely The utilization of simulated postoperative images derived
Random Forest, Logistic Regression, and Support Vector from preoperative images can serve as a valuable tool for pa-
Machine. The empirical findings demonstrated that Support tients in making informed decisions regarding their treatment
Vector Machine exhibited better time efficiency compared options. Shah et al. [126] have demonstrated the ability to
to other methods, while Random Forest (RF) marginally generate a precise three-dimensional facial image subsequent
outperformed other classifiers in terms of classification ef- to the Rejuvenation procedure. The model has been utilized
ficacy. The utilization of genetic data and the consideration as an input for generating 3D facial scan images. Facial
of trade-offs between phenotypic demands resulted in an landmarks were identified as injection sites for dermal fillers.
improvement in the recommendation performance. Their study introduced a model that forecasts the quantity
Ray et al. [30] proposed a scheme for recommending cos- of dermal filler required for facial application by utilizing
metic products that utilize Convolutional Neural Networks a multi-layered neural network architecture comprising two
(CNN). The study employed image analysis techniques to concealed layers. Their approach yielded an accuracy of
predict various categories of consumer facial images based 62.5%, surpassing that of the baseline methods 3D-Div by
on skin health. This was achieved by extracting relevant 51.5% and 3D-Vor by 55.8%. Shah et al. [127] proposed
features such as shape, texture, and color from the pho- enhancement to their simulation model for postoperative
tographs. The algorithm that was proposed attained a success rejuvenation image prediction using ML techniques. In this
rate of 97.38% accuracy in recommending items on the test study, a deep neural network model, Rejuv3DNet, and a
data. Zhang et al. [125] utilized knowledge graphs to de- kernel regression-based (KR) model were developed and
velop a recommendation system for cosmetic sequences. The demonstrated accuracy rates of 62.5% and 66.7%, respec-
construction of the knowledge graph for skincare products tively. In addition, they produced the initial 3D facial dataset
was achieved through a combination of manual screening that includes 3D facial images before and after receiving
and multi-label classification techniques applied to an open treatments. Lin et al. [128] utilized cosmetic laser therapy to
dataset. A ranking algorithm was developed with the aim of modify the melanin and hemoglobin components of the skin,
suggesting the optimal product based on the specific needs of resulting in the desired outcome of the treatment. In their
consumers and their individual skin types. study, ML algorithms were employed to retouch freckles and
ML techniques were employed to evaluate skin images, adjust skin tone by considering variations in melanin and
genetic data, and other pertinent factors for the purpose of hemoglobin levels based on the training data.
providing personalized treatment recommendations in the While the aforementioned articles focus on simulating the
field of aesthetic dermatology. This methodology facilitated posttreatment outcomes, other studies in treatment outcome
the development of a therapeutic regimen that is both ef- prediction also investigated the possibility of using AI tech-
fective and tailored to specific customer requirements while niques to quantify the treatment success chances. Akben
considering the unique clinical manifestations, medical back- [129] employed a decision tree-based fuzzy informative ap-
ground, genetic makeup, surroundings, and lifestyles. These proach to predict the success of various wart treatments. The
techniques aim to enhance the accuracy and efficiency of utilization of an automated prediction model has been pro-
the system in aligning customers with the optimal products posed as a computer-assisted tool for medical professionals.
and treatments based on their individual requirements, skin The variables used in their study consisted of the patient’s
characteristics, and dermatological conditions. gender, age, duration of time before treatment, quantity and
classification of warts, surface area, and the induration diam-
VOLUME 4, 2016 13
eter of the initial test. These features were utilized to forecast preoperative image compared between the two groups.
the outcome of the treatment as a dichotomous variable, In the cosmetic businesses, customer satisfaction held a
namely, positive or negative. The findings indicated that the significant important attribute. In this direction, The study
duration between onset and treatment provided the most conducted by Kim et al. [134] investigated the emotional
information gain, followed by age, as determined through responses of customers during the use of cosmetic cream
a comparative analysis with established classification tech- through the analysis of EEG data. The study was conducted
niques, including SVM, KNN, Random Forest, and Logistic by assessing the electroencephalogram (EEG) activity of par-
Regression. The Decision Tree approach demonstrated the ticipants during the administration of four distinct categories
highest level of accuracy of 94.4%. of topical skincare products. Subsequently, participants were
Erdoǧan et al. [130] constructed the post-operation evalu- administered a questionnaire to assess their level of satisfac-
ation of the FUE hair transplantation procedure. Their algo- tion with the cream. The proposed features were extracted
rithm was implemented as part of KEBOT, a comprehensive from the EEG signal and processed in a CNN-based model
device designed for hair transplantation. The KEBOT system to predict satisfaction. The findings revealed that the stacked
comprised an operational infrared-based depth camera that CNN model yielded an accuracy of 75.4%, surpassing all
was utilized to produce a three-dimensional model of the other selected models for the experiment.
user’s head. The acquired data was processed to extract Reviewed papers in this section demonstrated the applica-
information. Subsequently, the DL algorithm was employed tions of ML models for predicting treatment outcomes, suc-
to conduct an analysis that commenced with object detection, cess rates, and postoperative changes. These models could be
followed by hair thickness estimation, and culminated in used to assist patients and dermatologists in making informed
metrical analysis. The investigation focused on RetinaNet, decisions about treatment options and to help dermatologists
M2Det, YOLO-v4, and EfficientDet during the object detec- select suitable courses of treatment for their patients. Addi-
tion stage. The hair thickness was estimated through the uti- tionally, research in this field may result in the development
lization of SegNet, UNet, and ERFNet for hair segmentation. of advanced simulation tools, such as those that can simulate
Finally, the surgeon was presented with the post-operative the outcomes of several cosmetic procedures simultaneously
prediction in order to strategize the surgical procedure. Shi or those that can take a wider range of patient preferences and
et al. [131] created SkincareMirror, a personalized appear- characteristics into account.
ance prediction after using skincare products. SkincareMirror
was developed for applications by males and females, regard- V. DISCUSSION
less of their knowledge pertaining to skincare products. The In the discussion section, a summary of the findings from the
study conducted on the cosmetic product website revealed reviewed articles is presented. The discussion is structured
that users exhibited different behavioral patterns when using into three distinct segments, namely trend, limitation, and
SkincareMirror. Specifically, the results indicated that users opportunity.
who utilized SkincareMirror tended to click on a greater The review analysis presented in the previous section
number of products, albeit spending comparatively less time reveals an apparent trend of heightened adoption of ML tech-
reading through the product descriptions. The results also niques in the domain of cosmetic dermatology. Historically,
indicated that the male cohorts who did not have skincare traditional ML models served as the primary means of anal-
knowledge exhibited higher levels of satisfaction with the ysis. However, modern practices have shifted to utilizing DL
system in comparison to the remaining groups. methods due to their various advantages in medical domains
Some cosmetic treatments, especially plastic surgery, can [135]. Numerous studies have conducted a comparative anal-
completely alter a patient’s appearance. K and Krishnaveni ysis between conventional ML and DL methodologies. Ana-
[132] pointed out that this change may affect the face lyzing facial images has historically posed a challenge due
recognition and identity identification system. Hence, pair- to the quality and quantity of input data. However, recent
ing posttreatment images with pretreatment images is also advancements in DL techniques offer potential solutions
an essential task. They compared the performance of two to these issues. Currently, there is a significant research
common feature extraction techniques: Extended Uniform emphasis on utilizing self-portrait photographs as input, as
Circular Local Binary Pattern (EUCLBP) and Scale Invari- evidenced by recent studies [112, 113]. Furthermore, the
ant Feature Transform (SIFT). The findings indicated that exponential growth in the population of smartphone users has
the optimal outcome was achieved through the combined brought automatic self-diagnosis and product recommenda-
utilization of SIFT and EUCLBP, as opposed to individual tions to the forefront of attention. As a result, AI-powered
models. Bahçeci Şimşek and Şirolu [133] studied the changes dermatology applications have been witnessed to continue
of patients who did upper eyelid blepharoplasty surgery and rising [136]. Numerous scholarly articles have extensively
compared the result with and without a Müller’s muscle- utilized ML methodologies as an integral component of their
conjunctival resection (MMCR). In the experiment, upper respective application or website systems [103, 117, 131].
eyelid blepharoplasty surgery patients were divided into two The issue of insufficient data persisted as a constraint in
groups, with and without MMCR. After six months, the full- the adoption of ML techniques for diagnosis and assessment
face image was analyzed by measuring the change from the of the severity of cosmetic dermatological conditions. Due
14 VOLUME 4, 2016
to ethical concerns, a majority of research studies have been sis of nationally representative data.” Journal of drugs
limited to small and specific datasets. Furthermore, such in dermatology: JDD, vol. 11, no. 4, pp. 466–473,
concerns also influence the sharing of data for research 2012.
purposes. The absence of diverse data may lead to overfitting [2] I. G. Ferreira, M. B. Weber, and R. R. Bonamigo,
of the model and a deficiency in its ability to generalize. For “History of dermatology: the study of skin diseases
example, the utilization of color and skin texture as extracted over the centuries,” Anais Brasileiros de Dermatolo-
features to indicate skin health in various models resulted in gia, vol. 96, pp. 332–345, 2021.
a constructed model that exhibited satisfactory performance [3] S. Tripathi, “Artificial intelligence: A brief review,”
solely for the learned data, which was predominantly derived Analyzing future applications of AI, sensors, and
from limited skin types. Some research has attempted to robotics in society, pp. 1–16, 2021.
solve this problem by varying the skin types for machines [4] Y. Mintz and R. Brodie, “Introduction to artificial
to learn, but this was still done in a limited variation. intelligence in medicine,” Minimally Invasive Therapy
Another significant critique regarding the utilization of & Allied Technologies, vol. 28, no. 2, pp. 73–81, 2019.
ML, particularly DL, approaches in the field of cosmetic der- [5] Y. LeCun, Y. Bengio, and G. Hinton, “Deep learning,”
matology pertains to their opaque nature, commonly referred nature, vol. 521, no. 7553, pp. 436–444, 2015.
to as ’black boxes.’ Despite achieving satisfactory levels of [6] S. Chibani and F.-X. Coudert, “Machine learning ap-
accuracy, the model’s incapacity to provide a clear expla- proaches for the prediction of materials properties,”
nation for its decision-making process during the prediction Apl Materials, vol. 8, no. 8, p. 080701, 2020.
of diagnoses, evaluations, or treatments could potentially [7] S. Khan, M. Sajjad, T. Hussain, A. Ullah, and A. S.
result in unforeseen consequences during practical applica- Imran, “A review on traditional machine learning and
tion [137]. Explainable artificial intelligence (XAI) has the deep learning models for wbcs classification in blood
potential to address this concern and presents a promising smear images,” Ieee Access, vol. 9, pp. 10 657–10 673,
avenue for further investigation within this domain. 2020.
[8] R. Brehar, D.-A. Mitrea, F. Vancea, T. Marita,
VI. CONCLUSION S. Nedevschi, M. Lupsor-Platon, M. Rotaru, and R. I.
The utilization of artificial intelligence and machine learning Badea, “Comparison of deep-learning and conven-
is a significant factor in numerous functions within the field tional machine-learning methods for the automatic
of cosmetic dermatology. This systematic literature review recognition of the hepatocellular carcinoma areas from
was undertaken to provide a comprehensive summary of ultrasound images,” Sensors, vol. 20, no. 11, p. 3085,
the contemporary research utilizing machine learning in this 2020.
domain in accordance with the PRISMA protocol. The 63 pa- [9] D. H. Murphree, P. Puri, H. Shamim, S. A. Bezalel,
pers that underwent review were categorized into five distinct L. A. Drage, M. Wang, M. R. Pittelkow, R. E. Carter,
groups according to their respective tasks, namely: cosmetic M. D. Davis, A. G. Bridges et al., “Deep learning for
product development, skin assessment, skin condition di- dermatologists: Part i. fundamental concepts,” Journal
agnosis, treatment recommendation, and treatment outcome of the American Academy of Dermatology, vol. 87,
prediction. The utilization of machine learning approaches no. 6, pp. 1343–1351, 2022.
in the domain of cosmetic dermatology was highlighted, [10] P. Puri, N. Comfere, L. A. Drage, H. Shamim, S. A.
with a focus on identifying trends, limitations, and future Bezalel, M. R. Pittelkow, M. D. Davis, M. Wang, A. R.
opportunities. The primary contribution of this article is a Mangold, M. M. Tollefson et al., “Deep learning for
methodical examination of existing recent studies aimed at dermatologists: Part ii. current applications,” Journal
the utilization of artificial intelligence (AI) technologies in of the American Academy of Dermatology, vol. 87,
cosmetic dermatology. We expect this study to provide an no. 6, pp. 1352–1360, 2022.
overview for researchers seeking to explore contribution gaps [11] F. Hashimoto, H. Ohba, K. Ote, A. Teramoto, and
in this area as well as medical and IT practitioners looking H. Tsukada, “Dynamic pet image denoising using
to utilize intelligent technologies to address real-world chal- deep convolutional neural networks without prior
lenges in cosmetic industries. training datasets,” IEEE access, vol. 7, pp. 96 594–
96 603, 2019.
ACKNOWLEDGMENT [12] M. Jiang, Z. Wang, and Q. Dou, “Harmofl: Harmo-
This research project is supported by Mahidol University nizing local and global drifts in federated learning on
(Fundamental Fund: fiscal year 2023 by National Science heterogeneous medical images,” in Proceedings of the
Research and Innovation Fund (NSRF)). AAAI Conference on Artificial Intelligence, vol. 36,
no. 1, 2022, pp. 1087–1095.
REFERENCES [13] S. Borade and D. Kalbande, “Survey paper based
[1] S. A. Davis, S. Narahari, S. R. Feldman, W. Huang, critical reviews for cosmetic skin diseases,” in 2021
R. O. Pichardo-Geisinger, and A. J. McMichael, “Top International Conference on Artificial Intelligence and
dermatologic conditions in patients of color: an analy- Smart Systems (ICAIS). IEEE, 2021, pp. 580–585.
VOLUME 4, 2016 15
[14] R. Bhardwaj, A. R. Nambiar, and D. Dutta, “A study [27] Y. Yang, L. Guo, Q. Wu, M. Zhang, R. Zeng, H. Ding,
of machine learning in healthcare,” in 2017 IEEE 41st H. Zheng, J. Xie, Y. Li, Y. Ge et al., “Construction
annual computer software and applications conference and evaluation of a deep learning model for assessing
(COMPSAC), vol. 2. IEEE, 2017, pp. 236–241. acne vulgaris using clinical images,” Dermatology and
[15] T.-C. Pham, C.-M. Luong, V.-D. Hoang, and Therapy, vol. 11, no. 4, pp. 1239–1248, 2021.
A. Doucet, “Ai outperformed every dermatologist in [28] F. Linming, H. Wei, L. Anqi, C. Yuanyu, X. Heng,
dermoscopic melanoma diagnosis, using an optimized P. Sushmita, L. Yiming, and L. Li, “Comparison of
deep-cnn architecture with custom mini-batch logic two skin imaging analysis instruments: The visia®
and loss function,” Scientific Reports, vol. 11, no. 1, from canfield vs the antera 3d® cs from miravex,”
p. 17485, 2021. Skin Research and Technology, vol. 24, no. 1, pp. 3–8,
[16] M. Xiong, J. Pfau, A. T. Young, and M. L. Wei, 2018.
“Artificial intelligence in teledermatology,” Current [29] T. LAOHAKANGVALVIT, T. ACHALAKUL, and
Dermatology Reports, vol. 8, pp. 85–90, 2019. M. OHKURA, “A method to obtain effective attributes
[17] A. Coustasse, R. Sarkar, B. Abodunde, B. J. Metzger, for attractive cosmetic bottles by deep learning,” In-
and C. M. Slater, “Use of teledermatology to improve ternational Journal of Affective Engineering, vol. 19,
dermatological access in rural areas,” Telemedicine no. 1, pp. 37–48, 2020.
and e-Health, vol. 25, no. 11, pp. 1022–1032, 2019. [30] S. Ray, M. Abinaya, A. K. Rao, S. K. Shukla, S. Gupta,
[18] Y. Goldust, F. Sameem, S. Mearaj, A. Gupta, A. Patil, and P. Rawat, “Cosmetics suggestion system using
and M. Goldust, “Covid-19 and artificial intelligence: deep learning,” in 2022 2nd International Conference
Experts and dermatologists perspective,” Journal of on Technological Advancements in Computational
cosmetic dermatology, vol. 22, no. 1, pp. 11–15, 2023. Sciences (ICTACS). IEEE, 2022, pp. 680–684.
[19] D. R. Crowe, M. B. Morgan, S. Somach, and K. Trapp, [31] A. Nightingale, “A guide to systematic literature re-
Deadly Dermatologic Diseases. Springer, 2016. views,” Surgery (Oxford), vol. 27, no. 9, pp. 381–384,
[20] A. Wells, S. Patel, J. B. Lee, and K. Motaparthi, 2009.
“Artificial intelligence in dermatopathology: Diagno- [32] M. J. Page, J. E. McKenzie, P. M. Bossuyt, I. Boutron,
sis, education, and research,” Journal of Cutaneous T. C. Hoffmann, C. D. Mulrow, L. Shamseer, J. M.
Pathology, vol. 48, no. 8, pp. 1061–1068, 2021. Tetzlaff, E. A. Akl, S. E. Brennan et al., “The prisma
[21] S. Zhang, Y. Wang, Q. Zheng, J. Li, J. Huang, and 2020 statement: an updated guideline for reporting
X. Long, “Artificial intelligence in melanoma: A sys- systematic reviews,” International journal of surgery,
tematic review,” Journal of Cosmetic Dermatology, vol. 88, p. 105906, 2021.
vol. 21, no. 11, pp. 5993–6004, 2022. [33] N. R. Haddaway, M. J. Page, C. C. Pritchard, and
[22] A. Mosquera-Zamudio, L. Launet, Z. Tabatabaei, L. A. McGuinness, “Prisma2020: An r package and
R. Parra-Medina, A. Colomer, J. Oliver Moll, C. Mon- shiny app for producing prisma 2020-compliant flow
teagudo, E. Janssen, and V. Naranjo, “Deep learning diagrams, with interactivity for optimised digital trans-
for skin melanocytic tumors in whole-slide images: parency and open synthesis,” Campbell Systematic
A systematic review,” Cancers, vol. 15, no. 1, p. 42, Reviews, vol. 18, no. 2, p. e1230, 2022.
2022. [34] N. K. Chauhan and K. Singh, “A review on conven-
[23] H. K. Jeong, C. Park, R. Henao, and M. Kheterpal, tional machine learning vs deep learning,” in 2018
“Deep learning in dermatology: a systematic review International conference on computing, power and
of current approaches, outcomes and limitations,” JID communication technologies (GUCON). IEEE, 2018,
Innovations, p. 100150, 2022. pp. 347–352.
[24] M. Alam, “Teaching cosmetic dermatology without [35] P. Wang, E. Fan, and P. Wang, “Comparative analysis
sacrificing the proper emphasis on medical dermatol- of image classification algorithms based on traditional
ogy,” JAMA dermatology, vol. 150, no. 2, pp. 123– machine learning and deep learning,” Pattern Recog-
124, 2014. nition Letters, vol. 141, pp. 61–67, 2021.
[25] S. Sood, M. Jafferany, and S. Vinaya Kumar, “De- [36] G. Hinton, Y. LeCun, and Y. Bengio, “Deep learning,”
pression, psychiatric comorbidities, and psychosocial Nature, vol. 521, no. 7553, pp. 436–444, 2015.
implications associated with acne vulgaris,” Journal [37] H. A. Benson, M. S. Roberts, V. R. Leite-Silva,
of Cosmetic Dermatology, vol. 19, no. 12, pp. 3177– and K. Walters, Cosmetic formulation: principles and
3182, 2020. practice. CRC Press, 2019.
[26] N. Alamdari, K. Tavakolian, M. Alhashim, and [38] S. Sunkle, D. Jain, K. Saxena, A. Patil, T. Singh,
R. Fazel-Rezai, “Detection and classification of acne B. Rai, and V. Kulkarni, “Integrated “generate, make,
lesions in acne patients: A mobile application,” in and test” for formulated products using knowledge
2016 IEEE International Conference on Electro Infor- graphs,” Data Intelligence, vol. 3, no. 3, pp. 340–375,
mation Technology (EIT). IEEE, 2016, pp. 0739– 2021.
0743. [39] X. Zhang, T. Zhou, and K. M. Ng, “Optimization-
16 VOLUME 4, 2016
based cosmetic formulation: Integration of mechanis- potency using machine learning approaches,” Journal
tic model, surrogate model, and heuristics,” AIChE of Applied Toxicology, vol. 37, no. 7, pp. 792–805,
Journal, vol. 67, no. 1, p. e17064, 2021. 2017.
[40] S.-J. Yeh, J.-F. Lin, and B.-S. Chen, “Multiple- [51] J. Strickland, Q. Zang, N. Kleinstreuer, M. Paris,
molecule drug design based on systems biology ap- D. M. Lehmann, N. Choksi, J. Matheson, A. Jacobs,
proaches and deep neural network to mitigate human A. Lowit, D. Allen et al., “Integrated decision strate-
skin aging,” Molecules, vol. 26, no. 11, p. 3178, 2021. gies for skin sensitization hazard,” Journal of Applied
[41] C. Johnson, E. Ahlberg, L. T. Anger, L. Beilke, Toxicology, vol. 36, no. 9, pp. 1150–1162, 2016.
R. Benigni, J. Bercu, S. Bobst, D. Bower, A. Brigo, [52] J. Strickland, Q. Zang, M. Paris, D. M. Lehmann,
S. Campbell et al., “Skin sensitization in silico proto- D. Allen, N. Choksi, J. Matheson, A. Jacobs,
col,” Regulatory Toxicology and Pharmacology, vol. W. Casey, and N. Kleinstreuer, “Multivariate models
116, p. 104688, 2020. for prediction of human skin sensitization hazard,”
[42] A. B. Raies and V. B. Bajic, “In silico toxicology: Journal of Applied Toxicology, vol. 37, no. 3, pp. 347–
computational methods for the prediction of chemical 360, 2017.
toxicity,” Wiley Interdisciplinary Reviews: Computa- [53] H. Li, J. Bai, G. Zhong, H. Lin, C. He, R. Dai,
tional Molecular Science, vol. 6, no. 2, pp. 147–172, H. Du, and L. Huang, “Improved defined approaches
2016. for predicting skin sensitization hazard and potency in
[43] D. Sreedhar, N. Manjula, S. Pise, and V. Ligade, “Ban humans,” ALTEX-Alternatives to animal experimen-
of cosmetic testing on animals: A brief overview,” Int. tation, vol. 36, no. 3, pp. 363–372, 2019.
J. Curr. Res. Rev, vol. 12, p. 113, 2020. [54] V. M. Alves, E. Muratov, D. Fourches, J. Strickland,
[44] N. Gilmour, P. S. Kern, N. Alépée, F. Boislève, N. Kleinstreuer, C. H. Andrade, and A. Tropsha,
D. Bury, E. Clouet, M. Hirota, S. Hoffmann, J. Kühnl, “Predicting chemically-induced skin reactions. part ii:
J. F. Lalko et al., “Development of a next genera- Qsar models of skin permeability and the relationships
tion risk assessment framework for the evaluation of between skin permeability and skin sensitization,”
skin sensitisation of cosmetic ingredients,” Regulatory Toxicology and applied pharmacology, vol. 284, no. 2,
Toxicology and Pharmacology, vol. 116, p. 104721, pp. 273–280, 2015.
2020. [55] S. O. Akturk, G. Tugcu, and H. Sipahi, “Development
[45] J. W. van der Veen, E. Rorije, R. Emter, A. Natsch, of a qsar model to predict comedogenic potential of
H. van Loveren, and J. Ezendam, “Evaluating the some cosmetic ingredients,” Computational Toxicol-
performance of integrated approaches for hazard iden- ogy, vol. 21, p. 100207, 2022.
tification of skin sensitizing chemicals,” Regulatory [56] N. Sharma, S. Patiyal, A. Dhall, N. L. Devi, and G. P.
Toxicology and Pharmacology, vol. 69, no. 3, pp. 371– Raghava, “Chalpred: A web server for prediction of
379, 2014. allergenicity of chemical compounds,” Computers in
[46] J. S. Jaworska, A. Natsch, C. Ryan, J. Strickland, Biology and Medicine, vol. 136, p. 104746, 2021.
T. Ashikaga, and M. Miyazawa, “Bayesian integrated [57] A. Wilm, M. Garcia de Lomana, C. Stork, N. Mathai,
testing strategy (its) for skin sensitization potency S. Hirte, U. Norinder, J. Kühnl, and J. Kirchmair,
assessment: a decision support system for quantita- “Predicting the skin sensitization potential of small
tive weight of evidence and adaptive testing strategy,” molecules with machine learning models trained on
Archives of toxicology, vol. 89, pp. 2355–2383, 2015. biologically meaningful descriptors,” Pharmaceuti-
[47] T. Luechtefeld, A. Maertens, J. M. McKim, T. Har- cals, vol. 14, no. 8, p. 790, 2021.
tung, A. Kleensang, and V. Sá-Rocha, “Probabilistic [58] B. Jeon, M. H. Lim, T. H. Choi, B.-C. Kang, and
hazard assessment for skin sensitization potency by S. Kim, “A development of a graph-based ensemble
dose–response modeling using feature elimination in- machine learning model for skin sensitization hazard
stead of quantitative structure–activity relationships,” and potency assessment,” Journal of Applied Toxicol-
Journal of Applied Toxicology, vol. 35, no. 11, pp. ogy, vol. 42, no. 11, pp. 1832–1842, 2022.
1361–1371, 2015. [59] A. Forreryd, U. Norinder, T. Lindberg, and
[48] D. Asturiol, S. Casati, and A. Worth, “Consensus of M. Lindstedt, “Predicting skin sensitizers with
classification trees for skin sensitisation hazard predic- confidence—using conformal prediction to determine
tion,” Toxicology in Vitro, vol. 36, pp. 197–209, 2016. applicability domain of gard,” Toxicology in Vitro,
[49] A. P. Toropova and A. A. Toropov, “Hybrid optimal vol. 48, pp. 179–187, 2018.
descriptors as a tool to predict skin sensitization in [60] A. K. Singh, “Chapter 7 - mechanisms of
accordance to oecd principles,” Toxicology Letters, nanoparticle toxicity,” in Engineered Nanoparticles,
vol. 275, pp. 57–66, 2017. A. K. Singh, Ed. Boston: Academic Press,
[50] Q. Zang, M. Paris, D. M. Lehmann, S. Bell, N. Klein- 2016, pp. 295–341. [Online]. Available:
streuer, D. Allen, J. Matheson, A. Jacobs, W. Casey, https://www.sciencedirect.com/science/article/pii/B9780128014
and J. Strickland, “Prediction of skin sensitization [61] F. Hu, S. F. Santagostino, D. M. Danilenko, M. Tseng,
VOLUME 4, 2016 17
J. Brumm, P. Zehnder, and K. C. Wu, “Assessment biology and medicine, vol. 57, pp. 173–181, 2015.
of skin toxicity in an in vitro reconstituted human [73] M. Ko, D. Kim, M. Kim, and K. Kim, “Illumination-
epidermis model using deep learning,” The American insensitive skin depth estimation from a light-field
Journal of Pathology, vol. 192, no. 4, pp. 687–700, camera based on cgans toward haptic palpation,” Elec-
2022. tronics, vol. 7, no. 11, p. 336, 2018.
[62] I. Furxhi, F. Murphy, M. Mullins, and C. A. Poland, [74] J. Sánchez-Monedero, A. Sáez, M. Pérez-Ortiz, P. A.
“Machine learning prediction of nanoparticle in vitro Gutiérrez, and C. Hervás-Martínez, “Classification of
toxicity: A comparative study of classifiers and melanoma presence and thickness based on computa-
ensemble-classifiers using the copeland index,” Tox- tional image analysis,” in Hybrid Artificial Intelligent
icology letters, vol. 312, pp. 157–166, 2019. Systems: 11th International Conference, HAIS 2016,
[63] S. Y. Jun and H. S. Shin, “Analysis and prediction of Seville, Spain, April 18-20, 2016, Proceedings 11.
surface condition of artificial skin based on cnn and Springer, 2016, pp. 427–438.
convlstm,” Biotechnology and Bioprocess Engineer- [75] D. Kucharski, P. Kleczek, J. Jaworek-Korjakowska,
ing, vol. 26, pp. 369–374, 2021. G. Dyduch, and M. Gorgon, “Semi-supervised nests
[64] E. Chirikhina, A. Chirikhin, S. Dewsbury-Ennis, of melanocytes segmentation method using convolu-
F. Bianconi, and P. Xiao, “Skin characterizations by tional autoencoders,” Sensors, vol. 20, no. 6, p. 1546,
using contact capacitive imaging and high-resolution 2020.
ultrasound imaging with machine learning algo- [76] I. Abunadi and E. M. Senan, “Deep learning and
rithms,” Applied Sciences, vol. 11, no. 18, p. 8714, machine learning techniques of diagnosis dermoscopy
2021. images for early detection of skin diseases,” Electron-
[65] R. Zegour, A. Belaid, J. Ognard, and D. B. Salem, ics, vol. 10, no. 24, p. 3158, 2021.
“Convolutional neural networks-based method for [77] F. W. Alsaade, T. H. Aldhyani, and M. H. Al-Adhaileh,
skin hydration measurements in high resolution mri,” “Developing a recognition system for diagnosing
Biomedical Signal Processing and Control, vol. 81, p. melanoma skin lesions using artificial intelligence al-
104491, 2023. gorithms,” Computational and mathematical methods
[66] C. Sreeja, S. S. Sumaya, B. Madhumitha, M. Jayaraj, in medicine, vol. 2021, pp. 1–20, 2021.
S. Francis, and S. Muthukumar, “Odland bodies: A [78] S. Boumaraf, X. Liu, Y. Wan, Z. Zheng, C. Ferkous,
review,” Journal of Advanced Medical and Dental X. Ma, Z. Li, and D. Bardou, “Conventional machine
Sciences Research, vol. 8, no. 11, pp. 12–15, 2020. learning versus deep learning for magnification depen-
[67] K. Koseki, H. Kawasaki, T. Atsugi, M. Nakanishi, dent histopathological breast cancer image classifica-
M. Mizuno, E. Naru, T. Ebihara, M. Amagai, and tion: A comparative study with visual explanation,”
E. Kawakami, “Assessment of skin barrier function Diagnostics, vol. 11, no. 3, p. 528, 2021.
using skin images with topological data analysis,” NPJ [79] Y. Huang, J. He, S. Zhang, Y. Tang, B. Wang, D. Jian,
systems biology and applications, vol. 6, no. 1, p. 40, H. Xie, J. Li, F. Chen, and Z. Zhao, “A novel multi-
2020. layer perceptron model for assessing the diagnos-
[68] S. Borade, D. Kalbande, K. Pereira, R. Patel, and tic value of non-invasive imaging instruments for
S. Kulkarni, “Deep scattering convolutional network rosacea,” PeerJ, vol. 10, p. e13917, 2022.
for cosmetic skin classification,” International Journal [80] A. K. Sameera, V. Samuktha, T. Akash, M. Sabeshnav,
of Engineering Trends and Technology, vol. 70, no. 7, and S. Veni, “Real time detection of the various sign
p. 10–23, 2022. of ageing using deep learning,” in 2022 International
[69] A. Kothari, D. Shah, T. Soni, and S. Dhage, “Cos- Conference on Wireless Communications Signal Pro-
metic skin type classification using cnn with product cessing and Networking (WiSPNET). IEEE, 2022,
recommendation,” in 2021 12th International Confer- pp. 38–43.
ence on Computing Communication and Networking [81] L. Liu, C. Liang, Y. Xue, T. Chen, Y. Chen, Y. Lan,
Technologies (ICCCNT). IEEE, 2021, pp. 1–6. J. Wen, X. Shao, and J. Chen, “An intelligent diag-
[70] A. Firooz, A. Rajabi-Estarabadi, H. Zartab, nostic model for melasma based on deep learning and
N. Pazhohi, F. Fanian, and L. Janani, “The influence multimode image input,” Dermatology and Therapy,
of gender and age on the thickness and echo-density vol. 13, no. 2, pp. 569–579, 2023.
of skin,” Skin Research and Technology, vol. 23, [82] A. Goldsberry, C. W. Hanke, and K. E. Hanke, “Visia
no. 1, pp. 13–20, 2017. system: a possible tool in the cosmetic practice.” Jour-
[71] C. L. Huang and A. C. Halpern, “Management of the nal of drugs in dermatology: JDD, vol. 13, no. 11, pp.
patient with melanoma,” Cancer of the Skin, pp. 265– 1312–1314, 2014.
275, 2005. [83] S. Aditya, S. Sidhu, and M. Kanchana, “Prediction of
[72] S. Vyas, J. Meyerle, and P. Burlina, “Non-invasive alopecia areata using machine learning techniques,” in
estimation of skin thickness from hyperspectral imag- 2022 IEEE International Conference on Data Science
ing and validation using echography,” Computers in and Information System (ICDSIS). IEEE, 2022, pp.
18 VOLUME 4, 2016
1–6. [95] M. Kim and M. H. Song, “High performing facial

[84] A. Alagić, S. Alihodžić, N. Alispahić, E. Bečić, skin problem diagnosis with enhanced mask r-cnn and
A. Smajović, F. Bečić, L. S. Bećirović, L. G. Pokvić, super resolution gan,” Applied Sciences, vol. 13, no. 2,
and A. Badnjević, “Application of artificial intelli- p. 989, 2023.
gence in the analysis of the facial skin health condi- [96] J.-I. Jeong, D.-S. Park, J.-E. Koo, W.-S. Song, D.-
tion,” IFAC-PapersOnLine, vol. 55, no. 4, pp. 31–37, J. Pae, and H.-J. Choi, “Artificial intelligence (ai)
2022. based system for the diagnosis and classification of
[85] K. Dubey, V. Srivastava, and D. S. Mehta, “Auto- scalp health: Ai-scalpgrader,” Instrumentation Science
mated in vivo identification of fungal infection on & Technology, pp. 1–11, 2022.
human scalp using optical coherence tomography and [97] N. Hayashi, H. Akamatsu, M. Kawashima, and A. S.
machine learning,” Laser Physics, vol. 28, no. 4, p. Group, “Establishment of grading criteria for acne
045602, 2018. severity,” The Journal of dermatology, vol. 35, no. 5,
[86] P. Jansen, D. O. Baguer, N. Duschner, J. Le’Clerc Ar- pp. 255–260, 2008.
rastia, M. Schmidt, B. Wiepjes, D. Schadendorf, [98] M. H. Yap, N. Batool, C.-C. Ng, M. Rogers, and
E. Hadaschik, P. Maass, J. Schaller et al., “Evaluation K. Walker, “A survey on facial wrinkles detection and
of a deep learning approach to differentiate bowen’s inpainting: datasets, methods, and challenges,” IEEE
disease and seborrheic keratosis,” Cancers, vol. 14, Transactions on Emerging Topics in Computational
no. 14, p. 3518, 2022. Intelligence, vol. 5, no. 4, pp. 505–519, 2021.
[87] Y. Wang, M. Sun, Y. Duan et al., “Metagenomic [99] J. Rew, H. Kim, and E. Hwang, “Hybrid segmentation
sequencing analysis for acne using machine learning scheme for skin features extraction using dermoscopy
methods adapted to single or multiple data,” Compu- images,” Computers, Materials and Continua, vol. 69,
tational and Mathematical Methods in Medicine, vol. no. 1, pp. 801–817, 2021.
2021, 2021. [100] M. I. M. Ismail and A. N. Sung, “Acne lesion and
[88] F. S. Abas, B. Kaffenberger, J. Bikowski, and M. N. wrinkle detection using faster r-cnn with resnet-50,” in
Gurcan, “Acne image analysis: lesion localization and AIP Conference Proceedings, vol. 2676, no. 1. AIP
classification,” in Medical imaging 2016: computer- Publishing LLC, 2022, p. 020007.
aided diagnosis, vol. 9785. SPIE, 2016, pp. 64–72. [101] C. Shih, C.-H. Lin, Y.-C. Weng, J. Y. Jian, and Y.-
[89] Y. Yang, Y. Ge, L. Guo, Q. Wu, L. Peng, E. Zhang, C. Lin, “Implementation of weakly supervised vitiligo
J. Xie, Y. Li, and T. Lin, “Development and validation treatment evaluation system,” in 2022 IEEE 4th Eura-
of two artificial intelligence models for diagnosing sia Conference on Biomedical Engineering, Health-
benign, pigmented facial skin lesions,” Skin Research care and Sustainability (ECBIOS). IEEE, 2022, pp.
and Technology, vol. 27, no. 1, pp. 74–79, 2021. 44–47.
[90] A. K. Huong, K. Tay, and X. T. NGU, “Customized [102] A. Gallucci, D. Znamenskiy, N. Pezzotti, and
alexnet models for automatic classification of skin M. Petkovic, “Hair counting with deep learning,” in
diseases,” Journal of Engineering Science and Tech- 2020 International Conference on Biomedical Innova-
nology, vol. 16, no. 4, pp. 3312–3324, 2021. tions and Applications (BIA). IEEE, 2020, pp. 5–9.
[91] J. A. López-Leyva, E. Guerra-Rosas, and J. Álvarez- [103] D. T. Phan, Q. B. Ta, T. C. Huynh, T. H. Vo, C. H.
Borrego, “Multi-class diagnosis of skin lesions using Nguyen, S. Park, J. Choi, and J. Oh, “A smart led
the fourier spectral information of images on additive therapy device with an automatic facial acne vulgaris
color model by artificial neural network,” IEEE Ac- diagnosis based on deep learning and internet of things
cess, vol. 9, pp. 35 207–35 216, 2021. application,” Computers in Biology and Medicine, vol.
[92] A. Jain, A. C. S. Rao, P. K. Jain, and A. Abraham, 136, p. 104610, 2021.
“Multi-type skin diseases classification using op-dnn [104] H. Wen, W. Yu, Y. Wu, J. Zhao, X. Liu, Z. Kuang, and
based feature extraction approach,” Multimedia Tools R. Fan, “Acne detection and severity evaluation with
and Applications, pp. 1–26, 2022. interpretable convolutional neural network models,”
[93] H. Ito, Y. Nakamura, K. Takanari, M. Oishi, K. Mat- Technology and Health Care, vol. 30, no. S1, pp. 143–
suo, M. Kanbe, T. Uchibori, K. Ebisawa, and 153, 2022.
Y. Kamei, “Development of a novel scar screening [105] L. Maknuna, H. Kim, Y. Lee, Y. Choi, H. Kim, M. Yi,
system with machine learning,” Plastic and Recon- and H. W. Kang, “Automated structural analysis and
structive Surgery, vol. 150, no. 2, pp. 465e–472e, quantitative characterization of scar tissue using ma-
2022. chine learning,” Diagnostics, vol. 12, no. 2, p. 534,
[94] S. Borade, D. Kalbande, H. Jakaria, and L. Patil, “An 2022.
automated approach to detect & diagnosis the type of [106] J. Redmon, S. Divvala, R. Girshick, and A. Farhadi,
cosmetic skin & its disease using machine learning,” in “You only look once: Unified, real-time object de-
2022 IEEE 3rd Global Conference for Advancement tection,” in Proceedings of the IEEE conference on
in Technology (GCAT). IEEE, 2022, pp. 1–8. computer vision and pattern recognition, 2016, pp.
VOLUME 4, 2016 19
779–788. and Technology, vol. 26, no. 2, pp. 169–186, 2020.

[107] Y.-H. Liao, P.-C. Chang, C.-C. Wang, and H.-H. [117] J. Wang, Y. Luo, Z. Wang, A. H. Hounye, C. Cao,
Li, “An optimization-based technology applied for M. Hou, and J. Zhang, “A cell phone app for fa-
face skin symptom detection,” in Healthcare, vol. 10, cial acne severity assessment,” Applied Intelligence,
no. 12. Multidisciplinary Digital Publishing Institute, vol. 53, no. 7, pp. 7614–7633, 2023.
2022, p. 2396. [118] D. Wang, X. Chen, Y. Wu, H. Tang, and P. Deng, “Arti-
[108] H. Ding, E. Zhang, F. Fang, X. Liu, H. Zheng, ficial intelligence for assessing the severity of microtia
H. Yang, Y. Ge, Y. Yang, and T. Lin, “Automatic via deep convolutional neural networks,” Frontiers in
identification of benign pigmented skin lesions from Surgery, p. 1385, 2022.
clinical images using deep convolutional neural net- [119] Q. Man, L. Zhang, and Y. Cho, “Efficient hair damage
work,” BMC biotechnology, vol. 22, no. 1, p. 28, 2022. detection using sem images based on convolutional
[109] A. Li, R. Fang, and Q. Sun, “Artificial intelligence neural network,” Applied Sciences, vol. 11, no. 16, p.
for grading in acne vulgaris: current situation and 7333, 2021.
prospect,” Journal of Cosmetic Dermatology, vol. 21, [120] W.-J. Chang, L.-B. Chen, M.-C. Chen, Y.-C. Chiu,
no. 2, pp. 865–866, 2022. and J.-Y. Lin, “Scalpeye: A deep learning-based scalp
[110] J. K. Tan, “Current measures for the evaluation of acne hair inspection and diagnosis system for scalp health,”
severity,” Expert Review of Dermatology, vol. 3, no. 5, IEEE Access, vol. 8, pp. 134 826–134 837, 2020.
pp. 595–603, 2008. [121] S. Tian, W. Yang, J. M. Le Grange, P. Wang,
[111] K. Zarchi and G. B. Jemec, “Severity assessment and W. Huang, and Z. Ye, “Smart healthcare: making
outcome measures in acne vulgaris,” Current Derma- medical care more intelligent,” Global Health Journal,
tology Reports, vol. 1, pp. 131–136, 2012. vol. 3, no. 3, pp. 62–65, 2019.
[112] R. Jiang, I. Kezele, A. Levinshtein, F. Flament, [122] C. S. Pattichis, C. Pitris, J. Liang, and Y. Zhang,
J. Zhang, E. Elmoznino, J. Ma, H. Ma, J. Coquide, “Guest editorial on the special issue on integrating
V. Arcin et al., “A new procedure, free from human informatics and technology for precision medicine,”
assessment that automatically grades some facial skin IEEE Journal of Biomedical and Health Informatics,
structural signs. comparison with assessments by ex- vol. 23, no. 1, pp. 12–13, 2019.
perts, using referential atlases of skin ageing,” Inter- [123] W.-S. Huang, B.-K. Hong, W.-H. Cheng, S.-W. Sun,
national Journal of Cosmetic Science, vol. 41, no. 1, and K.-L. Hua, “A cloud-based intelligent skin and
pp. 67–78, 2019. scalp analysis system,” in 2018 IEEE Visual Commu-
[113] F. Flament, Y. W. Lee, D. H. Lee, T. Passeron, nications and Image Processing (VCIP). IEEE, 2018,
Y. Zhang, R. Jiang, A. Prunel, S. Dwivedi, C. Kroely, pp. 1–5.
Y. J. Park et al., “The continuous development of a [124] X. Liu, C.-H. Chen, M. Karvela, and C. Toumazou,
complete and objective automatic grading system of “A dna-based intelligent expert system for person-
facial signs from selfie pictures: Asian validation study alised skin-health recommendations,” IEEE journal of
and application to women of three ethnic origins, dif- biomedical and health informatics, vol. 24, no. 11, pp.
ferently aged,” Skin Research and Technology, vol. 27, 3276–3284, 2020.
no. 2, pp. 183–190, 2021. [125] Y. Zhang, J. Zhou, and W. Lin, “Knowledge
[114] F. Flament, L. Jacquet, C. Ye, D. Amar, D. Kerob, graph-based sequence recommendation,” in 2021
R. Jiang, Y. Zhang, C. Kroely, C. Delaunay, and 8th International Conference on Computational Sci-
T. Passeron, “Artificial intelligence analysis of over ence/Intelligence and Applied Informatics (CSII).
half a million european and chinese women reveals IEEE, 2021, pp. 12–17.
striking differences in the facial skin ageing process,” [126] S. A. A. Shah, M. Bennamoun, and M. Molton, “A
Journal of the European Academy of Dermatology and fully automatic framework for prediction of 3d facial
Venereology, vol. 36, no. 7, pp. 1136–1142, 2022. rejuvenation,” in 2018 International Conference on
[115] F. Flament, R. Jiang, J. Houghton, Y. Zhang, Image and Vision Computing New Zealand (IVCNZ).
C. Kroely, N. G. Jablonski, A. Jean, J. Clarke, J. Steeg, IEEE, 2018, pp. 1–6.
C. Sehgal et al., “Accuracy and clinical relevance [127] S. A. A. Shah, M. Bennamoun, and M. K. Molton,
of an automated, algorithm-based analysis of facial “Machine learning approaches for prediction of facial
signs from selfie images of women in the united rejuvenation using real and synthetic data,” IEEE Ac-
states of various ages, ancestries and phototypes: A cess, vol. 7, pp. 23 779–23 787, 2019.
cross-sectional observational study,” Journal of the [128] T.-Y. Lin, Y.-T. Tsai, T.-S. Huang, W.-C. Lin, and J.-H.
European Academy of Dermatology and Venereology, Chuang, “Exemplar-based freckle retouching and skin
vol. 37, no. 1, pp. 176–183, 2023. tone adjustment,” Computers & Graphics, vol. 78, pp.
[116] A. Seck, H. Dee, W. Smith, and B. Tiddeman, “3d sur- 54–63, 2019.
face texture analysis of high-resolution normal fields [129] S. B. Akben, “Predicting the success of wart treatment
for facial skin condition assessment,” Skin Research methods using decision tree based fuzzy informative
20 VOLUME 4, 2016
images,” biocybernetics and biomedical engineering, SIRAWICH VACHMANUS received the M.E.
vol. 38, no. 4, pp. 819–827, 2018. and Ph.D. degrees from Hokkaido University,
Japan, in 2019 and 2022, respectively. He is cur-
[130] K. Erdoǧan, O. Acun, A. Küçükmanísa, R. Duvar, rently a lecturer at the Faculty of Information and
A. Bayramoǧlu, and O. Urhan, “Kebot: an artificial Communication Technology, Mahidol University,
intelligence based comprehensive analysis system for Thailand. His research interests include computer
fue based hair transplantation,” IEEE Access, vol. 8, vision, deep learning, machine learning, sensor
fusion, and artificial intelligence for robots.
pp. 200 461–200 476, 2020.
[131] C. Shi, Z. Jiang, X. Ma, and Q. Luo, “A personal-
ized visual aid for selections of appearance building
products with long-term effects,” in Proceedings of the
2022 CHI Conference on Human Factors in Comput-
ing Systems, 2022, pp. 1–18.
[132] S. S. K and D. S. Krishnaveni, “An analytical approach
for reconstruction of cosmetic surgery images using
euclbp and sift,” Sep 2022. [Online]. Available:
https://doi.org/10.14445/23488379/IJEEE-V9I8P107
[133] İ. Bahçeci Şimşek and C. Şirolu, “Analysis of surgical
outcome after upper eyelid surgery by computer vision
algorithm using face and facial landmark detection,” THANAPON NORASET is a faculty member at
Graefe’s Archive for Clinical and Experimental Oph- the Faculty of Information and Communication
Technology, Mahidol University, Thailand. He re-
thalmology, vol. 259, no. 10, pp. 3119–3125, 2021.
ceived his BSc degree from the same faculty in
[134] J. Kim, D.-U. Hwang, E. J. Son, S. H. Oh, W. Kim, 2007. He received his Ph.D. degree in Computer
Y. Kim, and G. Kwon, “Emotion recognition while Science from Northwestern University, USA, in
applying cosmetic cream using deep learning from eeg 2017. His research work and interests are in the
field of natural language processing and machine
data; cross-subject analysis,” Plos one, vol. 17, no. 11,
learning.
p. e0274203, 2022.
[135] X. Chen, X. Wang, K. Zhang, K.-M. Fung, T. C. Thai,
K. Moore, R. S. Mannel, H. Liu, B. Zheng, and Y. Qiu,
“Recent advances and clinical applications of deep
learning in medical image analysis,” Medical Image
Analysis, p. 102444, 2022.
[136] F. Kaliyadan and K. T. Ashique, “Use of mobile appli-
cations in dermatology,” Indian Journal of Dermatol-
ogy, vol. 65, no. 5, p. 371, 2020.
[137] L. J. Caffery, M. Janda, R. Miller, L. M. Abbott,
C. Arnold, T. Caccetta, P. Guitera, S. Shumack,
P. Fernández-Peñas, V. Mar et al., “Informing a po-
sition statement on the use of artificial intelligence SUPPAWONG TUAROB received his PhD in
computer science and engineering and MS in in-
in dermatology in australia,” Australasian Journal of dustrial engineering both from the Pennsylvania
Dermatology, vol. 64, no. 1, pp. e11–e20, 2023. State University and his BSE and MSE both in
computer science and engineering from the Uni-
versity of Michigan-Ann Arbor. Currently, he is
an Associate Professor of Computer Science at
the Faculty of Information and Communication
Technology, Mahidol University, Thailand. His re-
PAT VATIWUTIPONG received a Master’s De- search involves data mining in large-scale schol-
gree in Applied Mathematics from Mahidol Uni- arly, social media, and healthcare domains, as well as applications of
versity, Thailand, in 2017. Then, he granted an intelligent technologies for social good.
Erasmus Mundus full scholarship to study a joint
program in Europe. He received two Master’s De-
grees there, Laurea Magistrale in Mathematical
Modelling, issued by the University of L’Aquila,
Italy, and Master Mathématiques issued by the
Université Côte d’Azur, France, in 2019. From
2019 to 2020, he worked as an assistant lecturer
at the Department of Mathematics, Faculty of Science, Mahidol University.
His research interests include mathematical modeling, time series analysis,
and computational mathematics applied in natural science.
VOLUME 4, 2016 21

Artificial Intelligence in Cosmetic Dermatology A

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Artificial Intelligence in Cosmetic Dermatology A

Uploaded by

Copyright:

Available Formats

This article has been accepted for publication in IEEE Access.

Artificial Intelligence in Cosmetic

I. INTRODUCTION revolution of the 19th century and has continued to undergo

P. Vatiwutipong et al.: Preparation of Papers for IEEE TRANSACTIONS and JOURNALS

P. Vatiwutipong et al.: Preparation of Papers for IEEE TRANSACTIONS and JOURNALS

P. Vatiwutipong et al.: Preparation of Papers for IEEE TRANSACTIONS and JOURNALS

FIGURE 1: The PRISMA flow diagram of the study.

TABLE 1: Number of papers of each category

P. Vatiwutipong et al.: Preparation of Papers for IEEE TRANSACTIONS and JOURNALS

FIGURE 4: Number of papers by country

IV. ANALYSIS OF REVIEWED PAPERS

P. Vatiwutipong et al.: Preparation of Papers for IEEE TRANSACTIONS and JOURNALS

P. Vatiwutipong et al.: Preparation of Papers for IEEE TRANSACTIONS and JOURNALS

P. Vatiwutipong et al.: Preparation of Papers for IEEE TRANSACTIONS and JOURNALS

P. Vatiwutipong et al.: Preparation of Papers for IEEE TRANSACTIONS and JOURNALS

P. Vatiwutipong et al.: Preparation of Papers for IEEE TRANSACTIONS and JOURNALS

P. Vatiwutipong et al.: Preparation of Papers for IEEE TRANSACTIONS and JOURNALS

P. Vatiwutipong et al.: Preparation of Papers for IEEE TRANSACTIONS and JOURNALS

P. Vatiwutipong et al.: Preparation of Papers for IEEE TRANSACTIONS and JOURNALS

P. Vatiwutipong et al.: Preparation of Papers for IEEE TRANSACTIONS and JOURNALS

P. Vatiwutipong et al.: Preparation of Papers for IEEE TRANSACTIONS and JOURNALS

P. Vatiwutipong et al.: Preparation of Papers for IEEE TRANSACTIONS and JOURNALS

P. Vatiwutipong et al.: Preparation of Papers for IEEE TRANSACTIONS and JOURNALS

P. Vatiwutipong et al.: Preparation of Papers for IEEE TRANSACTIONS and JOURNALS

P. Vatiwutipong et al.: Preparation of Papers for IEEE TRANSACTIONS and JOURNALS

1–6. [95] M. Kim and M. H. Song, “High performing facial

P. Vatiwutipong et al.: Preparation of Papers for IEEE TRANSACTIONS and JOURNALS

779–788. and Technology, vol. 26, no. 2, pp. 169–186, 2020.

P. Vatiwutipong et al.: Preparation of Papers for IEEE TRANSACTIONS and JOURNALS

You might also like