Transforming Aesthetic Dermatology: The Role of Artificial Intelligence in Skin Health

Abstract

Artificial intelligence (AI) has profoundly impacted various medical fields, including dermatology, which is to a large extent an image-based discipline. Initially applied in general dermatology, AI has expanded into aesthetic dermatology, where it assists in integrating objective approaches with the aesthetic judgment of dermatologists. This review explores the role of AI in aesthetic dermatology, focusing on skin condition assessment, diagnosis, treatment optimization, and patient education, as well as the challenges and future directions in this evolving field.

Key Summary Points

Why carry out this study?

Traditional aesthetic dermatology relies on subjective evaluations and lacks objective quantification tools, which Artifical Intelligence (AI) can address through data-driven solutions.

Although AI has shown preliminary applications in aesthetic dermatology, comprehensive reviews are lacking. This review addresses this gap by evaluating AI’s potential in this field.

What was learned from this study?

AI plays a critical role in aesthetic dermatology, encompassing skin condition analysis, diagnostic support, personalized treatment optimization, and patient education.

AI enhances treatment innovations spanning optimized phototherapy regimens, personalized injection protocols, and automated hair density quantification.

Critical challenges still remain to be addressed, including improving model generalizability and explainability, establishing unified evaluation standards, ensuring data diversity and multimodal interoperability, and addressing security concerns.

Introduction

Aesthetic dermatology is a rapidly evolving discipline, driven by changing aesthetic standards, expanding market demand, and advancing technology. Unlike general dermatology, which addresses conditions with systemic or even life-threatening implications, aesthetic dermatology focuses on skin appearance and overall skin health, which directly impacts individuals’ psychological well-being through enhanced self-perception and social belongingness [1]. To address these concerns effectively, aesthetic dermatologists must integrate medical expertise with a refined sense of aesthetics, yet the variability in individual aesthetic standards means that the processes of aesthetic assessment, diagnosis, treatment selection, and postprocedure evaluation often lack universally accepted criteria and are subject to personal judgment.

Artificial intelligence (AI) refers to the simulation of human cognitive functions by machines, enabling them to perform tasks that typically require human intelligence, such as learning and decision-making [2]. In dermatology, a specialty inherently reliant on image-based analysis, the further development of deep learning, especially convolutional neural networks (CNN), enables AI to handle clinical, dermoscopic, and histopathological images—and even self-captured images. In general dermatology, a landmark study published in Nature (2017) demonstrated deep neural networks’ efficacy in classifying skin cancer, achieving performance comparable to dermatologists [3]. This breakthrough in AI applications provides a foundation for its integration into aesthetic dermatology, where it may combine objective, science-backed methodologies with intuitive creativity and aesthetic judgment. AI models trained on specific datasets may objectively assess skin biophysical parameters and diagnose conditions, replacing traditional approaches that rely on semi-quantitative grading scales, costly equipment, or subjective clinical judgment. In addition, AI can provide data-driven recommendations for various aesthetic applications such as filler volume calculations [4] and treatment path planning in phototherapy [5].

We conducted our review by searching PubMed, IEEE Xplore, and Web of Science (2015–present), as it involves the intersection of medicine and computer science. The search strategies for PubMed included: (((Artificial intelligence) OR (machine learning) OR (deep learning) OR (neural network))) AND ((((Phototherapy) OR (Filler) OR (Injection) OR (Hair) OR (plastic surgery)) OR ((Vitiligo) OR (Pigmentary disease) OR (Acne Vulgaris)) OR ((Cosmetic Dermatology) OR (Aesthetic Dermatology)))). This article is based on previously conducted studies and does not contain any new studies with human participants or animals performed by any of the authors.

AI Fundamentals in Dermatology

In this section, AI fundamentals relevant to dermatology will be discussed, as summarized in Fig. 1. While these topics might not be at the top of a clinician’s to-do list, we aim to provide an accessible introduction to the concepts of artificial intelligence, machine learning (ML), neural networks (NN), and deep learning (DL), along with their basic functions, offering a foundational understanding for colleagues and researchers interested in engaging with this field.

Fig. 1.

AI fundamentals and structure of neural networks. While often conflated, artificial intelligence (AI), machine learning (ML), neural networks (NNs), and deep learning (DL) form a hierarchical continuum: ML specializes AI, NNs advance ML, and DL employs multilayered NNs—notably convolutional neural networks (CNNs), generative adversarial networks (GANs), and transformers—for complex pattern recognition. Neural networks process data through biologically inspired computational layers. Each node applies activation functions to weighted inputs, propagating refined signals to subsequent layers for progressive feature abstraction. Training constitutes dynamic parameter optimization via backpropagation, where weights/biases are iteratively adjusted to minimize prediction errors

Artificial intelligence is often used interchangeably with machine learning or deep learning. However, machine learning is a subfield of AI, while deep learning is a further specialization within it. Machine learning focuses on developing algorithms and statistical models that allow machines to learn from data without explicit programming. The key approaches include supervised, unsupervised, semi-supervised, and reinforcement learning, with supervised and unsupervised learning being the most commonly utilized [6]. Unsupervised learning aims to uncover latent patterns in unlabeled data, with clustering being one of the most widely used techniques. However, owing to the lack of labels, the results may lack practical or clinical significance and require careful interpretation. In contrast, supervised learning leverages labeled datasets, which are typically annotated by experts. The performance of supervised models heavily depends on the quality of the training data, as any systematic errors in labeling will be reflected in the model’s output. Classification is one of the most common tasks in supervised learning, aiming to assign data points to predefined categories. Among the various methods used for classification, decision trees are widely recognized for their intuitiveness and interpretability. A decision tree constructs its decision paths through step-by-step feature-based splits, where each node represents a feature condition and each branch corresponds to a decision path. The process continues until a leaf node is reached, producing the final classification outcome.

Neural networks can be seen as an extension of node-based decision logic. They consist of interconnected nodes (neurons) that mimic the structure and function of the biological nervous system [7]. Each neuron receives input signals, applies an activation function, and produces an output signal, which is passed to the next layer. The network is composed of multiple layers of neurons, with each layer extracting features of increasing complexity to gradually process the data. The training process of a neural network is fundamentally a dynamic optimization task. Using the backpropagation algorithm, the model adjusts the weights and biases of its connections on the basis of the error between its predictions and the true labels, aiming to minimize this error. When a neural network’s layers and nodes increase significantly, it surpasses shallow networks in handling complex tasks. This forms the essence of deep learning, where hierarchical architectures progressively refine features layer by layer. Starting with basic edge detection, these layers evolve into advanced semantic pattern recognition, ultimately forming robust representations for complex data analysis.

In the medical field, several common neural network architectures have proven to excel in specific tasks. CNNs are specialized for processing visual data by detecting patterns in images, similar to how the human visual cortex functions. Using convolutional layers, CNNs apply filters to extract local features, enabling tasks such as image segmentation and recognition. For example, segmentation algorithms such as U-Net or R-CNN can identify and label regions of interest, such as pigmentation or acne lesions, distinguishing healthy skin from abnormal areas.

Generative adversarial networks (GANs) operate through a competitive process between a generator and a discriminator. By generating synthetic skin images, such as those representing rare skin conditions, GANs enhance dataset diversity and improve model robustness.

Transformer models (notably GPT-based architectures and the more recent DeepSeek) excel at capturing long-range dependencies in data. They can assist in patient education, providing explanations of skin conditions and treatment plans to patients either before or after consultations [8]. Transformers can also support physicians by extracting key information from patient histories or generating personalized treatment recommendations, but patient data privacy and security must be carefully considered. These architectures will collectively form the foundation of multi-modal AI systems, integrating visual and textual data to deliver comprehensive AI-driven insights based on clinical information and image analysis for dermatology.

Skin Measurements in Aesthetic Consultations

The skin assessment process involves evaluating basic skin properties, such as hydration, oiliness, and thickness to assess the suitability of the patient’s skin for aesthetic procedures such as phototherapy or filler injections.

Skin hydration is a key indicator of skin barrier function, with low levels often linked to dryness, sensitivity, and aging. Traditional methods for assessing skin hydration, such as electrical capacitance measurement, trans-epidermal water loss (TEWL) measurements, high-frequency ultrasound, and optical techniques, are widely used in clinical practice and offer distinct advantages, but they also have potential limitations. Emerging AI-based approaches present a promising alternative with unique capabilities for hydration assessment. Topological data analysis is employed to extract skin texture features, which have been shown to correlate with TEWL and are subsequently used as inputs for machine-learning models to predict TEWL levels [9]. Furthermore, Zhang et al. employed AlexNet as a classification model to directly predict skin hydration levels from in vivo capacitive skin images, focusing on features such as skin texture and tone [10]. However, the sample size was small, consisting of only 1000 skin capacitive images from four healthy volunteers, with images taken from different skin areas. A larger sample size may be needed to improve generalizability. Another study developed an AI model using visible-light skin images and feature factors, significantly improving the prediction of skin moisture while reducing equipment requirements [11]. However, this relaxation in requirements emphasized the importance of selecting relevant input features, with combinations of brown spot count, pore count, and skin roughness showing strong correlations with moisture levels.

Like skin hydration, skin type also plays a crucial role in personalized dermatological treatments. The Baumann Skin Type Indicator (BSTI) is a widely used system that categorizes skin into four primary types—neutral, dry, oily, and combination—on the basis of hydration levels and oil balance. Advancements in AI have introduced more precise and efficient methods for skin type assessment. Chang et al. explored the use of spectral data combined with a CNN model to classify skin types based on the Fitzpatrick scale, showcasing the potential of AI in utilizing spectral measurements, with an accuracy rate of 92.59% [12]. In contrast, Borade et al. focused on image-based analysis, employing a hybrid model of wavelet scattering networks and ResNet, which demonstrated high accuracy in classifying dry and oily skin types, even with limited datasets [13]. Building on this, Kothari et al. applied an enhanced CNN model to high-resolution facial images, extracting features such as texture and oiliness to achieve consistent and reliable skin type classifications [14].

Skin thickness is crucial for both pathological evaluations and aesthetic treatments, particularly for treatments such as rejuvenating lasers and fillers designed to restore a fuller and thicker appearance. AI-based noninvasive estimation methods provide a promising alternative to traditional techniques that typically rely on skin biopsies. A model utilizing machine learning and hyperspectral imaging was proposed for the noninvasive estimation of human skin thickness, demonstrating accuracy comparable to ultrasound, the clinical gold standard [15]. It successfully estimated the thickness of five skin layers and four dermal layers across various anatomical locations, including the dorsal forearm, upper inner arm, calf, and cheek, offering a promising tool for precise and noninvasive skin analysis.

From Aesthetic Concerns to Diseases: AI in Diagnosis

Vitiligo

Vitiligo, although a benign condition, can significantly impact patients’ quality of life and impose considerable psychosocial stress, particularly when lesions are visible on areas such as the face, neck, and hands. While the diagnosis of vitiligo is relatively straightforward, its assessment poses challenges such as variability in lesion number, size, location, shape, and degree of depigmentation, especially when considering post-treatment repigmentation.

Accurate identification and evaluation of lesion size and pigmentation changes are essential for monitoring disease progression and assessing therapeutic outcomes. Several AI models have been applied to the detection and accurate segmentation of vitiligo lesions. In our recent work, we developed and validated a deep learning-based hybrid AI model for the objective morphometric and colorimetric assessment of vitiligo [16]. The process begins with YOLO v3, which rapidly identifies potential lesion areas and generates bounding boxes to narrow the scope for segmentation. Within these regions, UNet++ performs pixel-level segmentation, extracting precise lesion contours to enable subsequent calculations of lesion area and pigmentation analysis. The hybrid AI model demonstrated high sensitivity (92.91%) and a strong segmentation score (JI, 0.79) in detecting and segmenting vitiligo lesions in Chinese patients with Fitzpatrick skin types III or IV.

In another study, Zhang et al. expanded upon traditional datasets that typically include only normal skin and vitiligo images by incorporating multiple differential diagnoses of vitiligo (e.g., nevus depigmentosus, tinea versicolor) [17]. This approach allowed for a more comprehensive evaluation of the DenseNet model, which achieved an F1 score of approximately 0.96 on the test set, demonstrating its robustness and potential for vitiligo differentiation. Furthermore, Zhong et al. focus on optimizing vitiligo diagnosis using deep learning models, with an emphasis on performance improvement and, more importantly, model interpretability [18]. The Swin Transformer large model achieved the best classification performance among the five models tested, while Class Activation Mapping (CAM) was used to enhance interpretability by highlighting lesion areas, emphasizing boundaries and pigmentation differences.

Benign Pigmentary Disorders

Benign pigmented skin lesions (PSLs) result from either excessive melanin production or an abnormal increase in the density of active melanocytes [19]. While groundbreaking advancements have been achieved in AI-assisted diagnosis of malignant pigmented skin diseases, the early and accurate diagnosis of PSLs remains crucial, and specialized AI tools can not only alleviate the workload of clinicians but also integrate into mobile applications, enabling broader accessibility for patients.

Yang et al. utilized DenseNet and ResNet, two CNN models, to train on pre-cropped local images of six common benign hyperpigmentation dermatoses, including acquired nevi of Ota, melasma, café-au-lait spots, freckles, seborrheic keratoses, and nevi of Ota, achieving high classification accuracy comparable to dermatologists [20]. However, as the training images in previous studies were pre-cropped to isolate individual lesions, these models face challenges in adapting to clinical scenarios where multiple pigmented lesions of varying types may need to be simultaneously detected and analyzed. To address this, Ding et al. proposed a YOLO v5-based object detection model that can automatically annotate lesions and detect different PSLs in the same image [21]. The model performed well not only on high-quality digital camera images but also on smartphone-captured photos, achieving an accuracy of 90.5% on the test set, demonstrating its utility for preliminary detection by general users.

Focusing specifically on melasma, Liu et al. utilized 8010 images captured using five different imaging modes from the VISIA system (normal, brown spots, UV spots, porphyrins, and red areas) to perform multichannel image fusion. This approach significantly enhanced model performance, achieving a maximum accuracy of 97.4% and an AUC of 99.28%, contributing to the development of a high-precision diagnostic model based on a large-scale dataset for melisma [22]. Notably, AI has also advanced research into the pathophysiological mechanisms of melasma. Chen et al. utilized a CNN model based on optical coherent tomography to automatically identify and quantify melanin in melasma lesions and surrounding skin [23]. This approach accurately analyzed melanin distribution, area, morphology, and brightness across different skin layers. Building on this, Wang et al. further explored the dynamic changes in melanin under ultraviolet stimulation, uncovering the “fixed” hyperactive state of melanocytes within melasma lesions and the dynamic UV-responsive behavior of surrounding skin [24].

Acne Vulgaris

Acne vulgaris is a common inflammatory condition of the pilosebaceous unit that typically begins during adolescence but may persist into adulthood. It presents with a variety of lesions, including comedones, papules, pustules, nodules, and cysts [25]. The earliest lesion-based grading system was initiated by Witkowski and Simons, which categorized acne severity on the basis of both the number and type of lesions. This method has since been adapted regionally, leading to a proliferation of complex and time-consuming classification standards.

Fortunately, AI has increasingly been applied to automate acne severity grading. Early research by Shen et al. utilized CNNs and a seven-class classifier to identify and categorize acne lesions, including blackheads, whiteheads, papules, pustules, cysts, nodules, and normal skin, showing potential for diagnostic and qualitative rather than quantitative grading applications [26]. Similarly, Yang and colleagues trained Inception-v3 using the Chinese Guidelines for the Management of Acne Vulgaris, achieving an average F1 score of 0.8 and demonstrating high consistency with attending dermatologists. The model was also used to retrospectively evaluate ten acne cases, validating its clinical value in guiding treatment [27]. Subsequent research shifted toward qualitative acne diagnostic models. To address the variation in acne grading standards, Lin et al. proposed the global local fusion network, pretrained on large datasets such as ACNE04 (Hayashi’s criteria) and fine-tuned on smaller ones such as PLSBRACNE01 (Pillsbury’s criteria) using transfer learning to refine feature parameters, exploring its potential for adapting to diverse grading systems [28].

The prevalence of acne and its chronic nature as an inflammatory condition have driven a growing demand for self-assessment tools via smartphone, leading to significant advancements in teledermatology. Seité et al. developed a smartphone-based AI algorithm that utilizes the Global Evaluation Acne (GEA) scale, a tool widely used in Europe, to assess the severity of facial acne [29]. In a similar vein, Huynh et al. introduced AcneDet, a smartphone-based acne severity grading system that utilizes a Faster R-CNN for detecting four types of acne lesions [30]. In addition, Wang et al. developed a smartphone-based system that integrates a comprehensive grading standard, considering lesion type, count, and regional weights for precise evaluation [31]. The system incorporates a lightweight model that enables deployment on low-end smartphones. Moreover, it integrates patient information, including age, gender, and skin type, to provide personalized assessments, distinguishing it from most studies that rely solely on image-based diagnoses.

Despite the development of numerous diagnostic models, real-world validation remains essential for their clinical application. Phan et al. introduced a smart light emitting diode (LED) therapy system combining deep learning and the Internet of Things (IoT) in 2021 [32]. The system detects acne regions and severity, dynamically adjusts LED light power density and wavelength, and provides personalized red and blue light therapy. Furthermore, they advanced this work by developing a low-level laser system using the U2-Net deep learning model [33]. This system expanded the diagnostic scope to include acne, HSV, and hyperpigmentation while leveraging an IoT platform to record patient data, such as skin type and medical history, for personalized laser treatments with optimized wavelengths and dosages.

Aesthetic Issues

Addressing aesthetic concerns such as wrinkles, spots, and pores often depends on specialized equipment, which is widely considered the gold standard but can come with high costs. Advances in AI have simplified measurement methods, with objective skin analysis software now accessible on portable platforms, which has been preliminarily validated in clinical settings. This AI-based app was further compared with the VISIA system, which utilizes multiple lighting techniques to evaluate skin features [34]. Scores for five characteristics—spots, wrinkles, redness, texture, and pores—showed an overall agreement rate of 67.7% between the two devices, with the highest consistency observed for texture (72.0%) and pores (68.2%). While the app’s performance remains limited, particularly due to its camera functionality and the quality and comprehensiveness of image acquisition, these results still highlight its potential for practical applications.

AI’s performance benefits from ongoing algorithmic advancements, refining detection standards, and improving noise reduction to ensure adaptability in complex imaging conditions. Further study by Georgievskaya et al. trained their AI-based software on a dataset comprising three million data points, to refine assessment categories into sagging, dark circles, eye bags, wrinkles, pores, uniformness, acne, pigmentation, redness, and translucency [35]. This tool demonstrated cost-effectiveness in skin assessment, though the detailed evaluation criteria for these categories were not explicitly disclosed. In contrast, Park et al. developed and validated an AI-based diagnostic system using the U-Net architecture, which provided more clearly defined ranking metrics for each category [36]. Analyzing facial images of Korean women, the system evaluated wrinkles, pigmentation, skin pores, and red spots through detailed metrics such as total count, average/total area, and density for each feature. Furthermore, Liao et al. introduced a method combining Mask R-CNN and YOLO v4 to detect facial symptoms [37]. Mask R-CNN was first applied for facial segmentation, eliminating nonfacial features and background noise, after which YOLO v4 extracted detailed features. This approach achieved high accuracy in detecting acne, freckles, and wrinkles, even with limited training data.

AI’s adaptability is further evident in its ability to address variations in skin characteristics across diverse populations and regions. Flament et al. validated an automated algorithmic system for analyzing facial signs using selfie images from women in the USA across diverse ages, ancestries, and phototypes [38]. To ensure inclusivity, the study further extended validation to women from East Asia [39] and South Africa [40], demonstrating its ability to analyze diverse populations. AI applications also address gender diversity, where the research by Jiang et al. developed a CNN-based software trained on Skin Aging Atlases to quantify nine facial signs by ethnicity, validated with selfies from women of African-American, Asian, and Caucasian backgrounds [41]. Zhang et al. expanded this approach by designing a similar system for male facial signs, tested on 330 men from the same ethnic groups, demonstrating AI’s ability to address gender-specific features in aesthetic dermatology [42].

AI-Assisted Treatment Innovations

Phototherapy

Light-based therapies, encompassing treatments such as lasers and intense pulsed light (IPL), serve as a cornerstone in aesthetic dermatology, addressing pigmentary disorders, vascular lesions, and skin rejuvenation [43]. The application of AI in phototherapy spans three key areas: first, intelligent skin diagnosis and classification (as discussed above), coupled with the generation of corresponding treatment recommendations, which still require further clinical validation. Second, the optimization of treatment trajectories and dynamic adjustment of therapeutic parameters to enhance precision and efficacy. Lastly, safety monitoring through innovations, such as real-time temperature control, enhances patient protection by preventing overheating.

Many laser-based aesthetic procedures rely on quasiperiodic and rhythmic flashes to induce photothermal effects in target areas. However, as laser pulses last only a few milliseconds and leave subtle traces that are difficult to observe, achieving uniform laser distribution remains a challenge. This limitation has prompted the development of automation and intelligence in laser applications. A novel robotic system for skin photo-rejuvenation calculates and segments the facial surface into seven regions, and ensures these areas are evenly filled with laser irradiation by intelligent path planning [44]. Compared with human practitioners, the system covers 20% more area and achieves a more uniform, nonoverlapping laser distribution. This method, however, relies on hard-coded robotic movements, restricting their adaptability to new cases and subjects. To overcome this, Duan et al. proposed a neural network-based framework, known as robot imitation learning, which extracts and transfers laser motion patterns from human experts to robots, enabling seamless adaptation for new treatments [5].

Personalizing laser treatment parameters with AI remains challenging due to factors such as target size, location, patient skin type, and treatment response, along with difficulties in building large and diverse training databases. Currently, AI systems for disease-specific parameter recommendations are still lacking. An early study used neural networks to predict clinical responses to excimer laser treatment in vitiligo, focusing on identifying responders and estimating treatment duration [45]. The model incorporated variables such as skin phenotype and lesion characteristics (e.g., location, duration, and size) to predict repigmentation time. However, validation revealed an inherent alignment in repigmentation timing across lesions within the same patient, partially independent of the included variables or influenced by factors not captured easily in the model. This interdependence complicates predictions and underscores the challenges in tailoring laser treatment parameters to individual patients and lesions.

Laser treatment safety remains a critical concern, particularly in ensuring that energy-based devices maintain optimal temperatures to prevent burns or skin irritation. Dynamic temperature control technologies play a pivotal role in addressing this issue. Gao et al. introduced a deep operator network capable of predicting subcutaneous temperatures, enabling real-time monitoring [46]. Similarly, Heikal et al. highlighted the integration of infrared thermal imaging with CNNs to classify skin thermal responses to laser irradiation, offering valuable insights for optimizing clinical applications and enhancing patient safety in laser-based therapies [47]. Furthermore, endpoint reactions, which refer to observable changes in the skin, such as redness or pigmentation shifts during or immediately after laser treatment, require real-time detection and analysis. This represents a promising area for further research and development.

Filler Injection

Filler injections, typically using hyaluronic acid or other substances, are used to improve wrinkles, facial hollows, and contours in areas such as the forehead, eyes, lips, and cheeks. However, traditional strategies for determining injection sites and volumes rely primarily on the practitioner’s subjective experience. This lack of objective evaluation can lead to over- or undertreatment, resulting in outcomes that may not align with patient expectations. AI now allows for precise prediction of both injection sites and volumes, providing support for personalized and standardized treatment approaches [48].

Alrabiah et al. proposed a deep learning framework that uses CNN to analyze facial images and crop target areas, such as the forehead. The cropped images are then classified with Inception-V3 to detect visible wrinkles, outputting “wrinkles%” and “no wrinkles%” to assess the need for further filler injection [49]. The model achieved 85.3% accuracy in wrinkle classification. Building on this, Ali Shah et al. further focused on predicting injection dosage. Using 3D facial modeling, they selected target areas on the basis of anatomical features (such as the cheekbones and temples) and extracted local 3D feature data [50]. They developed the Rejuv3DNet and Kernel Regression models to estimate dosage by learning the underlying structural mapping from pre- and post-treatment 3D face images. The two models achieved prediction accuracies of 62.5% and 66.7%, respectively, on a real-world dataset. Given the limited scale of the training data, the authors plan to further improve accuracy by constructing a larger and more diverse real-world dataset. In addition, Ali Shah et al. made progress in simulating the effects of filler injections by using mesh morphing techniques to simulate facial changes under different injection volumes [4]. While this approach is currently used to generate synthetic data for dataset augmentation, it has the potential to create intuitive post-treatment visualizations through, for example, mesh deformation techniques and deep AI interaction. In the future, these techniques could offer patients a clear and objective answer to the question, “How will I look afterward?”

Moreover, injection safety is a critical concern, particularly in filler injections. Accidental injection into blood vessels can lead to embolism, skin necrosis, or even blindness. A study by Waked et al. proposed using magnetic resonance angiography (MRA) to map the superficial arteries of the face and convert this data into augmented reality (AR) images to visualize the facial artery locations and minimize the risk of injection errors [51]. It is noteworthy that AI’s involvement in AR can include the following: first, using image processing to convert MRA images into 3D models; second, applying image segmentation to extract key vascular structures from the MRA images; and finally, using facial recognition and positioning to align the facial contour with the artery locations in real-time.

In addition to filler injections, improper administration of botulinum toxin (BoNT-A) can also lead to excessive relaxation of facial muscles, affecting facial expressions and resulting in facial stiffness or loss of expressive ability. The study by Aktar Ugurlu et al. evaluated the changes in facial expressions following BoNT-A injections using deep learning technology, analyzing photos of patients’ neutral, happy, surprised, and angry expressions taken before and 14 d after the procedure [52]. However, since the use of CNNs for facial expression recognition is already a well-established AI application, it is also important to focus on developing objective post-injection expression assessment standards when using such tools.

Hair

Hair loss significantly impacts an individuals confidence, making hair transplantation an increasingly popular aesthetic procedure. AI-based robotic hair transplantation, especially through follicular unit extraction, optimizes follicular unit harvesting using advanced AI algorithms [53]. Robotic systems such as ARTAS integrate visual recognition and decision-making algorithms to automatically select the most suitable follicular units, determining the optimal angle and location for extraction with precision. These systems prioritize units with higher hair density, maximizing efficiency and minimizing follicular damage.

Recent advances in deep learning have automated hair density measurements, which are essential for assessing the severity of hair loss and planning of transplantation procedures. A study utilizing YOLO v4 on hair scalp images demonstrated strong performance in hair density measurement, achieving an average precision of 58.67% [54]. Similarly, a mask region-based CNN can assess hair loss severity by calculating follicle count, hair thickness, and the number of hairs per follicle, visualizing the results in heatmaps to illustrate the overall type and condition of hair loss [55]. Furthermore, Urban et al. combined deep learning with optical coherence tomography to accurately count scalp hair and follicles, improving the precision of hair loss evaluations [56].

Plastic Surgery

Although it extends beyond the scope of aesthetic dermatology, many principles of AI in plastic surgery are applicable. Current AI research focuses on surgeries such as rhinoplasty, using simulation models to predict the most aesthetically pleasing nasal shape, determine appropriate implant sizes, and analyze factors contributing to enhanced facial attractiveness postsurgery [57]. In addition, a craniofacial plastic surgical robot has been developed for genioplasty, combining AI for image-guided 3D reconstruction, real-time feedback, and surgical path planning, enabling precise operations [58].

One study that exemplifies these principles is the development of a fully automated 3D facial aesthetic surgery simulation [59]. The process begins with 3D facial registration and alignment before and after surgery, creating a delta image that captures the structural differences. This delta image is then used to train a convolutional autoencoder model, which selects the most similar face from a database. Finally, on the basis of this matched face, the model simulates the postsurgery results by adjusting the surgical procedures accordingly.

Patient Education

Transformer models based on deep learning, notably GPT-based architectures, can serve as interactive educational tools for patients, providing information about the procedures, risks, benefits, and potential outcomes of aesthetic treatments, thus enhancing patient consultations. Researchers have conducted preliminary investigations into the use of ChatGPT in relation to neurotoxin and filler procedures, as well as rhinoplasty [60].

In one study, 14 common surgical-related questions (7 for botulinum toxin and 7 for fillers), such as “What is Botox/filler?” and “What can I expect during a Botox/filler procedure?” were input into ChatGPT and its output was further evaluated [8]. While it was encouraging that the majority of physician reviewers either agreed or strongly agreed that the information was medically accurate, absolute agreement on medical accuracy, on a per-question basis, was poor. This further emphasizes the need for expert review of clinical accuracy and relevance before further use of GPT-like models as educational tools.

Challenges and Future Directions

AI applications in aesthetic dermatology are advancing rapidly but also face challenges that need to be overcome, both technically and in terms of application. On the technical side, concerns about the generalizability of closed-loop systems for neural network training and testing, as well as the lack of explainability resulting from the depth of training, must be addressed [61].

On the application side, establishing a unified and comprehensive AI evaluation standard to ensure objective assessments across various treatments is an important challenge that needs resolution. As previously mentioned, the use of CNN for facial expression evaluation after BoNT-A injections, while innovative in terms of evaluation tools, requires more qualitative evaluation standards. Furthermore, in order to promote the global application of AI models, it is essential to create a diverse dataset covering different races, genders, ages, and cultural backgrounds. Encouragingly, such work is already underway.

In addition, AI models that integrate patient history and clinical images, along with multimodal data, should be developed to accurately predict and address diverse aesthetic needs [62]. Finally, the requirement for large databases for AI training, especially in facial information in aesthetic dermatology, makes data security crucial. Ensuring privacy while using AI, and clearly defining legal responsibility when issues arise following AI-driven decisions, will be a major challenge for future applications.

The conceptual framework of AI applications in aesthetic dermatology, along with the challenges and future directions discussed above, is concisely illustrated in Fig. 2.

Fig. 2.

Applications of AI in skin health. Artificial intelligence (AI) revolutionizes aesthetic dermatology through comprehensive skin analytics, diagnostic augmentation, and personalized therapeutic optimization. Convolutional neural network (CNN)-based architectures enhance diagnostic precision for both cosmetic concerns (wrinkles, hyperpigmentation, pore visibility) and pathological conditions (vitiligo, benign pigmentary disorders, acne vulgaris). Clinical innovations include optimized phototherapy regimens, personalized injection protocols, and automated hair density quantification

Conclusions

In this review, we have endeavored to comprehensively present the applications of AI in aesthetic dermatology, covering areas such as detection, diagnosis, treatment assistance, and patient education. Overall, the healthcare space has seen the most focus in two main scenarios: AI as a decision-support tool for clinicians and its role in directly enhancing patient-facing medical services.

For clinicians, AI continues to expand its capabilities across the breadth of disease spectra, the reliability of decision-making, and the interpretability of tools. On the patient side, a direct-to-consumer model is already in existence to some extent; smartphone apps that enable individuals to assess and track their skin lesions are increasingly covering a wider range of aesthetic dermatology conditions. In addition, home treatment devices based on such apps are gradually being developed.

Finally, AI is unlikely to supplant dermatologists, but rather supplement their work. The adoption of technology could improve clinical pathways, making aesthetic dermatology services more accessible and efficient, and potentially transforming the future of our healthcare system.

Author Contributions

All authors contributed to the study conception and design. Xiang Chen drafted the manuscript, with Zhaokang Zhou, Hui Ding, Huiying Zheng, and Yiping Ge revising prior versions. Yiping Ge finalized and approved the manuscript.

Funding

This study was supported by CAMS Innovation Fund for Medical Sciences (CIFMS) (2023-I2M-C&T-B-113). The journal’s rapid service fee was funded by the authors.

Data Availability

Data sharing is not applicable to this article as no datasets were generated or analyzed during the current study.

Declarations

Conflicts of Interest

Xiang Chen, Zhaokang Zhou, Hui Ding, Huiying Zheng, and Yiping Ge declare that they have no competing interests.

Ethical Approval

This article is based on previously conducted studies and does not contain any new studies with human participants or animals performed by any of the authors.