**3. AI for extracting and quantitating the feature information of skin**

Artificial intelligence (AI) came out in 1956. This technology is widely used in various fields, and realizes intelligent diagnosis and treatment in the medical field through the screening, diagnosis, and management of diseases. Machine learning (ML), which is a subset of artificial intelligence, is represented by mathematical algorithms that improve learning through experience. There are two main types of machine learning algorithms (**Figure 9**): (i) unsupervised (ability to spot patterns), (ii) supervised (classification and prediction algorithms based on previous examples) [68–70]. ML has gradually become a common method for solving difficult problems in artificial intelligence because computer algorithms can be automatically improved through previous experience [71]. There are dozens of algorithms in ML, including deep learning, decision trees, clustering, and Bayesian. For example, the use of decision trees to monitor the depth of anesthesia is a type of ML [72]. Artificial neural networks (ANNs) are mathematical models of information processing based on structures similar to the brain's synaptic connections. ANNs have performances such as self-learning, associative storage, and fast finding optimal solutions, which are far superior to ML algorithms, and are especially suitable for dealing with cluttered and unstructured data (such as images, audio, and text) [73]. As researchers delved into the structure of ANNs, Deep neural networks (DNN) with more and more complex network hierarchies were produced [74]. It also means that DNNs are more capable of modeling or abstract representations of things, as well as simulating more complex models.

#### **3.1 Supervised learning for dermatology**

Skin cancer is a common cancer type [75]. Melanoma and non-melanoma are the two main types of skin cancer, with melanoma being the most dangerous type of skin cancer with a high mortality rate [76]. Traditional methods for early detection of skin cancer include skin self-examination and skin clinical examination [77]. However, skin self-examination is a random method and its accuracy depends on how well people know about skin cancer. In addition, the use of professional medical tools such as dermoscopy and microspectroscopy for clinical examinations is not only expensive but also requires professionals to operate [78]. Therefore, using AI to identify patients and upload shared images for diagnosis has become a more convenient method.

The most commonly used machine learning algorithm in dermatology is supervised learning. It is mainly related to retrieval-based AI, where we need to input already labeled data in advance [79]. The goal is to analyze this training data and produce an inferred feature that can be used to map out new instances. For example, when identifying benign and malignant skin lesions, we need to label the skin lesions images as benign and malignant in advance. In this approach, automatic classification of new and unlabeled images can be achieved once training on these images is complete. Esteva et al. explored the accuracy of this skin cancer classification algorithm by comparing deep-learning diagnoses with labeled results from 21 dermatologists. They used approximately 1.28 million images (1000 object categories) from the 2014 ImageNet Large-Scale Visual Recognition Challenge as pre-training objects to form validation and testing datasets. **Figure 10** shows the working system. The area under the curve (AUC) of the CNN algorithm exceeds 91%, indicating that the sensitivity and specificity in the classification of epidermal and melanocytic lesions is superior to

**Figure 9.**

*Machine learning algorithms (a) supervised learning (b) unsupervised learning. Reprinted with permission from reference [68].*

that of dermatologists. The results of this study show that, well-trained deep learning enables highly accurate diagnostic classification [80]. At present, ML has been gradually applied in combination with optical technology, which is mainly manifested in the use of AI-assisted analysis of OCTA data to achieve advanced diagnosis and correction of dispersion problems in OCT images to improve axial resolutio [81, 82]. These findings will help advance the application of AI in wound healing monitoring.
