1 INTRODUCTION

Since its beginnings, the use of artificial intelligence (AI) has brought advancements of high importance, which have enhanced our daily life and everyday activities in many ways (facial recognition, self-driving cars, and image classification, among others). A growing number of fields can benefit from AI support, including the surgical field (e.g., intelligent systems for assisted surgery, and video-surgery)1, 2 automatic disease diagnosis (e.g., decision support diagnosis systems from images)3-5 and the recently developed personalized medicine, which provided the predisposition to diseases, diagnosing, and selection of the best treatment for a given individual6-11 Although, dentistry might not seem to be greatly impacted by the advances in AI, certain areas such as image-based automatic detection of diseases and other diagnosis-support systems,12, 13 image segmentation for automatic detection of oral traits,14, 15 and resolution enhancement of dentistry related images,16 are undergoing significant improvements thanks to the use of AI.17 On the robotics side, several advances are similarly enabling the utilization of robotic support in dentistry.18 Either way, the door is still wide-open to AI techniques in many areas of dentistry, all under the emerging digital dentistry paradigm.

Several factors have contributed to this recent wave of AI revolution in biomedicine. First, data collection has increased exponentially over the last decades. Yet, data by itself is not enough. Thanks to the developments in high-performance computing (HPC), new powerful AI techniques have enabled a thorough and insightful extraction of information from collected data. This information extraction process is normally referred to as machine learning (ML), that is, the data-driven part of AI, whose objective is to allow the machines (algorithms executed in computer systems) to learn about a specific topic from a certain available dataset. This type of information extraction is usually performed by using supervised learning techniques, which have solved many problems with great success.19 Supervised learning is the task of learning a function that maps an input sample to a desired output, all based on a database of examples of input–output pairs. Once this function has been learned by using training data, new predictions over new incoming samples can be performed.19

Within this dramatic expansion across all the biomedical sciences, ML has reached a number of important milestones. These include the resurgence in recent years of neural networks under the new paradigm of deep learning (DL), the incursion of fuzzy logic (FL) for the treatment of uncertainty and the labeling of numerical data through “linguistic” terms, and the boom in kernel methods (KMs) and other specific ML techniques such as XGBoost.

KMs revolutionized the ML field in the late 1990s and the beginning of the 2000s,20 being a highly considered technique for pattern recognition tasks for middle sized datasets.21 Other pattern recognition techniques that strongly impacted the ML field include random forest (RF) and XGBoost,22, 23 decision-tree based methods frequently applied to a wide range of biomedical problems, including dentistry.24

FL is a well-known paradigm, which has been widely used in the design of decision-making support systems and other applications.25, 26 Its main advantages are related to their capacity to deal with uncertainty in data and with its ability to provide interpretable solutions to the experts in the form of rule bases. Specifically, an area with important applicability to esthetic dentistry is the so-called color naming (i.e., color designation) technique. Early results indicate that bridging the gap between the computational representation of colors in digital devices and subjective human perception of color may be possible.27 Fuzzy colors, defined as fuzzy sets, allow semantics to be introduced in the automatic operation and description of color.28

Finally, DL29 has matched and improved human performance for very complex tasks in areas such as image processing (e.g., object detection, and facial identification) and sound processing (e.g., speech synthesis and processing). In dentistry, the first models based on convolutional neural networks (CNNs) and 2D and 3D photography are emerging for the 3D design of dental prostheses, with very encouraging results.14, 15 Industrial initiatives to store information virtually from a large number of cases for subsequent processing enable a knowledge base to be built to help with the design of optimized treatments based on big data and AI techniques.30 Recently, several reviews have been published relating AI and ML with dentistry. They have approached the topic from a clinical point of view,31-34 either focusing on specific dentistry research areas (AI for dental and maxillofacial radiology,35 forensic odontology,36 orthodontics,37, 38 dental caries39) or on specific AI working areas of dentistry (DL in dentistry,40 AI for dental imaging).39, 41-43 Other review works have approached future trends and challenges.44-48

This comprehensive narrative review provides an insight into the different applications of ML in dentistry from a ML-focused approach to the problem. Special attention was paid to the area of esthetic dentistry and color research, and the great benefits that techniques such as DL and FL bring to this area.

The manuscript is organized as follows: Section 2 presents the methodology carried out to perform the present review. Then, the articleis organized based on the classification of ML techniques, on the three aforementioned main milestones and their application in dentistry. Section 3 presents the well-known and break-through paradigm of DL, which has dramatically changed the way computation and science is done, and summarizes the up-to-date dental applications that make use of deep neural networks (DNNs) to solve a variety of problems. An introduction to other ML techniques in provided in Section 4, such as KMs and gradient-boosting decision tree techniques, and their applications. Finally, Section 5 introduces the FL paradigm, including use of fuzzy systems for color naming in dentistry and fuzzy systems for diagnosing dental diseases. Section 5 is dedicated to some recent clinical assistance software initiatives, which have gained attention in the last few years, and that claim to apply AI techniques in their operation. Finally, Section 6 is dedicated to the future scope of data-driven AI techniques in dentistry, both from the computational and clinical points of view.

2 MATERIALS AND METHODS The review included studies that reported on AI and ML methodologies and applications in dentistry. In addition, the datasets and the comparison (expert opinion or reference standards) used for the model had to be indicated, and studies outcomes had to be quantified (predictive or measurable outcomes). In contrast, the exclusion criteria were as follows: Type of study: animal studies, forensic studies, literature reviews of AI applications for dentistry, letter to editors, comments, questionnaire-based studies, and conferences abstracts. Methodology: AI studies not applied to dentistry, robotics, AI model not described. Outcome: studies that did not report numerical or measurable outcomes. Studies using supervised learning that did not provide information on the data sets used for either training-test or cross validation for the assessment of the methodology

A systematic search was conducted in three different databases (MEDLINE/PubMed, Web of Science and Scopus). All studies have been published in the English language within the last 20 years, and the last search was performed on January 1, 2021. Table 1 shows the search strategy and the terms used for PubMed. The search strategy performed on Web of Science, and Scopus were adapted for each database.

TABLE 1. Structured search strategy carried out in MEDLINE/PubMed database. Searches on Scopus, and Web of Science were adapted according to the respective database Search Topic and terms #1 Artificial Intelligence: “artificial intelligence” OR “machine learning” OR “neural networks” OR “deep learning” OR “Fuzzy logic” OR “computational intelligence” OR “machine intelligence” OR “computer reasoning” OR “Support Vector Machines” OR “generative adversarial networks” OR “color naming” OR “TSK fuzzy system” OR “Computer Vision Systems” OR “Supervised Machine Learning” OR “Fuzzy C-means” OR “Unsupervised Machine Learning” OR “Clustering” OR “Natural Language Processing” OR “TSK fuzzy system” OR “Computer Vision Systems” OR “Supervised Machine Learning” OR “Fuzzy C-means” OR “Unsupervised Machine Learning” OR “Clustering” OR “Natural Language Processing” #2 Dentistry: “dentistry” (Mesh) OR “dentistry” OR “operative dentistry” OR “esthetic dentistry” OR “orthodontics” OR “pediatric dentistry” OR “oral pathology” OR “periodontics” OR “preventive dentistry” OR “prosthodontics” OR “oral surgery” OR “oral medicine” OR “endodontics” OR “oral cancer” OR “tooth segmentation” OR “prosthodontics” OR “dental materials” OR “tooth color” OR “orthodontics” OR “pediatric dentistry” OR “oral pathology” OR “periodontics” OR “preventive dentistry” OR “oral surgery” OR “oral medicine” OR “endodontics” OR “oral cancer” #3 Search #1 AND #2

After searching each database, Mendeley software was used to eliminate duplicates. Two reviewers (FCP and OEP) independently selected the studies analyzing the title and abstract, according to criteria previously described. In case of disagreement, this was resolved by the consensus of a third reviewer (Luis Javier Herrera). During full text reading, the reasons for excluding any paper was recorded (Figure 1).

Flow diagram of the electronic search

A descriptive analysis of the findings was used to evaluate the data. As the selected studies had a large diversity of objectives and the objective of the present study is to analyze the different methodologies of ML applied in dentistry, a quantitative analysis was considered impractical. Therefore, a qualitative data synthesis was performed for this comprehensive narrative review based on a systematic search.

Table 2 shows the AI methodologies and applications in dentistry reported in the included studies (120), which are also organized based on their data type. Table 3 shows a glossary with AI and ML terms used along the manuscript.

TABLE 2. Included studies organized by the AI methods and techniques, target problems, and data type used Technique Application Target problem and studies number Data type used Deep learning Disease identification Dental caries,12, 49-53 oral cancer,54-62 gingivitis,63, 64 other diseases65-80 Radiography,12, 49, 50, 57, 58, 65-73, 77 CT images,59, 74, 76, 80 Other image formats,51, 53-56, 61, 63, 64, 75, 78, 79 clinical data52, 60, 62 Image segmentation 3D tooth segmentation,14, 15, 81-84 2D tooth segmentation,85-89 teeth classification and numbering,90-95 segmentation for disease diagnosis,53, 61, 72, 73, 79, 80 segmentation of other oral surfaces,96-102 metal artifacts,103-105 root morphology,106 teeth alignment107 Radiography,72, 73, 85-88, 91-93, 95, 99-102, 106 CBCT images,14, 80, 82-84, 89, 90, 94, 98, 103, 104 other image formats15, 53, 61, 79, 81, 96, 97 Image correction Image enhancement16, 108, 109 CBCT images16, 108, 109 Other applications Dental implants classification,110-112 landmark detection,113-116 forecast cutting forces,117 need of orthodontic treatment,118, 119 dental artifact status prediction.105 Color matching120 Radiography,110-112, 115 CBCT Images,105, 116 clinical/other types of data,117-120 other image formats113, 114 ML techniques Disease identification Dental caries,13, 121, 122 periodontal disease,123-127 oral cancer,128-135 dental pain,136 oral malodour,137 oral clefts detection.138 Oral disease prevention139 Radiography,122 other image formats,121, 128, 129 clinical/biological data13, 123-127, 130-139 Other applications Dental restoration detection,140 dental deformities,141 failure of dental implants,142 tooth segmentation and numbering,143, 144 predict implant bonelevels,145 shade matching,146 dental care and tooth extraction needs24, 147-150 Radiography,140, 143 CBCT images,144 clinical/biological data,24, 142, 147-149 other image formats141, 146, 150 Fuzzy logic Disease identification Periodontal disease,151 Candidiasis risk factor,152 other diseases and applications153-158 Clinical/biological data,151-154 radiography155-158 Biomimetic color analysis and modeling Color naming,28 color threshold calculation,159, 160 shade guide optimization161 Tooth color measurements28, 159-161 TABLE 3. Glossary Acronym Term Definition AI Artificial intelligence The theory and development of computer systems able to perform tasks normally requiring human intelligence, such as visual perception, speech recognition, decision-making, and translation between languages Supervised learning ML approach where labeled data is used for predicting labels or outcomes NN Neural network ML technique inspired in biological neurons where the input is fed to one or multiple layers to produce an output FFNN Feed forward neural network ML technique inspired in biological neurons where the input is fed to one or multiple layers to produce an output DNN Deep neural network NN with multiple hidden layers, allowing more complex feature construction TL Transfer learning Technique for DNN that used the previously learnt weights from a bigger dataset to learn in an smaller one CNN Convolutional neural network Special type of NN. It can extract spatial information by means of filters, which use the convolution operator GANs Generative adversarial networks Methodology that is used to generate data similar to the input data. Make use of two different models that compete against each other SVMs Support vector machines ML technique where for classification a maximum margin separating hyperplane is built so that the samples of different categories are divided by a clear gap that is as wide as possible DT Decision tree Flowchart-like structure in which each internal node represents a “test” on an attribute, each branch represents the outcome of the test, and each leaf node represents a class label Bagging ML methodology that combines the prediction of multiple weak classifiers in order to improve classification performance Boosting ML methodology that builds classifiers sequentially based on the error of the previous classifier in order to improve classification performance RFs Random forest Methodology that combines the prediction of a high number of weak decision trees, averaging their predictions to perform a final prediction GB Gradient boosting Methodology that builds classifiers sequentially based on the error of the previous ones FL Fuzzy logic A form of many-valued logic in which the truth values of variables may be any real number between 0 and 1 both inclusive. It is used to handle the concept of partial truth, where the truth value may range between completely true and completely false Fuzzification process Converts the given numbered-valued inputs into fuzzy sets according to their membership functions for its later operation using fuzzy logic Knowledge database Provides the definition of the linguistic values of each of the variables considered in a problem, together with the rules making up the rule base of the system Inference engine Operates according to the input values provided and the rule base. It is itself the core of the fuzzy system and resembles the human capability to take decisions Biomimetic Is defined as the examination of nature, its models, systems, processes, and elements to emulate or take inspiration from nature in order to solve human problems 3 DEEP LEARNING APPLICATIONS IN DENTISTRY

Artificial neural networks (ANNs) are learning algorithms based on the functioning of biological neural networks. They can be used for supervised, unsupervised, and reinforcement learning problems and been used to solve many different problems. The basic structure of an ANN is a set of interconnected layers of operating neurons, and the term deep refers to ANN with a large (deep) number of both layers and neurons per layer.

Applications of DL in dentistry are probably the most promising area of research in this field. This type of techniques can contribute to the design of high-performance de