Artificial intelligence in orthodontics: Where are we now? A scoping review

Objective: This scoping review aims to determine the applications of Artificial Intelligence (AI) that are extensively employed in the field of Orthodontics, to evaluate its benefits, and to discuss its potential implications in this speciality. Recent decades have witnessed enormous changes in our profession. The arrival of new and more aesthetic options in orthodontic treatment, the transition to a fully digital workflow, the emergence of temporary anchorage devices and new


| INTRODUC TI ON
The last decades have witnessed enormous changes in our profession. The arrival of new and more aesthetic options in orthodontic treatment, the transition to the fully digital workflow, the emergence of temporary anchorage devices and new imaging methods all work to provide both patients and professionals with a new focus in orthodontic care. 1 To make the diagnostic process more accurate and efficient, the use of Artificial Intelligence (AI) in orthodontics has grown significantly in recent years. This knowledge is fundamental for predicting treatment prognosis. However, the addition of this AI-based knowledge does not change the fact that the health professionals, with their own knowledge gained through specialized education and years of experience, are the ones that ultimately have to diagnose and determine the best treatment plan. Nevertheless, AI can be useful when making specific clinical decisions in a limited time. AI applications can guide clinicians to make better decisions and perform better, because the results obtained from AI are highly accurate and therefore, in some cases, can prevent human errors. 2 To appreciate the impact of AI on orthodontics, it is first important to discern some key terms related to AI: • AI's main objective is to offer a machine the ability to have its own intelligence. Put another way, AI aims for a machine to be able to learn through data, to solve problems by itself.
• Machine learning (ML) is the main backbone of AI. It depends on algorithms to predict outcomes based on data sets and draws influence from many research disciplines. Its purpose is to facilitate machines to learn from data so they can resolve issues without human input. The most commonly used techniques of ML include the support vector machine (SVM), logistic regression (LR), naive Bayesian classifier, decision tree (DT), random forest (RF), extreme learning machine (ELM), fuzzy k-nearest neighbour (FKNN) and convolution neural network (CNN). 2,3 • Neural networks are a set of algorithms that calculate signals through artificial neurons that try to imitate the functioning of human neurons.
• Deep learning is an integral part of ML. It uses networks with different computer layers in deep neural networks to analyse input data. Its purpose is to build a neural network that can automatically recognize patterns to improve feature detection. 2,3 • Big data refers to large data sets and/or the combination of all available data points drawn from multiple sources which can be used to recognize patterns that inform a customized experience for different individuals. 1 Orthodontic treatments are usually long procedures with an average treatment duration of nearly 29 months, 4 which is why orthodontists must become more efficient to adapt to the needs of society. The application of ML techniques can help to solve this issue.
Recent technological innovations in orthodontics, including cone beam computed tomography (CBCT) and 3D visualizations, intraoral scanners, facial scanners, instant teeth modelling software capabilities and new appliance developments using robotics and 3D printing, are changing the face of medical care and are quickly becoming integrated into dentistry. 5 These tools enable a better understanding of the patient's anatomy and are able to create dynamic anatomical reconstructions for the specific patient, and therefore accommodate the possibility of 3D treatment planning. Convolutional neural networks (CNNs) are increasingly applied for medical image diagnostics, most frequently for the detection, segmentation or classification of anatomical structures. Deep learning has also recently been used for geometric feature learning and classification. 6 Machinelearning approaches, which are algorithms trained to identify patterns in large data sets, are ideally suited to facilitate data-driven decision-making. 7 This scoping review aims to determine the applications of AI that are extensively employed in the field of orthodontics, to evaluate the benefits of AI and to discuss its potential implications in this speciality.

| Protocol
This review was performed following the Preferred Reporting Items for Systematic reviews and Meta-Analyses extension for Scoping Reviews (PRISMA-ScR) guidelines. 8

Study question
What is the applicability of Artificial Intelligence in the field of Orthodontics?

Intervention
Artificial intelligence-based forms of diagnosis and treatment planning.

Comparison
Reference standards and existing literature.

Outcome
Measurable or predictive outcomes such as accuracy, sensitivity and specificity.

TA B L E 1 Description of PICO elements
PubMed) was conducted to prepare the study protocol. The data extraction forms were constructed after the initial results of the pilot search.
The search was based on the PICO (problem/patient/population, intervention/indicator, comparison and outcome) elements (Table 1).

| Literature search
The electronic literature search was performed through MEDLINE/ PubMed, Scopus, Web of Science, Cochrane and IEEE Xplore databases between November 2020 and March 2021.
A specific combination of words was introduced in order to complete a specific and reproducible search (

| Eligibility criteria
First, search engine results were evaluated for relevance based on their title and abstract. The studies whose titles or abstracts contained different information that was not related to the study question were excluded. An 11-year restriction was determined, from January 2010 to March 2021, to ensure the review was based on the most up-to-date information. Only fully available articles were considered. Articles focused on AI in the field of orthodontics were included. Only those publications that used some predictive measurable outcomes such as accuracy, sensitivity and specificity, and those with adequate documentation of the data sets they employed, were considered. All relevant publications and studies whose abstracts did not provide enough information to justify an exclusion decision were obtained in full text to determine their eligibility. Articles wrote in any language other than English, Spanish, Portuguese, Italian, German or French were excluded, as well as studies related to non-AI areas. Table 3 depicts how we collected select information from the included studies. The type of ML method, the number and type of images used for testing AI software, the accuracy of the technique, and its benefits to the field of orthodontics were extracted from the articles.

| Search and study selection
The flowchart of the articles conforming to the PRISMA-ScR and included in this scoping review study selection is shown in Figure 1.
The electronic literature search initially returned 311 records, which was reduced to 115 after removing duplicate references.
After reviewing the titles and abstracts, all 115 studies were examined in more detail. Two articles were excluded as their full text was not available. Ninety records were excluded because they did not Of the 17 studies included in this scoping review (Table 3) Lastly, one publication quantified the 3D asymmetry of the maxilla in patients with unilateral cleft lip and palate.

| Outcome domains of included studies
Considering the selected articles, a total of 472 lateral cephalometric radiographs were used in two of the studies to analyse the accuracy of using neural network ML to decide whether to use extractions to reduce discrepancy in different orthodontic malocclusions. 7

TA B L E 3 (Continued)
(Continues) 400 lateral cephalometric radiographs. 10 Following the same line of research, 500 radiological images of the head profile were used in two articles to study the viability of automatic detection of anatomical reference points on radiological images using a CNN (U-Net) 11 and Bayesian network. 12 The accuracies reported in these studies were 90.11% and 92%, respectively. 11 with an intra-class correlation coefficient (ICC) greater than 0.90.
Using a CNN, they also determined the existence of significant maxillary hypoplasia on the cleft side of those patients. However, it is important to remember that there is no singularly correct answer for the diagnosis of extractions. 7 Generally, most orthodontists decide whether an extraction is necessary based on their experience and knowledge by analysing data from their patients' clinical evaluation, photographs, dental casts and radiographs. One problem is that this often causes intra-and inter-clinician variability in the treatment planning process. 25 By mimicking the decisionmaking of human experts, an AI expert system could be developed based on various philosophies of diagnosis to assist the decisionmaking process. 7 Nevertheless, the final decision will always belong to the clinicians.

| D ISCUSS I ON
Various studies have been conducted to demonstrate the efficacy of AI applications in identifying cephalometric landmarks.
The diagnostic value of the analysis depends on the accuracy and the reproducibility of landmark identification. In orthodontic practice, lateral cephalometry has been widely used for skeletal classification and treatment planning. The incorporation of a CNN can provide an accurate and robust skeletal diagnostic system. 12  AI has also been used to automatically identify and classify skeletal malocclusions from 3D CBCT craniofacial images. In 2020, Kim et al proposed a method that aimed to assist orthodontists in determining the best treatment path for the patient, be it orthodontic treatment, surgical treatment, or a combination of both. 28 Fast and efficient CBCT image segmentation would allow for large clinical data sets to be analysed effectively. 29 ML can help to determine the cephalometric predictors of the future need for orthognathic surgery, as in patients with repaired unilateral cleft lip and palate (UCLP). 30 Thus, the use of AI definitely reduces doctor assessment workload and improves diagnostic accuracy. 20 The assessment of bone age and skeletal maturity and its comparison to chronological age is an important task for the diagnosis of paediatric endocrinology, orthodontics and orthopaedic disorders. Dental segmentation is one of the key steps in computerassisted orthodontic technology and its accuracy is closely related to treatment outcome. This procedure requires precise positioning and extraction of tooth shapes on the patient's 3D digital dental cast (or intraoral scan). ML using a CNN-based model for tooth segmentation and identification achieved performance improvements when compared with the state-of-the-art general mesh segmentation method for both tooth segmentation and identification tasks. 18 Deep learning systems work in distinct areas of orthodontics.
Orthodontists can use AI systems as an ancillary tool for increasing the accuracy of diagnosis, treatment planning and for predicting treatment outcomes. Automated systems can save a lot of time and increase the efficiency of the clinicians. For example, the use of automated cephalometric points identification or automated teeth segmentation to enable a treatment preview outcome helps reduce orthodontic treatment planning times. 5,10-13 Additionally, with deep learning techniques it is possible to eliminate the subjectivity associated with human decision-making; traditional manual methods are likely to incorporate a relatively higher degree of intra-and interobserver errors due to that subjectivity, which can lead to an increase in the prediction error. 32 Likewise, these systems can be used for secondary opinions, in order to improve the accuracy of diagnosis.
Nevertheless, clinicians should always trust their clinical judgment above all.
AI could become a valuable tool to use in those procedures that require high precision and are more time consuming, such as indirect bonding, precise Bolton Analysis or wire bending, in order to increase the quality of the treatments we offer to our patients.

| Limitations
This review presents two main limitations: First, being a scoping review, the review question has to be more generally defined when compared to a systematic review. Whereas scoping reviews assess where consolidated knowledge ends and additional research is needed, systematic reviews clarify whether existing knowledge is reliable. 33 AI embraces many different fields and applications, and therefore, it adjusts with the aim of a scoping review, which is to provide an overview of the evidence.
Second, the search was limited to the last 11 years, because the authors agreed that it would be more useful to describe only the latest applications of AI in orthodontics, rather than making an historical review and thereby including obsolete technologies.
Despite these limitations, the authors expect this to be a useful overall introduction to understand the recent past of AI and the actual present (as well as the near future) of its applications in orthodontics. There is no doubt that there is still a long road ahead.
Many of the results published in the papers used in this scoping review must be thoroughly and carefully analysed. However, those who are already used to work with intraoral scanners and facialdriven smile designs know exactly how limited and at the same time how useful all these new technologies are. Therefore, all the tools available to the clinicians are of great value, and AI is one of them.
Nevertheless, the authors truly believe that despite all the future advancements in AI, it will never substitute human reasoning; however, it will definitely help.

| CON CLUS IONS
The analysed studies demonstrated that CNNs can be used for the automatic detection of anatomical reference points on radiological images. For growth and development areas, the CVM can be determined using an ANN model and obtain the same results as human observers. AI technology can also help improve the accuracy of diagnoses for orthodontic treatment, therefore, helping orthodontists work more efficiently. However, although the improvement of AI is definitely a great help for orthodontists and other health professionals, the final decisions on health matters will always be the clinicians' responsibility.

ACK N OWLED G EM ENTS
The authors received no financial support for this work.

CO N FLI C T O F I NTE R E S T
The authors declared no conflict of interest.

AUTH O R CO NTR I B UTI O N S
Monill-González A. and Rovira-Calatayud L performed equally with the study selection, data extraction and data presentation. They also completed the initial manuscript draft and data analysis. d'Oliveira NG conceptualized the study and resolved disagreements in study selection. He analysed all data and prepared the presentation of the final manuscript. Ustrell-Torrent JM supervised the research activity and provided the necessary resources to conduct the review. All authors contributed to critical revision of the article. All authors read and approved the final manuscript.

DATA AVA I L A B I L I T Y S TAT E M E N T
The data that support the findings of this study are available from the corresponding author upon reasonable request.