Trends and Application of Artiﬁcial Intelligence Technology in Orthodontic Diagnosis and Treatment Planning—A Review

: Artiﬁcial intelligence (AI) is a new breakthrough in technological advancements based on the concept of simulating human intelligence. These emerging technologies highly inﬂuence the diagnostic process in the ﬁeld of medical sciences, with enhanced accuracy in diagnosis. This review article intends to report on the trends and application of AI models designed for diagnosis and treatment planning in orthodontics. A data search for the original research articles that were published over the last 22 years (from 1 January 2000 until 31 August 2022) was carried out in the most renowned electronic databases, which mainly included PubMed, Google Scholar, Web of Science, Scopus, and Saudi Digital Library. A total of 56 articles that met the eligibility criteria were included. The research trend shows a rapid increase in articles over the last two years. In total: 17 articles have reported on AI models designed for the automated identiﬁcation of cephalometric landmarks; 12 articles on the estimation of bone age and maturity using cervical vertebra and hand-wrist radiographs; two articles on palatal shape analysis; seven articles for determining the need for orthodontic tooth extractions; two articles for automated skeletal classiﬁcation; and 16 articles for the diagnosis and planning of orthognathic surgeries. AI is a signiﬁcant development that has been successfully implemented in a wide range of image-based applications. These applications can facilitate clinicians in diagnosing, treatment planning, and decision-making. AI applications are beneﬁcial as they are reliable, with enhanced speed, and have the potential to automatically complete the task with an efﬁciency equivalent to experienced clinicians. These models can prove as an excellent guide for less experienced orthodontists.


Introduction
Artificial intelligence (AI) is a new breakthrough in technological advancements based on the concept of simulating human intelligence. AI models are designed to assist humans in completing tasks more efficiently and are developed through training and programming, using a large number of datasets, and then executing these models to turn the data into actionable information in performing a specific task [1]. These emerging technologies highly influence the diagnostic process in medical sciences, with enhanced accuracy in diagnosis [2]. AI models have been applied in healthcare in order to assist healthcare professionals and enhance the accuracy in diagnosis, decision making, and predicting prognosis, with the ultimate goal of enhancing patient care and treatment outcomes. AI models have also demonstrated performances that are on par with trained and experienced healthcare professionals [3].
AI technology has been widely applied in dentistry. These models are based on Machine Learning (ML) systems, which are designed using algorithms that can be trained using a large number of data sets and later applied to new sets of data for testing and evaluation. ML based AI models have displayed excellent accuracies in making predictions based on the datasets without human interventions [4][5][6]. AI models based on the newly developed and advanced system known as Deep Learning (DL) are designed to mimic the functioning of the neural system of the human brain; hence they are termed artificial neural networks [7].
The AI system has been gaining high popularity in dentistry, as they have been widely applied for diagnosing and predicting oral diseases such as dental caries, periodontal diseases, and oral cancer [8][9][10][11]. AI models have demonstrated exceptional performances in diagnosis, determining the treatment needs, and predicting the prognosis of diseases [1,5,[8][9][10][11][12][13]. Along with this, AI models have also been designed and applied in the field of forensic odontology, where the models have demonstrated excellent precision in performing tasks [14]. Considering these developments in AI, this review article intends to report on the trends and application of AI models designed for diagnosis and treatment planning in orthodontics.

Search Strategy
Data search for the articles that have reported on the application of AI models in orthodontics was carried out in the most renowned electronic databases. These databases mainly included PubMed, Google Scholar, Web of Science, Scopus, and Saudi Digital Library. This search was mainly carried out for original research articles published over the last 22 years, from 1 January 2000 until 31 August 2022. Several Medical Subheadings (MeSH terms) were used for searching the articles in the electronic databases. MeSH terms included: artificial intelligence, automated diagnosis, computer-assisted diagnosis, digital diagnosis, supervised learning, orthodontics, dentofacial orthopedics, unsupervised learning, diagnosis, prognosis, prediction, deep learning, machine learning, artificial neural networks, and convolutional neural networks. Boolean operators AND/OR were used to generate a combination of these key words in the advanced search. Language filters for the English language were also used. A manual search for articles was also performed simultaneously in the college central library after screening the references of the articles obtained from the electronic search.
At this stage, article selection was based on the title and abstract; 628 articles related to our research topic were extracted. Later, 436 articles were excluded due to duplication. The remaining 192 articles were applied for the eligibility criteria for being included in this review article.

Eligibility Criteria and Study Selection
Original research articles that reported on the application of AI models in orthodontics were included in this systematic review. Articles with no full text, narrative reviews, scoping reviews, letters to editors, opinion letters, case reports, short communications, conference proceedings, and articles other than English, were excluded ( Figure 1).
A total of 56 articles that fulfilled the eligibility criteria were subjected to qualitative synthesis (Table 1).   A total of 56 articles that fulfilled the eligibility criteria were subjected to qualitative synthesis (Table 1). Summary of AI models designed for different diagnostic tasks is presented in Table 2.

Qualitative Synthesis of the Included Studies
The research trend shows a gradual increase in the number of research publications over the last two decades. However, in the last two years, the number of articles reported on the application of AI models in orthodontics has rapidly increased ( Figure 2

Qualitative Synthesis of the Included Studies
The research trend shows a gradual increase in the number of research publications over the last two decades. However, in the last two years, the number of articles reported on the application of AI models in orthodontics has rapidly increased ( Figure 2).

Discussion
Digitalization in medical diagnosis is a new breakthrough in the field of health sciences. In dentistry, digital technology has been widely applied in clinical practice, especially with the use of 3D intra oral scanners designed for scanning dental arches. These technical advancements have simplified the process of impression making and fabrication of prostheses [71,72]. The recent advancements in the field of medical sciences and computer technology have resulted in significant developments in AI based models that have been designed for diagnosing oral diseases, treatment planning, clinical decision making, and predicting treatment outcomes; these models have demonstrated excellent performances [1,[9][10][11].

AI Models Designed for Automatically Identifying Cephalometric Landmarks
In orthodontics, cephalometric analysis is considered one of the most significant tools for evaluating the skeletal profile of the craniofacial region. The manual tracing of X-ray films and plotting the landmarks requires skill and expertise. Advancements in AI technology have resulted in the development of automated models that can predict landmarks without human assistance. Nishimoto S et al. [15] reported on an automated landmark predicting system based on a deep learning neural network, where the model demonstrated a performance equivalent to manual plotting. Park JH et al. [16] reported on the performance of two automatic deep learning models You-Only-Look-Once version 3 (YOLOv3) and Single Shot Multibox Detector (SSD) designed for identifying cephalometric landmarks. The YOLOv3 algorithm outperformed SSD in accuracy and displayed 5% higher accuracy than the benchmark methods reported in the literature. Kunz F et al. [18] reported on an automatic cephalometric identification model using an (AI) algorithm; this model demonstrated quality levels equivalent to experienced professional examiners. Hwang HW et al. [19] reported on the performance of an automated identification system,

Discussion
Digitalization in medical diagnosis is a new breakthrough in the field of health sciences. In dentistry, digital technology has been widely applied in clinical practice, especially with the use of 3D intra oral scanners designed for scanning dental arches. These technical advancements have simplified the process of impression making and fabrication of prostheses [71,72]. The recent advancements in the field of medical sciences and computer technology have resulted in significant developments in AI based models that have been designed for diagnosing oral diseases, treatment planning, clinical decision making, and predicting treatment outcomes; these models have demonstrated excellent performances [1,[9][10][11].

AI Models Designed for Automatically Identifying Cephalometric Landmarks
In orthodontics, cephalometric analysis is considered one of the most significant tools for evaluating the skeletal profile of the craniofacial region. The manual tracing of X-ray films and plotting the landmarks requires skill and expertise. Advancements in AI technology have resulted in the development of automated models that can predict landmarks without human assistance. Nishimoto S et al. [15] reported on an automated landmark predicting system based on a deep learning neural network, where the model demonstrated a performance equivalent to manual plotting. Park JH et al. [16] reported on the performance of two automatic deep learning models You-Only-Look-Once version 3 (YOLOv3) and Single Shot Multibox Detector (SSD) designed for identifying cephalometric landmarks. The YOLOv3 algorithm outperformed SSD in accuracy and displayed 5% higher accuracy than the benchmark methods reported in the literature. Kunz F et al. [18] reported on an automatic cephalometric identification model using an (AI) algorithm; this model demonstrated quality levels equivalent to experienced professional examiners. Hwang HW et al. [19] reported on the performance of an automated identification system, You-Only-Look-Once version 3 (YOLOv3), for the automatic identification of cephalometric landmarks. This model displaced excellent accuracies, which were similar to experienced human examiners. Zeng M et al. [20] reported on an AI based model designed for automatically predicting cephalometric landmarks; this model demonstrated competitive performance compared with other models. Bulatova G et al. [22] reported on an AI based model for automated cephalometric landmark identification; this model demonstrated a performance similar to a calibrated senior orthodontist. Hwang HW et al. [23] compared the performance of a deep learning based AI model with previously reported AI models in the literature designed to identify cephalometric landmarks automatically. This model demonstrated superior results in comparison with the previously reported AI models. Kim J et al. [24] reported that a CNN model designed to identify cephalometric landmarks displayed excellent accuracies. Kim YH et al. [25] reported on a deep learning based fully automatic AI model for identifying cephalometric landmarks; this model demonstrated better performance than two calibrated and experienced examiners. Kim MJ et al. [26] reported on a CNN model designed for the automated identification of cephalometric landmarks using conebeam computed tomography (CBCT) scans; this model displayed better consistency in identifying the landmarks in comparison with the experienced human examiners. Another study conducted by Kim MJ et al. [27] reported on an automatic cephalometric landmark identification system using CBCT images; this model displayed a performance equivalent to the experienced examiners. Yao J et al. [28] reported on a CNN based AI based model for automatically identifying the cephalometric landmarks; this model demonstrated a higher accuracy. Le VNT et al. [29] reported on a human-AI collaboration for identifying cephalometric landmarks. The collaborative system was effective in identifying landmarks. Gil SM et al. [30] reported on a convolution neural network (CNN) based AI model for identifying cephalometric landmarks; this model demonstrated excellent performance, similar to a human examiner's performance.

AI Models Designed for Bone Age and Maturity Estimation
The timing of orthodontic treatment is crucial in achieving the desired clinical outcomes. In treatment planning, quantifying the skeletal growth, mainly the mandible, impacts the diagnosis, treatment planning, and treatment outcomes [73]. Therefore, if orthodontic treatment is initiated during the optimum development phase, it will produce more favorable results. Otherwise, a much longer treatment time and surgical intervention may be needed to correct the deformities of the jaw [74,75]. The standard method for estimating bone maturity uses a hand-wrist radiograph and a lateral cephalogram to estimate cervical vertebral maturity (CVM). However, studies have reported that the reproducibility of the CVM varies among examiners [76,77]. In recent developments, AI has been widely used to estimate bone maturation using hand-wrist radiographs or CVM. Kok H et al. [32] reported on the efficiency of seven AI classifiers designed to determine growth; based on cervical vertebrae stages. These models demonstrated acceptable performance in determining the stages. Another study conducted by Kok H et al. [33] also reported the application of an artificial neural network (ANN) model for determining growth based on cervical vertebrae. This model demonstrated satisfactory results in determining the growth-development periods. Kok H et al. [34] also reported on the success rates of two AI models, artificial neural network models (NNMs) and naïve Bayes models (NBMs), designed for determining the growth and development based on the cervical vertebrae. NNMs displayed a higher success rate than the NBMs. Amasya H et al. [35] reported on the performance of five AI based Machine Learning (ML) models designed for CVM analysis. These AI models displayed more accuracy in comparison with the human observer. Amasya H et al. [36] also reported on the validation of the ANN model designed for CVM analysis; this model displayed better performance in comparison with four trained human observers. Seo H et al. [37] reported on the performance of six CNN based deep learning models designed for CVM analysis on cephalometric radiographs. These six models demonstrated more than 90% accuracy in performing the task. Zhou J et al. [38] reported on the performance of an AI model designed for automatically determining the CVM status using cephalometric radiographs. This model demonstrated good agreement with the human examiners. Kim DW et al. [39] reported on the AI model designed for predicting the hand-wrist skeletal maturation stages. The model demonstrated excellent accuracy in predicting the skeletal maturation stages. Kim EG et al. [40] reported on an AI based CNN model designed for estimating the CVM using lateral cephalograms. The model displayed 93% accuracy in performing the task. Mohammad-Rahimi H et al. [41] reported on an AI model for determining CVM and growth spurts using lateral cephalograms. The model demonstrated reasonable accuracy in determining the CVM stage and displayed high reliability in estimating the pubertal stage. Li H et al. [42] reported on AI based CNN model for CVM classification. This model demonstrated good accuracy in classifying the CVM. Atici SF et al. [43] reported on an AI based deep learning model designed for estimating the CVM stages; this model displayed higher accuracy in determining the CVM stages.

AI Models Designed for Palatal Shape Analysis
The palate has a complex structure, and its shape varies among individuals. The palate's shape is related to the facial pattern and a wide range of factors such as breathing pattern and occlusion [78][79][80]. The shape of the palate is of great interest to orthodontists as it is a potential area for evaluation and assessing the outcome of orthodontic procedures like maxillary expansion [81,82].
Palatal measurements usually include palatal surface area, volume, and depth [78]. These measurements often require greater experience and are often subjected to observer errors. The recent technological advancements have resulted in the development of AI based models designed for palatal shape analysis. Croquet B et al. [44] reported on a deep learning model designed for analyzing the palatal shape. This automatic model demonstrated excellent repeatability with promising accuracy. Nauwelaers N et al. [45] reported on the application of an AI based deep learning model for palatal shape analysis, and this model achieved results similar to conventional approaches.

AI Models Designed for Determining the Need for Extractions
Determining the need for tooth extraction and deciding which teeth need to be extracted is a critical decision in orthodontic treatment planning as it is irreversible [83]. The orthodontists' decision regarding extraction is based on their training, clinical experience, and treatment philosophies [84]. AI technology has been applied to designing models which can be used as an axillary tool for deciding on the need for orthodontic extractions. Xie X et al. [46] reported on the AI based ANN model for determining the need for extraction; this model was very efficient and displayed an accuracy of 80% in determining the need for extraction. Jung SK et al. [47] reported on the performance of an AI model designed to diagnose the need for orthodontic extraction. The model was very efficient in diagnosing extraction and non-extraction cases. Li P et al. [48] reported on the multilayer perceptron ANN model for predicting extraction and non-extraction cases. The model demonstrated an excellent accuracy of 94% for predicting the extraction and non-extraction cases. Choi HI et al. [49] reported on an AI based model designed to determine the need for extraction. This model displayed a success rate of 91% for determining extraction decisions. Suhail Y et al. [50] reported on a machine learning model to diagnose the need for extraction. The model demonstrated a performance that was in agreement with the trained examiners. Etemad L et al. [51] reported on a machine learning model for predicting the need for extraction and non-extraction. The model displayed acceptable results; however, the authors suggested a need for improvisation in the algorithms to improve generalizability.

AI Models Designed for Planning Orthognathic Surgeries
Individuals presenting dentofacial deformities, either due to congenital or acquired conditions, may require orthognathic surgeries in order to reposition the jaws into a functional relationship. In recent years, several articles have reported on the application of automated computerized methods designed for analyzing dentofacial deformities and elaborating treatment plans [85]. Patcas R et al. [57] reported on an AI model designed for predicting facial appearance post orthognathic surgery; this model displayed an acceptable performance in predicting facial attractiveness and appearance. Knoops PGM et al. [58] reported on a machine learning model designed for automated diagnosis and treatment planning. This model demonstrated an excellent sensitivity of 95.5% and specificity of 95.2% in diagnosing the patients. Stehrer R et al. [59] reported on an AI model for predicting perioperative blood loss following orthognathic surgery. The model demonstrated an excellent result and efficiently predicted the perioperative blood loss prior to surgery. Jeong SH et al. [60] reported on an AI model designed for judging the soft tissue profiles requiring orthognathic surgery. The model displayed an accuracy of 89.3% in judging the soft tissue profiles requiring surgery. Lee K-S et al. [61] reported on an AI based deep CNN model for differential diagnosis of orthognathic surgery. The model was successful and can be applied for differential diagnosis of orthognathic surgery. Tanikawa C et al. [62] reported on an AI model designed for predicting facial morphology post orthognathic surgery. The model demonstrated an excellent success rate in predicting facial morphology and can be applied for clinical purposes. Xiao D et al. [63] reported on an AI model designed for virtually simulating the surgical plan. The model showed higher accuracy in generating the shape models. Shin W et al. [68] reported on an AI model that automatically predicts the need for orthognathic surgery. The model was efficient with relative accuracy in predicting the need for surgery. Kim YH et al. [69] reported on an AI based deep learning model designed to diagnose orthognathic surgery. The model demonstrated excellent performance in predicting the diagnosis of orthognathic surgery.
A few of the limitations of this paper might be with the search strategy. Even though we have performed a comprehensive search for articles, some might have been missed. In general, these AI models' limitations are mainly due to the limited amount of data sets that have been applied for training these models, and validating and testing. Another limitation of the data sets is the standardization since the data sets applied for assessing the performance of these AI models are obtained from one diagnostic center in most cases. Hence, the performance of these models may vary when exposed to different data sets from multiple centers. However, considering the performance of these AI models, there is an urgent need to develop and implement policies to accelerate the process of approval of these models for marketing and usage in clinical scenarios, which can help clinicians in the diagnosis and decision making process.

Conclusions
The past decade has witnessed tremendous advancements in digital diagnostic techniques. AI is a major development that has been successfully implemented in a wide range of image-based applications. These applications can facilitate clinicians in diagnosing, treatment planning, and decision making. These applications are extremely useful as they are reliable and fast methods that have the potential of automatically completing the task with an efficiency equivalent to experienced clinicians. These models can prove to be an excellent guide for less experienced orthodontists. However, there are a few limitations with most of these models, with respect to the limited number of datasets used for training and validating these models, and the reliability of the data, as they are obtained from a single hospital/institution or a single machine. Hence, greater improvisation needs to be conducted in this area for better reliability and generalizability.

Conflicts of Interest:
The authors declare no conflict of interest.