Artificial intelligence is not a new term. Alan Turing wrote in his paper “Computing Machinery and Intelligence” (9) in the 1950 issue of Mind:

Back then, there was no term to interpret AI; Turing described AI as “machines thinking”. He mathematically investigated the feasibility of AI and explored how to construct intelligent machines and assess machine intelligence. He proposed that humans solve problems and make decisions by utilising available information and inference, machines also can do the same thing.

In the paper (9), Turing proposed setting a test as to whether a machine can achieve human-level intelligence. This test is known as the Turing Test. It lies on the following lines: Assuming a human evaluator could distinguish natural language communications between a human test taker and a machine. It is given that a human evaluator knows that the conversation is between a human and a machine, and the human evaluator, human test taker and machine are separated from one another. The conversation between the human test taker and the machine is limited to plain text, i.e., keyboard input, instead of speech. This is to make the test only focus on the machine's ability to answer the questions logically instead of testing its speech interpretation ability. If the human evaluator cannot distinguish the human test taker and the machine, the machine can be viewed as having passed the Turing Test, and such a machine is said to have “machine intelligence”.

Later, in 1955, the term AI was first proposed in a 2-month workshop: Dartmouth Summer Research Project on Artificial Intelligence (10) led by John McCarthy, Marvin Minsky, Nathaniel Rochester, and Claude Shannon. However, the concept was only on paper. Certain restrictions stopped researchers from developing real AI machines in the 1950s. Firstly, computers before 1949 lacked a fundamental prerequisite for AI tasks: there was no storage function, which means the codes could not be stored, they could only be executed. Secondly, computers were costly at that time. Lastly, funding sources had conservative attitudes towards this new field back then (11).

From 1957 to 1974, the AI field was fast-growing because of the growth of computer power, its accessibility, and AI algorithms. Examples include ELIZA, a computer program that could interpret spoken language and solve problems via text (12). Two “AI Winters” arrived after the first wave of development due to insufficient practical applications and research funding reduction in the mid-1970s and late 1980s (13). However, AI had its breakthrough between the two periods with very few developments. In the 1980s, it developed through two paths: machine learning (ML) and expert systems. They are two opposite approaches to AI considering their theory. ML allows computers to learn by experience (14); expert systems, on the contrary, simulate the decision-making process of human experts (15). In other words, ML finds the solution by learning and summarizing the experience by itself, while expert systems need human experts to input all possible situations and solutions in advance. Expert systems have largely been used in industry since then. The example includes R1 (Xcon) program, an expert system with around 2,500 rules for assisting components selection for computer assembly was developed (16) and used by DEC, a computer manufacturer.

Two important time points in computer vision are 2012 and 2017. In 2012, a graphics processing unit (GPU)-implemented deep learning (DL) network with eight layers was developed (17), The work won the ImageNet Large Scale Visual Recognition Challenge (ILSVRC) and achieved a classification top-5 error of 15.3%. The error rate was more than 10.8% lower than the runner-up. In 2017, SE-NET further lowered the top-5 error to 2.25%, surpassing the human top-5 error (5.1%) (18).

Later famous AI examples include Deep Blue—a chess-playing expert system, which defeated chess champion of the time Gary Kasparov in 1997 (19); 20 years later in 2017, Google's AlphaGo, a DL program, defeated the world No. 1 ranked player Jie Ke in a Go match (20); recently in late 2022, OpenAI launched ChatGPT (Chat Generative Pre-trained Transformer), it is a text-generation model that can generate human-like responses based on text input, the model received extensive discussion since its launch (21). These examples used different AI approaches to operate.

There are many approaches whereby AI can be achieved; different types of AI can achieve different tasks, and researchers have created different AI classification methods.

AI is a generic term for all non-human intelligence. As Figure 1 shows, AI can be further classified as weak AI and strong AI. Weak AI, also called narrow AI, uses a program trained to solve single or specific tasks. The AI of today is mostly weak AI. Examples include reinforcement learning, e.g., AlphaGo, and automated manipulation robots; natural language processing, e.g., Google translation, and Amazon chat robot; computer vision, e.g., Tesla Autopilot, and face recognition; data mining, e.g., market customer analysis and personalised content recommendation on social media (22). Strong AI refers to the ability and intelligence of AI equalling that of humans—it has its own awareness and behaviour as flexible as humans (23). Strong AI aims to create a multi-task algorithm to make decisions in multiple fields. Research on strong AI has to be very cautious as there might be ethical issues, and it could be dangerous. Thus, there are no strong AI applications up to now.

ML and expert systems are two different subgroups of weak AI. As shown in Table 1, ML can be further classified as supervised, semi-supervised and unsupervised learning based on the theory of the methods. Supervised learning uses labelled datasets for training, and these labelled datasets are the “supervisor” of the algorithm. The algorithm learns from the labelled input, and extracts and identifies the common features of the labelled input to make predictions about unlabelled input (24). Examples of supervised learning includes k-nearest neighbors, logistic regression, random forest, and support-vector machine (25). Unsupervised learning, on the contrary, works on its own to find the various features of unlabelled data (26). Semi-supervised learning lies between those two, which utilises a small amount labelled data together with a large amount of unlabelled data during training (27). Recently, a new method called weakly-supervised learning became increasingly popular in the AI field to alleviate labelling costs. In particular, the object segmentation task only uses image-level labels (i.e., only knowing what objects are in the images) instead of object boundary or location information for training (28).

Deep learning is currently a very prominent research area and forms a subset of ML. It can involve both supervised and unsupervised learning. As Figure 2 shows, “deep” represents an artificial “neural network” consisting of a minimum of three nodal layers—input, multiple “hidden”, and output layers such that each layer consists of various numbers of interconnected nodes (artificial neurons) whereas each node x has an associated weight (w_i) and biased threshold (t) from m decisive factors, given by its own (simplified) linear regression model . The weight is assigned when there is an input of the node. If ∑i=1mwixi+t≥0, then the output = 1, meaning the data is passed to another node in another layer. The process of passing data from one layer to the next defines the neural network as a feedforward network, similar to a decision tree model.

As mentioned above, a deep neural network can extract features from the imported data, which does not require human intervention. Instead, it can learn those features from large datasets. On the other hand, expert systems require human intervention to learn, which indeed tuning the w_i and t manually. So, less data is required.

Neural networks (NNs) are biologically inspired networks that can be regarded as the pillars of deep learning algorithms. There are different variations of NNs, among which the most important types of neural networks are artificial neural networks (ANNs), convolution neural networks (CNNs), and generative adversarial networks (GANs).

Traditionally, dentists diagnose caries by visual and tactile examination or by radiographic examination according to a detailed criterion. However, detecting early-stage lesions is challenging when deep fissures, tight interproximal contacts, and secondary lesions are present. Eventually, many lesions are detected only in the advanced stages of dental caries, leading to a more complicated treatment, i.e., dental crown, root canal therapy, or even implant. Although dental radiography (whether panoramic, periapical, or bitewing views) and explorer (or dental probe) have been widely used and regarded as highly reliable diagnostic tools detecting dental caries, much of the screening and final diagnosis tends to rely on dentists' experience.

In operative dentistry, there has been research on the detection of dental caries, vertical root fractures, apical lesions, pulp space volumetric assessment, and evaluation of tooth wear (39–44) (Table 2). In a two-dimensional (2D) radiograph, each pixel of the grayscale image has an intensity, i.e., brightness, which represents the density of the object. By learning from the above-mentioned characteristics, an AI algorithm can learn the pattern and give predictions to segment the tooth, detect caries, etc. For example, Lee et al. (45) developed a CNN algorithm to detect dental caries on periapical radiographs. Kühnisch et al. (46) proposed a CNN algorithm to detect caries on intraoral images. Schwendicke et al. (47) compared the cost-effectiveness of AI for proximal caries detection with dentists' diagnosis; the results showed that AI was more effective and less costly.

Several studies mentioned above showed that AI has promising results in early lesion detection, with accuracy the same or even better compared with dentists. This achievement requires interdisciplinary cooperation between computer scientists and clinicians. The clinicians manually label the radiographic images with the location of caries while the computer scientists prepare the dataset and ML algorithm. Finally, clinicians and computer scientists jointly check and verify the accuracy and precision of the training results (48).

Periodontitis is one of the most widespread diseases. It is a burden for billions of individuals and, if untreated, can lead to tooth mobility and even tooth loss (49). To prevent severe periodontitis, early detection and treatment are needed. In clinical practise, periodontal disease diagnosis is based on evaluating pocket probing depths and gingival recession. The Periodontal Screening Index (PSI) is frequently used to quantify clinical attachment loss. However, this clinical evaluation has low reliability: the screening for periodontal disease is still based on the experience of dentists, and they may miss localized periodontal tissue loss (50).

In periodontics, AI has been utilised to diagnose periodontitis and classify plausible periodontal disease types (51, 52). In addition, Krois et al. (50) adopted CNN in the detection of periodontal bone loss (PBL) on panoramic radiographs. Lee et al. (53) evaluated the potential usefulness and accuracy of a proposed CNN algorithm to detect periodontally compromised teeth automatically. Yauney et al. (54) claimed that periodontal conditions could be examined by a CNN algorithm developed by their research group using systemic health-related data (Table 3).

Orthodontic treatment planning is usually based on the experience and preference of the orthodontists. As every patient and orthodontist is unique, the treatment is decided mutually by both sides. Traditionally, it takes a lot of effort for orthodontists to diagnose malocclusion, as many variables need to be considered in the cephalometric analysis, such that it is difficult to determine the treatment plan and predict the treatment outcome (55). AI is an ideal tool for solving orthodontic problems. In orthodontics, AI has applications (Table 4) in treatment planning and prediction of treatment results, such as simulating the changes in the appearance of pre- and post-treatment facial photographs. The impact of orthodontic treatment, the skeletal patterns, and the anatomic landmarks in lateral cephalograms (67) can be clearly seen with the aid of AI algorithms, greatly assisting communication between patients and dentists.

A Bayesian-based decision support system was developed by Thanathornwong (57) to diagnose the need for orthodontic treatment based on orthodontics-related data as input. Xie et al. (58) proposed an ANN model to evaluate whether extractions are needed from lateral cephalometric radiographs; A similar evaluation system was proposed by Jung et al. (59). Apart from the application in predicting the extractions needed for orthodontic purposes, AI has been adopted to locate cephalometric landmarks. Park et al. (60, 61) demonstrated a DL algorithm for the automatically identifying cephalometric landmarks on radiographs with a high accuracy. Bulatova (68) et al. and Kunz et al. (69) developed similar AI algorithms, with accuracies comparable with human examiners in identifying those landmarks. An automatic system for skeletal classification using lateral cephalometric radiographs was proposed by Yu et al. (63).

Besides locating multiple cephalometric landmarks and classification, AI systems have been used in orthodontic treatment planning. Choi et al. (64) proposed an AI model to judge whether surgery is needed using lateral cephalometric radiographs. It looks like most of the orthodontic applications are on landmarking identification and treatment planning, which are tedious procedures for orthodontists. A basic task for orthodontic treatment planning is to segment and classify the teeth. AI has also been used for these purposes on multiple sources, such as radiographs and full-arch 3D digital optical scans (65, 66). Cui et al. proposed several AI algorithms to automatically segment teeth on a digital teeth model scanned by a 3D intraoral scanner (65) and CBCT images (66, 70). In addition to tooth segmentation, they also segmented alveolar bone, the efficiency exceeded the radiologists' work (i.e., 500 times faster). The paper also claimed that the algorithm works well in challenging cases with variable dental abnormalities (66).

Oral and Maxillofacial Pathology (OMFP) is a specialty for examining pathological conditions and diagnosing diseases in the oral and maxillofacial region. The most severe type of OMFP is oral cancer. Statistics from the World Health Organization (WHO) show that every year there are over 657,000 patients diagnosed with oral cancer globally, among which there are more than 330,000 deaths (71). In OMFP, as shown in Table 5, AI has been researched mostly for tumour and cancer detection based on radiographic, microscopic and ultrasonographic images. Besides, abnormal locations can also be detected from radiographs by AI (72), such as nerves in the oral cavity, interdigitated tongue muscles, and parotid and salivary glands. CNN algorithms were demonstrated to be a suitable tool for the automatically detecting cancers (73, 78). It is worth mentioning that AI also plays a role in managing cleft lip and palate in risk prediction, diagnosis, pre-surgical orthopaedics, speech assessment, and surgery (79).

Early detection and diagnosis of various mucosal lesions are essential to classify benign or malignant. Surgery resection is required for malignant lesions. However, some of the lesions behave similarly in appearance, thus requiring the diagnosis by biopsy slides and radiographs. Pathologists diagnose disease by observing the morphology of stained specimens on glass slides using microscopic (80). It is tedious work that requires much of effort for pathologists. Of all the biopsies that need to be examined, only around 20% of them are found to be malignancies. Thus, AI can be a suitable tool for aiding pathologists in this task.

Warin et al. (74) used a CNN approach to detect oral potentially malignant disorders (OPMDs) and oral squamous cell carcinoma (OSCC) in intraoral optical images. In addition to intraoral optical images, OCT has been used in identify benign and malignant lesions in the oral mucosa. James et al. (75) used ANN and SVM models to distinguish malignant and dysplastic oral lesions. Heidari et al. (76) used a CNN network, AlexNet (17), to distinguish normal and abnormal head and neck mucosa. Abureville et al. (73) used a CNN algorithm to automatically diagnose oral squamous cell carcinoma (SCC) from confocal laser endomicroscopy images; the study showed that the CNN algorithm used in the study was especially suitable for early diagnosis of SCC. Poedjiastoeti et al. (77) also used a CNN algorithm to identify and distinguish ameloblastoma and keratocystic odontogenic tumour (KCOT). The two oral tumours with similar features in radiographic images. By comparing the computer-generated results with the biopsy results, the accuracy of the CNN algorithm was found to be 83% and the diagnostic time 38 s. These values were similar to those of oral and maxillofacial specialists.

In prosthodontics, a typical treatment process to prepare a dental crown includes tooth preparation, impression taking, cast trimming, restoration design, fabrication, try-in, and cementation. The application of AI in prosthodontics mainly lies in the restoration design (Table 6). CAD/CAM has digitalised the design work in commercialized products, including CEREC, Sirona, 3Shape, etc. Although this has dramatically increased the efficiency of the design process by utilising a tooth library for crown design, it still cannot achieve a custom-made design for individual patients (81). With the development of AI, Hwang et al. (82) and Tian et al. (83) proposed novel approaches based on 2D-GAN models to generate a crown by learning from technicians' designs. The training data was 2D depth maps converted from 3D tooth models. Ding (84) reported a 3D-DCGAN network in the crown generation, which utilised 3D data directly in the crown generation process, the morphology of generated crowns was similar compared with natural teeth. Integrating AI with CAD/CAM or 3D/4D printing can achieve a more desirable workflow with high efficiency (88). AI has also been used in shade matching (85) and debonding prediction of CAD/CAM restorations (86).

Apart from fixed prosthodontics, the design in removable prosthodontics is more challenging as more factors and variables need to be considered. No ML algorithm is available for the purpose of designing removable dentures while several expert (knowledge based) systems have been introduced (89–91). Current ML algorithms are more focused on assisting the design process of removable dentures, e.g., classification of dental arches (87), and facial appearance prediction in edentulous patients (92).