Teeth Reconstruction Using Artificial Intelligence: Trends, Perspectives, and Prospects

ABSTRACT

Background

Artificial intelligence (AI) applications in dental restorative procedures have significantly developed over recent years. However, there is a lack of documentation regarding the types of AI used in tooth reconstruction.

Types of Studies Reviewed

Studies using AI in tooth reconstruction were electronically searched over three databases PubMed, Cochrane library, and Google Scholar. The relevancy of theme tooth reconstruction was prioritized in searching. Only original research was included without the restriction of publication time or any filters.

Results

A total of 18 studies were included in this review: 5 reported programmable AI systems based on mathematical models such as statistical estimation, 9 studies used biogeneric tooth libraries, and 4 presented the deep learning models.

Practical Implications

AI has gained significant progress over the past two decades as a powerful tool for automated tooth reconstruction. However, further studies are required to compare different forms of AIs and to assess their clinical performance in the reconstruction of occlusal surfaces.

KEYWORDS:

• Artificial intelligence
• computer neural networks
• Dentistry 4.0
• prosthodontics
• restorative dentistry

Background

Loss of sound tooth substances and even complete loss of teeth are common oral conditions that dental diseases, jaw pathologies, traumatic injuries, or congenital conditions such as hypodontia and anodontia may cause.^1–4 If untreated, missing teeth/tooth substances tend to deteriorate the oral functions, systemic health, as well as the well-being and quality of life of individuals.^5,6

In restorative dentistry, dental prostheses are a common management approach to restore the appearance as well as oral functions of a patient through the replacement of missing teeth and tooth substances.^7,8 Such treatments should take serious consideration regarding the occlusal surface of dental prostheses because patients can sense minor discrepancy at around 10 µm. The resulting dental prostheses should be as close to the natural teeth as possible regarding occlusal morphologies for the best treatment outcomes.^9–11

Reconstructing original occlusal morphology on a dental prosthesis has always been a challenge for dentists, and in case of failure, it tends to increase the failure and complication rate and may result in oral dysfunctions and impaired oral health-related quality of life (OHRQoL).^12,13

Application of data-driven technologies in the tackling of dental problems is seen as the trend of oral healthcare in the 21^st century, as the dental community strives to become more personalized, as visioned by the Ottawa Charter.^14,15 It is also stated by World Health Organization (WHO) that dentists should apply such technologies, such as artificial intelligence (AI), to achieve quality, accessible and equity of oral and dental care.^16–19

There has been a variety of studies and literature on the application of AI in restorative dentistry in recent years, with most of the efforts involved in diagnosing and detecting dental diseases such as caries, periodontal diseases, and tooth fractures.^20–22 To the best of the authors’ knowledge, studies regarding AIs in teeth reconstruction, especially AIs with machine learning (ML) abilities, are scarce.²¹ This paper aimed to outline an overview of the application of ML techniques in the reconstruction of the occlusal surface, thus providing insights into existing and potential ML applications in restorative dentistry and general dental practice, which may facilitate the adoption of AI technologies in dentistry.

Studies using AI in tooth reconstruction were electronically searched over three databases PubMed, Cochrane library, and Google Scholar, without the restriction of publication time or any filter. The relevancy of theme tooth reconstruction was prioritized in searching. The keywords used were: (“artificial intelligence” or “machine learning” or “deep learning”) AND (“tooth morphology” or “dental crown” or “dental prosthesis” or “prosthodontics”). In this paper, studies of original research data were included and discussed. Review articles were excluded to prevent overlapping of data. Only the most recent piece was included if a single study presented multiple papers at different time points.

Principles of Artificial Intelligence

According to Oxford English Dictionary, AI is “the theory and development of computer systems able to perform tasks that normally require human intelligence, such as visual perception, speech recognition, decision-making, and translation between languages.”²³ It is at the forefront of a new method of computing, known as cognitive computing (CC), and is different from programmable systems, which are based on mathematical principles.^24,25

Categories of Machine Learning

Machine learning is a branch of artificial intelligence (AI) and computer science that focuses on using data and algorithms to imitate how humans learn, gradually improving its accuracy.^26,27 However, machines do not have innate intelligence or cognitive abilities. Humans give them the power to learn and self-improve through experience through machine learning. The various ways of ML may be categorized into three main groups, namely Supervised Learning (SL), Unsupervised Learning (UL), and Reinforcement Learning (RL).^28,29 Each category of Machine Learning caters to different natures of problems and tasks.²⁹

Supervised Learning

Supervised Learning refers to the scenarios in which datasets are labeled for the machines to learn, and such datasets are used to construct a classification model. It also means that Supervised Learning is only applicable when the feature and label of each data point are already known. Supervised Learning may be further categorized into three groups, namely (i) Regression, (ii) Classification, and (iii) Neural Networks.^28,30

Regression

Regression AI models are built on the relationship between input feature(s) χ and result γ, where both are known quantities and γ is a continuous variable (Figure 1). The Regression AI models can be used to estimate continuous values based on input feature(s).³¹ In dentistry, Regression AI models have been adapted for epidemiological studies like predictions of caries to dental image diagnoses.^32–35

Figure 1. Illustration of a regression AI model.

Classification

Classification AI models focus on discrete values that they have identified. They operate by assigning class labels to result γ based on input feature(s) of χ (Figure 2). For example, a classification AI model is given a set of features χ, like fitness, cementation, bonding of the posts, etc. The algorithm classifies the output γ into two labels, True or False, predicting whether the crown will fail. Such models may have more than two outcome labels. They may have multiple forms, including decision trees (Figure 3a), support vector machines (Figure 3b), logistic regression (Figure 3c), and random forests (Figure 3d).^28,30 There have been applications of such AI systems such as detecting cancer cells from magnetic resonance images and labeling dental restorations from radiography.^36–38

Figure 2. Illustration of a classification AI model.

Figure 3. (a) Illustration of a decision tree model. (b) Illustration of a support vector machine model. (c) Illustration of a logistic regression model. (d) Illustration of a random forest model.

Neural networks

Neural Networks mimic the biological structure of human brains. They have their neurons compute information and estimate results (Figure 4). Neural Networks learn from input data and learn to make decisions over time. Such a learning process is called Backpropagation, meaning that the AI matches its outputs to desired outputs to calculate an error function, and the AI then adjusts its results to reduce the error. One collection of neurons, or a Network, is also called a layer, and one layer takes in only one input and gives only one output. If there are multiple layers of Neural Networks, they are called Deep Neural Networks, and the learning process is called Deep Learning.^39–41 Application of neural networks in dentistry started as early as the 1990s, and such explorations included exploring its potential in assisting clinical decision-making and dental age estimation.^42,43

Figure 4. Illustration of a neural networks AI model.

Unsupervised Learning

Unsupervised Learning refers to the case that data is untagged and unstructured; thus, there is no label for the machines. The AI systems in such scenarios need to learn the data and identify any features or labels within the dataset, usually through clustering techniques. Other common techniques in Unsupervised Learning are dimension reduction, density estimation, and market basket analysis. In most scenarios, Unsupervised Learning requires the adoption of neural networks.^44,45

Clustering

Clustering is considered to be one of the most popular unsupervised machine-learning techniques used for grouping data points or somehow similar objects. It is used mostly for discovering structure, summarization, and anomaly detection.^44,46 In medicine and dentistry, it may be used to characterize patients’ behavior based on similar characteristics. It may also be able to identify successful therapies, or causes of treatment failures, for different illnesses.^{47, 48}

For datasets with large amounts of data, partition-based clustering techniques, like K-means, K-Medians, or Fuzzy c-Means, are usually deployed (Figure 5).^49,50 For datasets with a small size of data, hierarchical clustering algorithms as such agglomerative or divisive algorithms are more likely to be used (Figure 6).^51,52 For datasets with significant noise, density-based algorithms may be used (Figure 7).^53,54

Figure 5. Illustration of partition-based clustering.

Figure 6. Illustration of Hierarchical Clustering.

Figure 7. Illustration of density-based clustering.

Means, are usually deployed0

Dimension reduction

Dimensionality reduction, also known as Feature Selection, is operated by reducing redundant features to make grouping easier. By transforming datasets with vast amounts of features, i.e., in high dimension, into refined datasets with selected portions of features, i.e., in low dimension, the AI may analyze data with improved accuracy with comparatively reasonable computing powers (Figure 8).^55,56 Such technique is commonly adopted in studies involving vast numbers of observations and variables, such as medical imaging, bioinformatics, and neuroinformatic.^57–59

Figure 8. Illustration of dimension reduction.

Density estimation

Density estimation is a common statistical concept that is mainly used to explore the data to find some structures within it, and it operates by constructing an estimated probability density function (Figure 9).^60,61 It may be applied in simulation/real data studies and outlier detection applications, like emotion detections, epilepsy and breast cancer diagnosis.^62,63

Figure 9. Illustration of density estimation.

Market basket analysis

Market basket analysis is a modeling technique based upon the theory that if an individual has bought a particular group of items, that person is more likely to buy another group of things, and the model operates by identifying the co-occurrence of matters for the discovery of patterns as well as the frequency of the elements involved.^64,65 This model has a wide range of applications in healthcare, including preventive medications, early warning of diseases, and cancer site detection.^66,67

Reinforcement Learning

Reinforcement Learning is the final of the three basic machine learning categories. It features using “scores” to penalize bad predictions or award good predictions, which then serves as the guidance of the AI model. Such a learning method usually adopts the Markov chains, a stochastic model describing a sequence of possible events in which the probability of each event depends only on the state attained in the previous event. It does not require labeled datasets like Supervised Learning but operates on a trial-and-error basis in the search for outputs with the most rewards (Figure 10).^68,69 Given its strong ability to interact with the environment, this model has been considered an important component of precision medicine, where an AI is trained to adopt an action depending on a patient’s conditions or to optimize treatments and therapies.^70,71

Figure 10. Illustration of reinforcement learning AI model.

Principles of Deep Learning

Deep Learning is a specialized subset within Machine Learning, with multiple layers of artificial Neural Networks to mimic human brains’ biological structures and functionality. Such a system enables an AI to “learn on the job” continuously and automates the improvement of outputs in terms of quality and accuracy. The multi-layer Neural Networks, or Deep Neural Networks, also allow an AI to compute unstructured data like photos, videos, audio, and three-dimensional (3D) models.^72–74

A Deep Learning AI does not directly map inputs to outputs but uses multiple layers of Neural Networks processing units. Each layer passes its respective output to the next. Such a mechanism is also why it is named Deep Learning.⁷³

When a Deep Learning AI is given thousands of datasets, like images, the system runs those datasets through its layered neural network. It adjusts the weights of the variables in each layer of the neural network to detect the common patterns that define images with similar features.^72,75

An advantage of Deep Learning over traditional Machine Learning is that it fixes one of the significant problems in older generations of learning algorithms. While the efficiency and performance of Machine Learning AI plateau after specific amounts of datasets, Deep Learning AI can continue to improve and grow as it is fed with more data (Figure 11).^72,73

Figure 11. Illustration of AI vs machine learning vs deep learning.

Deep Learning AI is proven very efficient in medical imaging and restorative dentistry. Some examples are AIs that may identify gingival inflammation site-specifically or AIs that may automatically reconstruct missing teeth in virtual environments.^22,76–78

Typical Types of Deep Neural Networks

Artificial Neural Networks (ANNs)

ANNs, also known as Feedforward Neural Networks (FNNs), are the simplest and oldest among the Deep Neural Networks families. They are called “feedforward” because the datasets are only processed in one direction, i.e., forward, from the input nodes (Figure 12).⁷⁹ They are adept at learning and modeling non-linear relationships, like oral submucous fibrosis, oral carcinogenesis, and diabetes.^80,81

Figure 12. Illustration of an ANN AI model.

Convolutional Neural Networks (CNNs)

CNNs are multilayer neural networks that take inspiration from animal visual cortex (Figure 13). CNNs adept at applications such as image processing, video recognition, and natural language processing. A convolution is a mathematical operation where a function is applied to another function, and the result is a mixture of the two functions. Convolutions are good at detecting features of an image or an object and putting those features together to reconstruct complex features. In a CNN, this process occurs over a series of layers, each conducting a convolution on the previous layer’s output. They are widely applied in nuclear medicine, dental implant systems, and some disease detection systems.^82–84 An example of such an application is a recent AI system that detects site-specific gingivitis from intraoral photographs.⁸⁵

Figure 13. Illustration of a CNN AI model, reprint under a CC BY-SA 4.0 License.^86,87

Recurrent Neural Networks (RNNs)

RNNs are recurrent because they perform the same task for every sequence element, with prior outputs feeding subsequent stage inputs. Input is processed within a processing unit for a predefined number of times, which is a varied form of backpropagation to optimize the output (Figure 14). RNNs can use information in long sequences, each layer of the network representing the observation at a specific time.^88–90 It is being adopted in radiology and mobile healthcare devices.^91,92

Figure 14. Illustration of an RNN AI model.

Trends of Machine Learning in Restorative Dentistry

Characteristics of a Dental Crown Prosthesis and Its Fabrication

Traditionally dental prostheses were crafted by dental technicians, so the accuracy of occlusal reconstruction was subject to human errors and depended on the experience and skills of dental technicians.⁹³ Moreover, using a semi-adjustable articulator to simulate an individual’s jaw movement and to design the occlusal morphology of dental prostheses is inadequate in reproducing the natural occlusal anatomy.^94,95 While fully adjustable articulators allow better simulation of jaw movements, they require expensive equipment, extensive training, and advanced knowledge. In recent years, such designs were facilitated by the development of computer-aided design-computer-aided manufacturing (CAD-CAM) technology with the advancement of bigeneric tooth models and programmable AI.^96,97 By 2022, CAD-CAM systems were still considered to be better than their AI counterparts in terms of morphology and fracture behaviors, as reported by Chen et al.⁹⁸ A brief overview of all collected studies were listed in Table 1.

Table 1. A brief overview of trend and development of tooth reconstruction AI.

Study	Model	Algorithm	LearnED data	Learning Method	Manual GUIDEs
Alhouri et al.^Citation99	Statistical Model	Mathematical Equations	\	\	Yes
Blanz et al.^Citation100	Statistical Model	Mathematical Equations	\	\	Yes
Mehl et al.^Citation101	Biogeneric Tooth	Mathematical Equations	\	\	Yes
Mehl et al.^Citation102	Biogeneric Tooth	Mathematical Equations	\	\	Yes
Richter et al.^Citation103	Biogeneric Tooth	Mathematical Equations	\	\	Yes
Buchaillard et al. ^Citation104	Statistical Model	Mathematical Equations	\	\	Yes
Steinbrecher et al.^Citation105	Statistical Model	Mathematical Equations	\	\	Yes
Ender et al.^Citation106	Biogeneric Tooth	Mathematical Equations	\	\	Yes
Zheng et al.^Citation107	Biogeneric Tooth	Mathematical Equations	\	\	Yes
Kollmuss et al.^Citation108	Biogeneric Tooth	Mathematical Equations	\	\	Yes
Kwon et al.^Citation109	Biogeneric Tooth	Mathematical Equations	\	\	Yes
Jiang et al.^Citation110	Biogeneric Tooth	Mathematical Equations	\	\	Yes
Hwang et al.^Citation111	Deep Learning	GAN	Technician Crown Designs	Supervised Learning	Yes
Choi et al.^Citation112	Statistical Model	Mathematical Equations	\	\	Yes
Yuan et al.^Citation93	Deep Learning	GAN	Technician Crown Designs and Extracted Occlusal Features	Supervised Learning	No
Wang et al.^Citation113	Biogeneric Tooth	Mathematical Equations	\	\	Yes
Tian et al.^Citation114	Deep Learning	GAN	Extracted Occlusal Features	Supervised Learning	No
Chau et al.^Citation77,Citation78	Deep Learning	GAN	Full Digital Casts	Unsupervised Learning	No

Mathematics Models and Biogeneric Tooth Models (Programmable Systems)

Mathematical analysis of crown contour in comparison with the contralateral tooth in the same arch was proposed in 2004 by Alhouri et al.⁹⁹ The objective was to analyze the oversize or over-contour of dental crowns. Statistical models of tooth morphology were also proposed for the reconstruction of a tooth three-dimensionally when partial information about the morphology of an individual tooth was given.¹⁰⁴ Similarly, the quasi-conformal theory was proposed to build a framework for tooth morphometry regarding planetology and anthropology.¹¹² Then, Blanz et al in 2004 presented an algorithm using statistical methods to reconstruct a 3D model based on identified coordinates of 2D images, which might be used to restore 3D surfaces of teeth through inlay designs.¹⁰⁰ Steinbrecher et al in 2008 proposed automatic construction of inlays and onlays using Laplacian Surface Editing technology.¹⁰⁵ Mesh models from tooth libraries were adopted onto the remaining tooth surface in that system. However, significant manual adjustments were needed to finalize a functional prosthesis design for CAD-CAM fabrication.

The Biogeneric tooth model was also a mathematical construction of a missing surface of a tooth by referring to the remaining tooth substance. Mehl et al reported calculating the average form of maxillary first molar crowns in 2005.¹⁰¹ Biogeneric tooth models using these mathematical calculations were developed to facilitate CAD-CAM designs, virtual reality (VR) applications in dental education, and parametric estimations of tooth morphology.¹⁰² In 2006, Richter et al applied biogeneric tooth models to design inlays and onlays of occlusal surfaces of multiple mandibular molars. They reported that reconstruction using biogeneric tooth models was feasible with an average deviation of 150 micrometers.¹⁰³

Another study reported that the resulting prostheses could reconstruct better morphology within a shorter period. They featured comparably fewer needs of occlusal adjustment time against conventionally CAD-CAM-designed prostheses.¹⁰⁶ Later, two studies compared the occlusal morphology of crowns developed by software based on the biogeneric tooth model and those designed by human dental technicians.^108,115 They reported that reconstructions made by the CAD-CAM software were closer to the original teeth than the human-designed counterparts. However, Kwon et al argued that the biogeneric tooth models were not as accurate as described when matched against the morphology of naturally healthy teeth, as the reproducibility of the CAD-CAM occlusal construction was relatively low.¹⁰⁹

Biogeneric tooth model technology was further developed and implemented into CAD-CAM software to fabricate inlays, onlays, and partial crowns that mimic natural tooth morphologies. While some clinicians found that the clinical results of using the biogeneric tooth model are less than satisfactory,¹¹⁶ there are several commercial CAD-CAM systems, including the CEREC system (Dentsply Sirona), E4D Dentist system (D4D Technology LLC), Procera Forte system (Nobel Biocare), and 3Shape Dental system.¹¹³ In 2011, Zheng et al introduced a 3D morphing technique for tooth reconstruction, especially for inlays and onlays, which included 3 steps: cavity contour extraction, points identification, and surface deformation.¹⁰⁷ This model was only for occlusal surface reconstruction of teeth with an existing cavity or defective areas on occlusal surfaces. Then, Jiang et al 2016 proposed another tooth reconstruction algorithm for all-ceramic crown design, with 3 steps: initial tooth positioning, points extraction, and tooth deformation.¹¹⁰ This system was reported to be able to generate a reconstruction of a tooth within 5 seconds, though chairside adjustments were still needed.

Machine Learning Models and Deep Learning Models

Deep learning was considered a promising way to automate the design of dental prostheses. However, it usually required a large number of training datasets to train the AI on dental prostheses designs and occlusal relationships.^73,75,114 Non-deep learning models for restorations seemed to be rare in literature, and it might be due to the limitation of single-layer neural networks, meaning they were lack of appropriate computing capabilities for complex feature extractions and interpretations.¹¹⁷

Hwang et al proposed a Generative Adversarial Network (GAN) model in 2018 to develop a crown design based on intraoral 3D scans. They created 2D scan images of the arch with prepared teeth, the opposite arch, and the occlusal gap between two arches from the original intraoral 3D scans. The GAN model then learned the 2D images and dental technicians’ designs for the teeth in question to generate crown designs using CAD-CAM tools. In that study, statistical calculations were performed regarding the occlusal morphology of restorations and the occlusal relationship between the prepared teeth and their opposing teeth. Still, the designs were not tested for biomimetic features or clinical applications.¹¹¹

Yuan et al proposed an improved GAN model in 2020 for tooth surface reconstruction based on conditional generative adversarial networks (CGANs), which considered the anatomical occlusal surfaces of teeth in question and their occlusal relationship with respective opposite teeth through 2D images of their occlusal surfaces.⁹³ The system was a combined application of CGANs and techniques of perceptual loss for reconstructing a defective tooth, i.e., referring to the opposite tooth for occlusal information. The proposed AI still learned the occlusal morphology of teeth from the designs of dental technicians and worked on prepared teeth, similar to the study of Hwang et al, though this system applied both 2D and 3D training datasets.

Another GAN proposed by Tian et al 2022 involved 2 steps: learning the occlusal relationships of a prepared tooth and its neighboring teeth, as well as those of the opposite tooth and its adjacent teeth; the system then used these “occlusal fingerprints” to enrich the functional characteristics of generated designs.¹¹⁴ This system was specifically designed for inlays, onlays, and single-crown fabrications, though it still required a prepared tooth to function, similar to Hwang et al and Yuan et al. However, this system considered natural occlusal morphologies of neighboring tooth structures. The outputs were compared to natural healthy teeth instead of technicians’ designs, making it different from the previous studies.

The most recent GAN development was proposed by Chau et al, which was designed to generate a single-tooth replacement for missing teeth.^77,78 The 3D GAN system learned the occlusal relationships among all remaining dentitions only through 3D datasets, which consisted of digital casts taken from authentic participants with naturally healthy teeth.

Opportunities and Challenges

Several forms of AIs have been used to assist in the design of tooth reconstruction over the past 20 years, though they still cannot substitute the roles of dentists or dental technicians at the moment. Several limitations are observed among the collected studies. One of such limitations is that most AI projects are based on mathematic models that do not self-improve with experience. Furthermore, the algorithms use supervised learning which learns the specific patterns decided by the dentists and thus cannot discover new patterns for better learning, if any, as in unsupervised learning.^72,73,75 More investigations into Deep Learning in tooth reconstruction and comparisons of different learning algorithms may be needed.

Data dimensions seem to be another potential room for improvement in AI studies. In most of the collected AI studies, only 2D data are used to train the AI system for 3D tooth reconstruction. The use of 3D teeth models for learning is considered better as they provide more information for learning and tend to be more accurate, avoiding the angular distortion of capturing 2D images.^118,119 Further investigations into the improvement of using 3D data in tooth reconstruction may be necessary to assess if the quality and accuracy of the AI system are improved.

The collection of adequate samples is another critical point in the field of machine learning. Unless the training datasets are homogenous and thus enable few-shot learning, i.e., training with a small sample size, a proper Deep Learning model with 3D data could require up to 1000 sets of training data to achieve minimum acceptable accuracy.^119–123

Lastly, all Deep Learning studies collected involve no clinical trials, and thus there are no assessments of their clinical performance, especially in terms of dynamic occlusion, aesthetic, patient acceptance, etc. Therefore, further investigations would be necessary before a broader application of Machine Learning models in restorative dentistry.

Acknowledgement

The work described in this paper was fully supported by a grant from the Research Grant Council of Hong Kong, China (Grant number 17126021).

Correction Statement

This article has been corrected with minor changes. These changes do not impact the academic content of the article.

Additional information

Notes on contributors

Reinhard Chun Wang Chau

Reinhard C.W. Chau, BMed, MSc is a clinical research coordinator in the division of restorative dental sciences at the Faculty of Dentistry, The University of Hong Kong. He specializes in community dentistry and holds a membership from the Royal College of Surgeons of England. He is a council member of the Section Council of Odontology of the Royal Society of Medicine. His research focuses on digital dentistry, especially the adoption and integration of data-driven technologies in the facilitation of population oral health.

Khaing Myat Thu

Khaing Myat Thu, BDS, DipDSc, PhD is a senior research assistant in the division of restorative dental sciences at the Faculty of Dentistry, The University of Hong Kong. He has been conducting research regarding oral rehabilitation since 2015.

Richard Tai Chiu Hsung

Richard T.C. Hsung, BEng, PhD is an associate professor at Chu Hai College of Higher Education, Department of Computer Science, and an honorary associate professor in the division of Oral and Maxillofacial Surgery at the Faculty of Dentistry, The University of Hong Kong. He is also a member of IEEE, IET. His research interests include wavelet theory and applications, tomography, fast algorithms, biomedical signal processing, computer vision, and neural network. He has developed natural head recording methods, CBCT segmentation of the human orbital cavity, virtual patient model-building methods, and acoustic-based obstructive sleep apnea detection.

Walter Yu Hang Lam

Walter Y.H. Lam, BDS, MDS, AdvDip Prosth is a clinical assistant professor in Prosthodontics at the Faculty of Dentistry, The University of Hong Kong, He is a prosthodontist and holds Fellowships from several Dental Colleges. He is an international dental advisor of the Royal College of Physicians and Surgeons of Glasgow and the Secretary of the Hong Kong Prosthetic Dentistry Society. He is the Editorial Board Member of the Digital Dentistry Section, Journal of Dentistry, and the Secretary of the Digital Dentistry Research Network, International Association of Dental Research (IADR). He is a founding member of the HKU Musketeers Foundation Institute of Data Science and a Hong Kong representative in the Dental CAD/CAM Systems (TC 06/SC 9), International Standard Organization (ISO). His research interests include Prosthodontics, Digital Dentistry, and Artificial Intelligence.