The concept of a strong artificial intelligence (AI) which is capable of mimicking the intelligence of humans has led to, the application of weak AI capable of answering specific questions or solving some disputes¹⁾. The concept of weak AI explores different methods for constructing algorithms which learns from a set of data and make predictions. Machine learning aids in developing algorithms supervised by data²⁾. Neural network (NN) was the earliest types of AI to be built consisting of nodes and weights²⁾. The training data attributes influences the computing skills of the network, allowing it to update the connections weight²⁾. A simple network structure involving a few layers is called a “shallow” learning neural networks, while a “deep” learning neural network structure involves a plentiful and large layers³⁾. Convolutional neural network (CNN) which is a deep learning structure, can be used for extracting many features from abstract filter layers. CNN are commonly useful for processing large and complex figures⁴⁾. The progression made in self-learning back-propagation algorithms has accelerated deep learning by refining the outcome from the data, thereby increasing the computing capability. The advancements in AI today have made it possible to be implemented for solving real-life problems and is applicable in various fields of study⁵⁾.

Since dental implants were first introduced in the 1960s, they have now become the standard treatment method for replacing edentulous area. There are many implant manufacturers, and there are subtle differences in treatment protocols for each company. Jokstad et al. reported that approximately 220 implant brands from 80 companies exist worldwide⁶⁾. If mechanical complications such as screw loosening or screw fracture occur in the implant, accessories provided by the manufacturer should be used to overcome the complications. However, it may be difficult for certain dentists to identify new implant systems simply by only viewing the images of the fixtures in radiographs. Therefore, radiographic identification of implants is especially important to provide appropriate diagnoses and treatments to patients.

For this reason, research to identify implant systems on radiographic images through AI is currently being actively conducted. In this article, papers about AI applied to find and detect dental implant system using dental plain radiography were reviewed.

1. Search strategy

In PubMed and Scopus, a search was performed for ‘Machine learning OR deep learning OR neural network’ and ‘dental implants AND (identification OR detection OR classification)’ until September 2020, and 7, 7, and 9 search results were obtained, respectively. A total of 4 peer-reviewed papers were obtained by removing articles not written in English, those focusing on radiological artifact, papers not related to dental implant, as well as reviews, editorials, and letters.

2. Data extraction

The following items were investigated within each paper; Author and year, Implant system, Data modality, Sample Size, Training time, Validation method, AI method, Accuracy, Sensitivity, Specificity, and Area under the receiver operating characteristic curve (AUC).

Among the items listed, Accuracy, Sensitivity, Specificity, AUC are defined as follows.

$$\text{Accuracy(\%)=}\frac{\text{TP+TN}}{\text{TP+FP+FN+TN}}\text{×100}$$

$$\text{Sensitivity(\%)=}\frac{\text{TP}}{\text{TP+FN}}\text{×100}$$

$$\text{Specificity(\%)=}\frac{\text{TN}}{\text{TN+FP}}\text{×100}$$

TP: True positive, TN: True negative, FP: False positive, FN: False negative.

ROC curve: a curve that visualizes how the evaluation value changes according to the change of each threshold after placing the sensitivity on the y-axis and (1-specificity) on the x-axis.

AUC: The value by obtaining the area of the lower area of the ROC curve has a value between 0~1, and a higher value means better classification performance.

The summary of the data derived from the selected papers is given in Table 1.

Table 1 . Summary of artificial intelligence articles in identification of dental implants system

Author and year	Implant system	Data modality	Sample size	Training	Validation method	AI method	Accuracy (%)	Sensitivity (%)	Specificity (%)	AUC
Lee et al.24) 2020	3 types (Osstem TSIII, Dentium Superline, Straumann Bone level)	Panorama, periapical	10770	1000 epoch	10-fold cross-validation	Pretrained CNN (GoogLeNet Inception-v3)	99.5	95.3	97.6	0.971
Takahashi et al.25) 2020	6 types (Nobel Biocare MK III/IIIG/MKIV/SG, Starumann Bone level, GC Genesio)	Panorama	1282	1000 epoch		Pretrained CNN (YOLO)	51~85	50~82		0.72
Kim et al.26) 2020	4 types (Nobel Biocare MK TiUnite, Dentium Implantium, Straumann Bone level/Tissue level)	Periapical	801	500 epoch	k-fold cross-validation	Pretrained CNN (SqueezeNet, GoogLeNet, ResNet-18, MobileNet-v2, and ResNet-50)	93~98	92~98	94~98
Sukegawa et al.27) 2020	11 types (Zimmer Full Osseotite, Dentsply Astra EV/TX/Microthread, Nobel Biocare MKIII/SG/CC, Kyocera Finesia, Straumann Tissue level)	Panorama	8859	700 epoch	4-fold cross-validation	Pretrained CNN (VGG16 and VGG19)	86.0~93.5	84.2~92.8	80.2~90.7	0.958~1.000

In all studies, CNN was adopted and used as a main network component. The size of the data set was different for each paper, and it ranged from 801 to 10770. In order to maximize the learning effect of the computer, the number of original data was replicated and increased through cross-validation. Then, deep learning was conducted on more replicated data. In the papers, the frequency of learning was set between 500 and 1000 times. Most of them used pretrained networks such as Alexnet, VGG, GoogLeNet, and Inception in existing servers. For radiographs, two-dimensional images such as panorama or periapical view were used alone or in combination. In the case of implant systems, global-market-leading products such as Straumann and Nobel Biocare were commonly investigated in all paper. Additionally, Osstem and Dentium's products were investigated in a Korean author's paper, and Zimmer, Astra, GC, and Kyocera's products were investigated in a Japanese author's paper. The accuracy of AI was 51∼99.5%, and the AUC ranged from 0.72∼1.00.

AI have been studied in various fields within dentistry. Most of them are machine or deep learning using radiographic images or clinical pictures. In the field of oral and maxillofacial surgery, studies have been reported to classify tongue muscles or to provide anatomical guides for segmenting the mandibular canal using AI^7,8). In addition, studies to diagnose oral cancer using tissue slides or cytology images have been reported⁹⁾. There was also a study analyzing the accuracy of automatic diagnosis using ultrasonography in patients with Sjogren syndrome¹⁰⁾. Additionally, there have been studies on the correlation between TMD and TMJ changes on CBCT or MRI¹¹⁾. As for orthodontics, studies on automating measurement points in lateral cephalometry or cone-beam CT have been reported^12-14). On top of that, studies have been reported to discriminate the dentition on which orthodontic treatment was performed, or to analyze the relationship between age and face^15,16). In cariology, studies to detect dental caries from radiographic or infrared images have been conducted^17,18). Also, studies to find normal root morphology, root fracture, and periapical pathogen have also been reported^19-21). In prosthodontics, there have been studies to distinguish between restorations and normal teeth in clinical pictures, or to predict the color tone of porcelain powder or composite resin^22,23).

Researches on dental implants using AI had yet to be conducted only on the determining the implant system. The reason that only studies on implant systems have been conducted is thought that AI research using radiographic images or clinical images is relatively easy to perform. However, clinicians' desperate need for grasping the implant system is probably the main reason. When a patient has undergone an implant procedure at another hospital, or when the clinical information is not available because of an implant treatment long ago, clinicians must infer the implant system only by looking at radiographs. As mentioned in the introduction, when prosthetic complications of implants occur, clinicians need to know the implant system to solve the complications. Considering these issues, an AI based approach seems to be a potentially suitable solution to solve that problem, and most studies were conducted to focus on developing an automated identification system of implants from radiographic images using object detection.

The accuracy of AI was 51∼99.5%, and a large difference in accuracy was observed between literatures. Takahashi et al.'s research showed relatively low accuracy of 51∼85%, while other papers had an accuracy of 90% or more, and Kim's research showed that the accuracy was 93∼98% even though the sample size was only 801^24-27). Presumably, this is likely to be the effect of sample size replication by cross-validation. In the study of Takahashi et al., the content of cross-validation was not mentioned²⁵⁾. Even though the sample size was not small, as a result, accuracy may be low because only a small scale of learning was performed.

Currently, the areas that AI must overcome are largely classified into four categories. The first is data acquisition. Establishing an open-access standard data set will play a more important role since the sample size is ultimately directly related to accuracy²⁸⁾. The second is interpretability. Data-based AI cannot explain the decision-making process because it uses a purely computational method, so it needs to be supplemented to compensate mistake²⁹⁾. The third is Computing Power. As the AI model becomes more complex, the output speed depends on the computing power of the computer, so the advancement of hardware must be accompanied³⁰⁾. The last is Ethical consideration. Since AI does not imply standards on human ethics, development with ethical issues must be done for symbiosis with humans³¹⁾.

Contemporary AI excels in utilizing formalized knowledge and extracting information from massive data sets. However, it is hard to involve human emotion in its algorithm. AI should be viewed as an augmentation tool and, at times, relieve dentists so that they can perform more valued tasks, such as integrating emotional sympathy of patient and improving dentist-patient interactions³²⁾. In dental implantology, where humans are currently in charge of a large part, it is expected that more development will be achieved through the help of AI in the near future.

No potential conflict of interest relevant to this article was reported.