Artificial intelligence (AI) has evolved from the concept of strong AI, which imitates human intelligence, to the implementation of weak AI that can solve certain problems.1 Studies of weak AI explore ways to construct algorithms that can learn from data and make predictions. Machine learning is a branch of computer science that builds algorithms guided by data.2 Among them, neural networks (NNs), which consist of nodes and weights, were one of the first types of AI algorithms to be developed. The computational power of these networks relies on the quality and quantity of training data, which allow these networks to update the weights of the connections. Simple network structures with only a few layers are known as “shallow” learning neural networks, whereas network structures that employ numerous and large layers are referred to as “deep” learning neural networks.3 Deep learning structures referred to as convolutional neural networks (CNNs), which can extract many features from abstracted layers of filters, are mainly used for processing large and complex images. Deep learning is being accelerated by the development of self-learning back-propagation algorithms that progressively refine the results from the data, as well as by increases in computational power. Due to these rapid technological advances, AI, represented by deep learning, can be used for real-life problems and is applied across all sectors of society.4 The diagnostic accuracy of deep learning algorithms in the medical field is approaching levels of human expertise, changing the role of computer-assisted diagnosis from a ‘second-opinion’ tool to a more collaborative one.3 The development of AI applications in the dental field is also remarkable.12 In this article, papers about deep learning applied to the field of oral and maxillofacial radiology will be reviewed.

The data extracted from the selected papers are summarized in Table 1.

In all studies, CNN was used as a main network component, and there were also studies using various other types of networks, such as long short-term memory and siamese networks, in addition to CNNs. CNN-based papers have appeared in the field of dentistry since 2016, and subsequently, more and more dentistry papers using CNN have been published (Fig. 1).

The median size of the datasets used for training also tended to increase, from 100 units to 1000 units (Fig. 2).

Many papers that used pretrained networks such as Alexnet, VGG, GoogLeNet, and Inception v3 showed good results for general purposes.31 However, the structure of CNN networks tends to change from networks with shallow layers to deeper or problem-specific home-made or complex networks.

These studies dealt with various field of dentistry. Most of them were related to teeth, but other subjects such as the gingiva and periodontium, the dental arch, osteoporosis, and anatomical landmarks were also studied using deep learning (Table 2).

Various imaging modalities have been studied in conjunction with the abovementioned subjects. Efforts are underway to diagnose dental disease using traditional 2-dimensional radiographs (intraoral and panoramic), as well as using 3-dimensional cone-beam computed tomography (CBCT). Other studies have investigated new modalities in dental applications, such as quantitative light-induced fluorescence, optical coherence tomography, and the use of intra-oral laser scanners.

Computer assisted diagnosis (CAD) software in the medical field has been used to obtain second opinions, but the design and tuning of conventional CAD tends to be very arduous. Recently, deep learning techniques have been integrated into CAD, with promising results for various medical applications.3233 The qualitative and quantitative applications of deep learning in dentistry are also expanding, but certain areas need to be complemented to promote the continued development of deep learning research in oral and maxillofacial radiology.

However, because all the data sets used in the research analyzed herein were in-house, objective comparison of the studies was difficult. Only a single study tried to evaluate the accuracy of developed networks using other public datasets.23 Efforts are needed to develop a public dataset, such as in the medical field,34 to develop algorithms that can be used in clinical applications. In order to achieve this, researchers need to release the data used in their papers with appropriate removal of personal information, and legal and institutional support from each country is also necessary.3536 There is also a need to build a common, free repository that can reliably collect, catalog, and archive publicly available data in the dental field.

The overall increase in the size of training datasets is desirable for clinical applications of deep learning to the dental field. However, most studies used relatively small data sets (fewer than 1000 units per group), and the accuracy of most studies was less than 90%. This is below the clinically expected accuracy of 98%–99%.37 Deep learning requires a large amount of data because it learns features directly from the data via an end-to-end process. In an anatomical classification study of CT data, at least 1,000 data sets per group were required to achieve 98% validation accuracy with deep learning, and 4,092 data sets per group were required to reach the desired accuracy of 99.5%.38 CBCT, which is the most popular 3D imaging modality in the dental field, does not utilize defined Hounsfield unit values like medical CT, and the pixel values of the acquired images change at every exposure.39 The image quality and magnification of panoramic radiographs, which are commonly used in dental practice, depend on the positioning of the patient.40 Therefore, to achieve clinically meaningful high accuracy, trans-hospital or hybrid data sets from multiple machines and conditions are likely to be needed due to the nature of dental images. For this reason, it is especially important to emphasize the need to construct a large-scale dental public dataset to make the clinical application of deep learning possible.

It is also necessary to emphasize the need for data standardization in the dental field, as well as for standardization of data set construction. In particular, CBCT exhibits large image variation according to brand, machine, and exposure conditions, which can be an obstacle to deep learning research. For example, collecting and learning data on a machine-by-machine basis is difficult because models learned on one machine do not apply to other machines. Although attempts have been made to develop guidelines in Europe, Germany, and England regarding the image quality of CBCT, no international standard has yet been established.41 Therefore, in order for 3-dimensional diagnosis using deep learning to be practical, an international standard for the quality of CBCT images needs to be established in the near future.

Many papers have used preprocessed images via manual cropping of the region of interest. This makes it difficult to analyze and compare results accurately due to errors in the manual process. Some papers91019 have described networks that learned by dividing images into patches of a certain size. However, this method is limited because the network cannot learn the whole image, and instead only focuses on a small part of the image. Some papers212224 used downsampling, which might delete important details of the image. These choices seem to have been made due to limitations in the amount of data or computational power, as indicated in the limitations sections of some papers.2122 However, as computing power per cost increases, it is necessary to use entire images to learn, without any artificial manipulation in the preprocessing stage, in order to obtain more accurate and general results.

Currently, the use of AI is expanding in the medical field. For example, Watson, developed by IBM, has been used to support doctors' clinical decisions.42 However, the clinical accuracy of AI in the dental field must be verified with a variety of cases and imaging modalities due to the difficulty of standardizing dental radiology before AI can take on a more important role in making diagnostic recommendations. Furthermore, current AI algorithms function as black boxes, making it difficult for humans to identify or adjust the criteria used for diagnoses.43 Therefore, in order to increase the reliability of AI, it is necessary to develop a visualization and modification tool for deep learning networks that can be easily understood and edited by humans.