This is an open-access article distributed under the terms of the Creative Commons Attribution License (https://creativecommons.org/licenses/by/4.0/), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work, first published in JMIR Medical Informatics, is properly cited. The complete bibliographic information, a link to the original publication on https://medinform.jmir.org/, as well as this copyright and license information must be included.
Early detection of tooth-related diseases in patients plays a key role in maintaining their dental health and preventing future complications. Since dentists are not overly attentive to tooth-related diseases that may be difficult to judge visually, many patients miss timely treatment. The 5 representative tooth-related diseases, that is, coronal caries or defect, proximal caries, cervical caries or abrasion, periapical radiolucency, and residual root can be detected on panoramic images. In this study, a web service was constructed for the detection of these diseases on panoramic images in real time, which helped shorten the treatment planning time and reduce the probability of misdiagnosis.
This study designed a model to assess tooth-related diseases in panoramic images by using artificial intelligence in real time. This model can perform an auxiliary role in the diagnosis of tooth-related diseases by dentists and reduce the treatment planning time spent through telemedicine.
For learning the 5 tooth-related diseases, 10,000 panoramic images were modeled: 4206 coronal caries or defects, 4478 proximal caries, 6920 cervical caries or abrasion, 8290 periapical radiolucencies, and 1446 residual roots. To learn the model, the fast region-based convolutional network (Fast R-CNN), residual neural network (ResNet), and inception models were used. Learning about the 5 tooth-related diseases completely did not provide accurate information on the diseases because of indistinct features present in the panoramic pictures. Therefore, 1 detection model was applied to each tooth-related disease, and the models for each of the diseases were integrated to increase accuracy.
The Fast R-CNN model showed the highest accuracy, with an accuracy of over 90%, in diagnosing the 5 tooth-related diseases. Thus, Fast R-CNN was selected as the final judgment model as it facilitated the real-time diagnosis of dental diseases that are difficult to judge visually from radiographs and images, thereby assisting the dentists in their treatment plans.
The Fast R-CNN model showed the highest accuracy in the real-time diagnosis of dental diseases and can therefore play an auxiliary role in shortening the treatment planning time after the dentists diagnose the tooth-related disease. In addition, by updating the captured panoramic images of patients on the web service developed in this study, we are looking forward to increasing the accuracy of diagnosing these 5 tooth-related diseases. The dental diagnosis system in this study takes 2 minutes for diagnosing 5 diseases in 1 panoramic image. Therefore, this system plays an effective role in setting a dental treatment schedule.
Several recent studies [
Dental caries is diagnosed using videos and radiographs, and studies [
Dental caries is one of the most common chronic diseases worldwide. Oral diseases are recognized as serious diseases like other systemic diseases and were classified by the World Health Organization in 2011 as serious noncommunicable diseases. The teeth are one of the most important organs in the body, and dental caries is one of the biggest causes of tooth disease [
Although various simple and easy AI diagnostic methods in the dental field have been studied, there are limitations [
Since this study evaluated 5 tooth-related diseases (ie, coronal caries or defects, proximal caries, cervical caries or abrasion, periapical radiolucency, and residual root), which are commonly diagnosed using dental imaging, training data were generated by collecting and labeling panoramic images with tooth-related diseases. This study used panoramic images provided by 50 dental clinics from 2001 to 2021. Data from 30 dental hospitals in Korea were collected, anonymized, and used for this study. Among the anonymized genders, there were 3702 males and 3783 females, with a total of 2515 unidentified persons who could not be identified. Population distribution by age group did not include teenagers; there were 1721 persons in their 20s, 956 persons in their 30s, 1134 persons in their 40s, 1351 persons in their 50s, 1914 persons in their 60s and older, and 2934 persons with unknown identities.
A total of 10,000 panoramic images with one or more of the following 5 tooth-related diseases were used for labeling: 4206 images of coronal caries or defects, 4478 images of proximal caries, 6920 images of cervical caries or abrasion, 8290 images of periapical radiolucency, and 1446 images of residual roots. As shown in
Findings of each tooth-related disease (ie, coronal caries or defect, proximal caries, cervical caries or abrasion, periapical radiolucency, residual root, in clockwise order from the top left).
Findings of each tooth-related disease.
Tooth-related diseases | Findings |
Coronal caries or defect | Defect or radiolucency that lacks density compared to normal in the coronal portion of a tooth |
Proximal caries | Radiolucency that lacks density compared to normal in the adjacent surfaces between teeth |
Cervical caries or abrasion | Radiolucency that lacks density compared to normal in the cervical area of the tooth |
Periapical radiolucency | Low density compared to normal in the periapical area of tooth |
Residual root | Coronal portion of tooth is completely lost and only the root portion remains |
Labeling was performed by data collection and preprocessing, and thus, an image classification model was used to learn about each of the 5 tooth-related diseases. This study learned dental diseases by using fast region-based convolutional network (Fast R-CNN), residual neural network (ResNet), and inception. The model with the highest accuracy in disease detection was selected. For training the model, 10,000 panoramic images were modeled in total: 4208 coronal caries or defects, 4478 proximal caries, 6920 cervical caries or abrasion, 8290 periapical radiolucency, and 1446 residual roots.
Fast R-CNN has increased accuracy compared to the existing object detection algorithms because it extracts the image features and minimizes the noise in image analysis. Fast R-CNN consists of a convolution feature map and a region of interest feature vector [
For image reading, a rectangular bounding box was first used, and segmentation was performed through an algorithm based on about 500 segmentation data. In the case of segmentation, accuracy was not calculated for the segmented data because it was used only for grasping the approximate accuracy. Thereafter, the coordinate values of the box-type tooth classes that are multilabeled in 1 tooth panoramic image were derived. Each disease corresponding to the derived coordinate value was classified by class. Then, each of the 5 tooth classes was applied to learning through the box coordinate values having the corresponding dental disease on the panoramic image. Through this, the input value for 1 model was constructed using the panoramic image data of 1 class and the box coordinate values corresponding to dental diseases. As shown in
ResNet derives a value through the weight layer to solve the problem of overfitting owing to increased dimensional depth in deep learning, which adds the result learned through the previous weight layer to the activation function and delivers it to the next layer [
Inception, like ResNet, is created to solve the overfitting problem and the increase in computational traffic through a lot of learning when the size of the model is increased by increasing the depth of the layer [
Case of misclassification of tooth-related diseases. The green boxes represent detected areas that are not included in the teeth.
Bounding box for each tooth-related disease.
Integrated detection system for the learning process. DICOM: digital imaging and communications in medicine.
While learning about the 5 tooth-related diseases through Fast R-CNN, ResNet, and inception, problems, as shown in
Based on the process shown in
In the training layer structure of each model of Fast R-CNN, ResNet, and inception, looking at the structure of the Fast R-CNN model (region proposal → CNN classification → region of interest projection), region of interest projection and bounding box regression were performed through region of interest pooling. The model is configured as shown in
ResNet improves the accuracy by reducing the depth of the learning layer and increasing the performance compared to the CNN model, which is an existing image analysis model, through residual learning. In order to increase the learning accuracy in general CNNs, many layers are stacked. However, such a deep layer can lower the accuracy of the learning model. When learning through residual learning, the positive error rate can be lowered even when learning in a deep layer. When ResNet derives a value from the weight layer through the activation function in the convolution operation, it imports the previously learned information as it is, as shown in
In this way, by adding information to the result derived from the weight layer, information can be added and computational complexity can be reduced so that a model with faster learning and better performance can be derived. Since ResNet learns 1 dental disease by using 5 models as 1 model, 50 hidden layers of each dental disease were designated for learning. For training, like Fast R-CNN, the training epoch was performed 100,000 times, and the learning rate was set to 0.001.
The inception model connects the highly correlated nodes when the correlation between each node is high in the fully connected architecture and does not connect the rest so that N clusters are created for each feature. When creating a connected architecture, we additionally convolve features that are far from each other through filters of various sizes for nonuniform and inefficient sparse structures and reduce the number of channels by using a 1 × 1 filter for nodes with high correlation. The inception model was constructed, as shown in
The overall flow diagram is shown in
Based on the process shown in
Flow diagram of the learning process. AI: artificial intelligence; PC: personal computer.
Schematic diagram of a detection system for tooth-related diseases. AI: artificial intelligence; API: application programming interface; DB: database; PACS: picture archiving and communication system; PC: personal computer.
Since the data is a retrospective study, it was processed in the direction of protecting personal information through database anonymization, etc. In addition, data collected for research purposes were collected through Cheongju University Bioethics Committee IRB (1041107-202208-HR-024-01).
The web service in this study was built based on the process presented in
Success and fail cases in the detection of the 5 dental diseases.
Web system for tooth-related diseases.
To detect tooth-related diseases, Fast R-CNN, which has the best performance among the classification models from the point of view of a dentist, was applied. For learning the model for judging the 5 types of tooth-related diseases, dentists can update the panoramic images in real time through the web application programming interface and continuously collect data by improving the accuracy through additional updates on the tooth-related disease detection labeling.
This study created a detection model for 5 dental diseases that are difficult to judge visually (ie, coronal caries or defect, proximal caries, cervical caries or abrasion, periapical radiolucency, and residual root) by using a dental panoramic image. Fast R-CNN, ResNet, and inception have previously been used to learn about dental disease detection [
Tooth-related disease detection results.
Model, diseases | Precision | Sensitivity | Specificity | |
|
||||
|
Coronal caries or defect | 0.785 | 0.708 | 0.982 |
|
Proximal caries | 0.484 | 0.792 | 0.918 |
|
Cervical caries or abrasion | 0.795 | 0.767 | 0.952 |
|
Periapical radiolucency | 0.824 | 0.953 | 0.895 |
|
Residual root | 0.640 | 0.904 | 0.972 |
|
||||
|
Coronal caries or defect | 0.253 | 0.609 | 0.848 |
|
Proximal caries | 0.327 | 0.783 | 0.883 |
|
Cervical caries or abrasion | 0.444 | 0.707 | 0.785 |
|
Periapical radiolucency | 0.371 | 0.946 | 0.556 |
|
Residual root | 0.232 | 0.893 | 0.873 |
|
||||
|
Coronal caries or defect | 0.2101 | 0.395 | 0.876 |
|
Proximal caries | 0.685 | 0.377 | 0.987 |
|
Cervical caries or abrasion | 0.378 | 0.011 | 0.996 |
|
Periapical radiolucency | 0.308 | 0.883 | 0.451 |
|
Residual root | 0.225 | 0.744 | 0.89 |
This study has several advantages. The use of panoramic images of individual patients in dentistry is a complex procedure. This study designed a model that could determine 5 types of dental caries by acquiring various panoramic image data and collecting 10,000 pieces of data with various oral structures and dental caries. Therefore, a tooth-related disease determination system with high accuracy and without complex procedures was developed. However, since there is a large deviation in the number of classes for each tooth-related disease, there was a problem in that the learning accuracy was slightly lowered where the number of analysis groups was small. The accuracy of the model is expected to be improved by collecting and supplementing data through continuous updates by using real-time panoramic images uploaded to the web.
In this study, the tooth-related disease judgment system identified 5 types of tooth-related diseases that are difficult to determine clinically (visually) by using an AI model, and this information was provided on the web to create a system that allows doctors and patients to make real-time judgments. The trained model labeled 5 dental caries through 10,000 panoramic images. Accuracy was compared using Fast R-CNN, ResNet, and inception models, which are good models for detection. Among these models, Fast R-CNN was finally used, which has the highest accuracy. Therefore, Fast R-CNN can be used to shorten the time required for the diagnosis and treatment of dental caries. In addition, by updating the captured panoramic images of patients on the web service developed in this study, the system can acquire new data and further increase the accuracy of diagnosing tooth-related diseases. Additionally, the patient can be aware of the tooth areas where he or she has received treatment, the treatment time, and the type of caries, so that he or she can adjust the schedule for the future dental visit, which will aid in continuous management of dental health. Thus, this study is meaningful as it collects learning data from cases embodied as actual services and implements a prototype-type service based on the collected data. In the future, it will be possible to develop a model for predicting overall oral diseases with panoramic images through additional learning of various dental diseases.
artificial intelligence
fast region-based convolutional network
residual neural network
None declared.