A deep learning approach to permanent tooth germ detection on pediatric panoramic radiographs

An accurate diagnosis can be characterized as the first and the most crucial step for performing successful dental treatments. Panoramic radiography has been a valuable tool that is widely used by dental practitioners to make correct diagnoses.1 Panoramic radiography makes it possible to visualize the entire maxillofacial region on a single image with minimal patient discomfort, fast and easy application, and a limited radiation dose.2 With these advantages, panoramic radiography does not adversely affect patient cooperation, which is particularly essential in pediatric dentistry.3 However, interpreting the detected conditions and lesions may be challenging because of superimpositions caused by the complexity of the 3-dimensional anatomic structures on the panoramic image.4 Computer-aided diagnosis systems have been developed to prevent clinicians from overlooking dental diseases and to increase the accuracy of radiological diagnoses.5 In recent years, computer-aided detection or diagnosis systems created by conventional methods have been replaced by artificial intelligence (AI)-based systems aiming to enhance radiological diagnoses.6

Deep learning is an AI-based approach employed in several clinical fields to make automated decisions. Convolutional neural networks (CNNs), a particular class of deep learning models, have gained tremendous attention in medical imaging.7 CNNs have the ability to “learn” how to make a diagnosis based on a sufficient amount of provided data, such as images, by mimicking the neurons of the human brain.8 Training a CNN from scratch, using off-the-shelf pre-trained CNN models, and performing unsupervised CNN pre-training with supervised fine-tuning are major techniques that have been employed for medical image classification.9 After validation of a trained CNN model with an independent dataset, performance metrics can be obtained.

CNNs have recently been applied in dental radiology for automatic segmentation of teeth,10, 11 detection of caries,12, 13 apical lesions,14 and periodontal bone loss.15 Tuzoff et al.16 employed the Faster R-CNN architecture for teeth detection and numbering on 1,352 adult panoramic radiographs and reported that the performance level of the CNN-based system was close to the level of radiologists. Although these studies indicated promising outcomes, this research area is not fully developed yet.16 The aforementioned studies all used CNN architectures to detect dental diseases or teeth on adults’ panoramic radiographs. However, there have been limited studies on employing AI-based systems for detecting dental diseases on panoramic radiographs of pediatric patients.17, 18, 19

Generally, CNN-based models are expected to contribute to making the correct diagnosis, thereby reducing the dentists’ workload and efforts.14 Particularly in pediatric dentistry, faster and more effective diagnoses help patients cooperate better with their dental practitioners and increase the success rate of dental treatment.20 The diagnosis of permanent tooth deficiency or excess directly affects the treatment choice. The early detection of such conditions allows less invasive restorative procedures and surgical removal of lesions, such as cysts and tumors, at early stages without damaging the surrounding healthy tissues.21 For this purpose, this study aimed to apply a CNN architecture for pediatric panoramic radiographs to help dental practitioners make effective diagnoses. To the best of the authors’ knowledge, this is the first study using a CNN model to detect permanent tooth germs on pediatric panoramic radiographs.

The trained model was tested with 523 panoramic radiographs that contained a total of 5,898 permanent tooth germ ground truth labels. The YOLOv4 model provided 6,043 detections, out of which 5,364 were TP detections and 679 were FP detections. The total number of FN detections was 534.

In the descriptive statistics, out of 6,043 detections, the minimum confidence score was calculated as 0.250 and the maximum confidence score was 0.999; the mean was 0.877±0.196. With 5,364 TP detections and 679 FP detections, the t-test showed statistical significance at P<0.05.

To evaluate the model’s success, a confusion matrix was used together with the following procedures and metrics: using 5,364 TP detections, 679 FP detections, and 534 FN detections in the above formulas, the precision, recall, and F1-score were computed to be 0.89, 0.91, and 0.90 respectively. Using the area under the precision-recall curve, the AP value was calculated as 94.16%. The AP, precision, and recall values indicated the success of the model for permanent tooth germ detection. It is significant when the F1-score yields a value above P=0.5. Since P>0.5 in this study, the F1-score was a strong indicator of the success and significance of this study (Fig. 3).

Fig. 3
Precision and recall curve. The area under this curve gives the average precision value.

• Click for larger image

Figure 4 provides sample permanent tooth germ detections from a test set image, together with the ground truth labels. All permanent tooth germs were correctly detected in this image, and the confidence scores of detection were very high.

Fig. 4
Visual results from the test set. A. Original pediatric panoramic radiograph. B. Permanent tooth germ detection using YOLOv4 (prediction scores are provided on top of each detection. The model predicts the ground truth annotations and bounding boxes with high accuracy).

• Click for larger image

The YOLOv4 inference time was 90 ms on average. Hence, the model can process about 11 images per second. Together with a 94.16% AP value and close-to-real-time inference speed, the YOLOv4 model has proven that very accurate results can be obtained at very fast speeds, when compared to other CNN-based object detection models. Moreover, the high precision and high recall values all indicate that the model presented in this study was statistically powerful.

Primary teeth and the permanent successors are interdependent components of the dental arch that closely interact with each other.27 Diagnosing developmentally absent or supernumerary permanent teeth is very crucial for determining the most appropriate dental procedure in pediatric dentistry. The choice of whether to perform endodontic and restorative rehabilitation for the involved tooth depends on the presence or absence of a permanent tooth germ.21 The development of a system that helps dental practitioners detect permanent tooth germs may enable more accurate diagnoses by reducing their workflow and human-induced errors. In this regard, the present study aimed to use an AI-based system to detect permanent tooth germs on panoramic radiographs.

The development of deep learning and neural methods accelerated the integration of AI into the field of medicine. In recent years, the use of deep learning methods in dentistry has gained attention, as in the medical field. Many studies have investigated deep learning methods for automated tooth segmentation and dental disease diagnosis on panoramic radiographs from adults.10, 11, 16 However, only 1 study has used a deep learning-based AI model for the automated tooth segmentation of pediatric panoramic dental radiographs.28 This is the first study to investigate the use of a CNN architecture to detect permanent tooth germs on pediatric panoramic images.

The benefits and limitations of the automatically detecting deep CNN algorithm YOLOv4 were experienced in the detection of permanent tooth germs on panoramic images in this research. YOLOv4 is a real-time detection system with some outstanding features, such as classifying targeted objects in a single forward pass.29 Moreover, YOLOv4 treats detection as a regression problem that does not need a complicated pipeline or semantic segmentation. YOLOv4 also learns the regions of interest and their background together and performs simple inference in a new image during the test process for predicting detections.30

The most important reason for choosing YOLO instead of other CNN models was that the algorithm is fast enough to detect objects at once and can run in real time. In addition, the most important feature that distinguishes YOLO from other algorithms is that it can accurately perform real-time object detection with its overall high mean AP.29 Yang et al.30 evaluated the performance of YOLO to detect and classify odontogenic cysts and tumors on 1,603 panoramic radiographs and revealed that YOLO provided at least similar diagnostic performance as experienced dentists on detection of maxillofacial lesions. Since this study evaluated a different dental issue and used panoramic images of adults, it is not possible to compare the findings directly.

Kuwada et al.6 reported that deep learning systems have potential for detecting maxillary impacted supernumerary teeth on panoramic radiographs, but for the mixed dentition, it would be more difficult because of the presence of unerupted permanent teeth. It was believed that learning the permanent tooth germs may be an important step toward solving this problem. Furthermore, applying a CNN to detect permanent tooth germs on panoramic radiographs may help diagnose impacted supernumerary teeth more accurately. The diagnostic performance of the deep-learning system for maxillary sinus lesions on panoramic radiographs was found to be sufficiently high.31, 32, 33 Similarly, a deep learning application provided high accuracy in the detection of impacted third molars and their relationship with anatomical structures.34, 35

Ahn et al.18 investigated the performance of a deep learning model to detect mesiodens on primary or mixed dentition panoramic radiographs and suggested that this method may help clinicians with insufficient clinical experience make more accurate and faster diagnoses. Ha et al.19 proposed a model based on YOLOv3 to detect mesiodens on panoramic radiographs of primary, mixed and permanent dentition groups. They compared the model performance between original panoramic radiographs and contrast-limited histogram equalization (CLAHE) and found that the accuracy and specificity of the original images was higher than that of the CLAHE images. This study reported that their model was successful for clinical practice to detect mesiodens on panoramic radiographs of all dentition types. However, no study has evaluated the performance of deep learning systems for detecting permanent tooth germs on panoramic radiographs. Therefore, it is not possible to make a direct comparison with the present study.

Although the number of studies on detecting dental diseases or tooth numbering with AI-supported systems on panoramic images of adult patients is rapidly increasing,6, 13, 14, 15, 31, 32, 33, 34 few studies have investigated AI-based systems with pediatric panoramic images.17, 18, 19, 28 This may be explained by the positioning artifacts that occur due to the superposition of anatomical structures and teeth on panoramic images of children. Since visualizing both jaws in children with severe maxilla-mandible discordance and positioning children correctly during imaging are often difficult, panoramic images do not provide detailed information.36 Caliskan et al.17 used CNN algorithms to detect and classify submerged molars on 74 panoramic images of 5- to 12-year-old children and reported that the system was extremely accurate. It may not be possible to make a full comparison since they investigated a different task and used a different algorithm.

Further research may be required to detect which germ is absent by using tooth numbering algorithms, thereby improving the performance of YOLO. The detection of missing tooth germ may help dental practitioners determine accurate treatment approaches. The number of annotated dental tissues or diseases can be increased as well. More advanced dental imaging methods may be used to enhance the results of YOLO.

In conclusion, deep learning systems are an encouraging approach and may have a crucial effect in terms of shortening the diagnosis time and effort in dental practice. Particularly for pediatric patients, shortening the time for diagnosis increases the cooperation of children and enables dental practitioners to perform more successful treatments. Thanks to YOLO’s fast real-time results, images were processed in less time than expected, and highly accurate results were obtained. To the best of the authors’ knowledge, this is the first study that used a CNN algorithm for detecting permanent tooth germ on pediatric panoramic images.