Detection of Periodontal Bone Loss Types and Furcation Defects from Panoramic Radiographs Using Deep Learning Algorithm: A Retrospective Study

doi:10.21203/rs.3.rs-3583126/v1

Download PDF

Research Article

Detection of Periodontal Bone Loss Types and Furcation Defects from Panoramic Radiographs Using Deep Learning Algorithm: A Retrospective Study

https://doi.org/10.21203/rs.3.rs-3583126/v1

This work is licensed under a CC BY 4.0 License

Journal Publication

published 31 Jan, 2024

Read the published version in BMC Oral Health →

You are reading this latest preprint version

Background

This retrospective study aimed to develop a deep learning algorithm for the interpretation of panoramic radiographs and to examine the performance of this algorithm in the detection of some periodontal problems such as horizontal alveolar bone loss, vertical bone defect, and furcation defect.

Methods

A total of 1121 panoramic radiographic images were used in this study. Total alveolar bone losses in the maxilla and mandibula (n = 2251), interdental bone losses (n = 25303), and furcation defects (n = 2815) were labeled using the segmentation method. In addition, interdental bone losses were divided into horizontal (n = 21839) and vertical (n = 3464) bone losses according to the defect types. A Convolutional Neural Network (CNN)-based artificial intelligence (AI) system was developed using U-Net architecture. The performance of the deep learning algorithm was statistically evaluated by the confusion matrix and ROC curve analysis.

Results

The system showed the highest diagnostic performance in the detection of total alveolar bone losses and the lowest in the detection of vertical bone defects. The sensitivity, precision, F1 score, accuracy, and AUC values were found as 1, 0.995, 0.997, 0.994, 0.951 for total alveolar bone loss; found as 0.947, 0.939, 0.943, 0.892, 0.910 for horizontal bone losses; found as 0.558, 0.846, 0.673, 0.506, 0.733 for vertical bone defects and found as 0.892, 0.933, 0.912, 0.837, 0.868 for furcation defects (respectively).

Conclusions

AI systems offer promising results in determining periodontal bone loss patterns and furcation defects from dental radiographs. This suggests that CNN algorithms can also be used to provide more detailed information such as automatic determination of periodontal disease severity and treatment planning in various dental radiographs.

Panoramic radiography

artificial intelligence

periodontitis

dentistry

Periodontitis is a chronic inflammatory disease that is characterized by damage to the supporting tissues of the teeth and can result in tooth loss if not controlled [1]. This common disease in society is associated with many systemic diseases in the body as well as dental problems such as chewing loss [2, 3]. In this respect, early diagnosis and treatment planning are very important in patients with periodontitis [2, 3].

Radiographic evaluations play an important role in the diagnosis of periodontitis in addition to clinical periodontal evaluations such as probing pocket depth, attachment loss, and gingival recession [4]. Intraoral radiographs such as periapical radiography and bitewing radiography are commonly used to determine periodontal status [4–6]. Panoramic radiographs, which is one of the extraoral dental radiographic techniques, have advantages such as working with low-dose radiation and providing quick and easy radiological imaging, and also helping in determining general dental problems and periodontal conditions [6–8]. In addition, these radiographs allow the observation of bone defect types (such as horizontal bone loss, vertical bone defects, and furcation defects) in cases of periodontitis [9].

In periodontitis cases, it is very important to determine the type of alveolar bone loss and bone morphology to make an appropriate treatment plan and achieve successful results [10, 11]. Horizontal bone loss is defined as the loss of the entire alveolar bone crest perpendicular to the long axis of the tooth, where the interdental bone is resorbed at the same level as the vestibule and lingual cortical layers [12–14]. Vertical bone defects are bone losses that occur obliquely and angularly in the interdental bone region [12–14]. In multi-rooted teeth, furcation defects occur when the alveolar bone between the tooth roots is also affected by periodontal disease. In furcation defects, both diagnosis and treatment planning become much more complex and difficult for dentists [15]. The correct estimation of bone types enables the correct treatment selection and facilitates the work of both patients and physicians. For example, flap surgery and/or resective surgical periodontal therapy may be preferred rather than regenerative treatment in horizontal bone losses, while regenerative treatment may be an alternative method in vertical bone defects [14]. In addition, the determination of defect types is important in terms of making the 2018 new classification complete and correct [16, 17].

Various studies about artificial intelligence (AI) have been carried out in dentistry, following the use of AI systems in the medical field for disease diagnosis and treatment planning [18, 19]. In these studies, it is seen that AI systems are used for the determination of many pathologies such as dental caries, apical lesions, tooth numbering, and root fractures on two-dimensional (2D) radiographs [20–23]. The main purpose of using AI in dentistry practice is to automatically detect pathologies, diseases, or anatomical structures and determine disease severity. Thus plans for preventing situations that are overlooked due to inexperience, intensity, and fatigue, diagnosing diseases early, and analyzing and recording patient data with digital systems [6, 19].

There are limited studies for evaluating alveolar bone losses and periodontal problems using AI systems [3, 6, 17, 24, 25–29]. However, as far as we know, there is no AI-based study aiming to determine the types of periodontal bone loss using the segmentation method. This study aimed to examine the performance of a convolutional neural network (CNN)-based AI system in the detection of periodontal problems such as horizontal bone losses, vertical bone defects, and furcation defects on panoramic radiographs by segmentation method.

One-thousand-one-hundred-twenty-one (n = 1121) anonymized panoramic radiographic images with periodontal bone destructions were used in this AI-based retrospective study. All of the images were obtained from the radiology archive of the Eskişehir Osmangazi University Faculty of Dentistry. Ethics board approval was obtained from the Non-interventional Clinical Research Ethics Committee of Eskişehir Osmangazi University before starting the study (decision number: 2019 − 227). The guidelines of the Declaration of Helsinki were followed throughout all phases of the study.

Radiographic Image Selection

All radiographic images were selected from images of individuals over the age of 18, without paying attention to age and gender differences. Even though images of individuals of a variety of ages and different genders were included in the dataset, all images were selected with the same image quality and resolution to increase the reliability of AI in diagnosis. All panoramic images used in the study were obtained with the same radiography device (Planmeca Promax 2D, Planmeca, Helsinki, Finland) with the following parameters: 68 kVp, 16 mA, and 13 s.

‘1. The images with dense artifacts, 2. The images with low quality due to patient positioning error or patient movement during acquisition, 3. The images of patients who received orthognathic treatment, 4. The images of patients with bone metabolism disorders, 5. The images of patients with unusual alveolar bone morphology, 6. The images of patients with cleft lip and palate, 7. The images of poor quality and blurred images in the alveolar bone region due to conditions such as dental crowding, 8. The images of patients with metal restorations that cause artifacts to complicate periodontal diagnosis’ were not included in this study.

Determination and labeling of periodontal problems

All images were uploaded into the CranioCatch labeling module (CranioCatch, Eskisehir, Turkey). All panoramic images were evaluated by 3 periodontists (SKB, MBY, NS) and an oral maxillofacial radiologist (İŞB), whose intra-observer and inter-observer compatibility were analyzed before. After the evaluations, total alveolar bone losses in the lower and upper jaws, interdental bone losses according to defect types, and furcation defects in the teeth were labeled using the segmentation method in this web-based labeling software system. (Fig. 1-a-c). Observed types of bone loss were classified according to the following criteria:

Total alveolar bone loss

All bone losses around the mandibular and maxillary teeth were labeled as a whole including all interdental spaces. Initially, a segmentation line was created from the cement-enamel boundary of all teeth. After that, the segmentation line was then combined to follow the line on the distal surfaces of the most distal teeth and the border of the bone crests in the relevant area (separately for the mandible and maxilla). In Fig. 1-a, an example image was presented to better understand the path followed when labeling this parameter. (Fig. 1-a).

Horizontal bone loss

It is a form of bone destruction in which the alveolar bone is observed more apically than it should be, and the imaginary line passing through the enamel-cement boundaries of the two adjacent teeth is parallel to the alveolar bone crest morphology [12–14].

In this study, in cases where the enamel-cementum border of two adjacent teeth and the imaginary line passing through the alveolar bone crest were parallel to each other, horizontal bone loss was decided and labels were performed. All horizontal bone loss in the interdental region had certain characteristics and the dentists who performed the labeling at the decision stage made their decisions in line with these criteria. In all cases, bone destruction was observed at almost the same rate and symmetrically in the interdental region between two adjacent teeth. When the labels were completed, they created a certain geometry, similar to geometric shapes such as squares, rectangles, and trapezoids. They certainly didn't have a triangle image.

The interdental spaces of all teeth were examined separately in the labeling stage. Areas of interdental bone losses with horizontal bone loss characteristics were labeled with the continuous segmentation technique, starting from the cement-enamel boundary of the two adjacent teeth, up to the alveolar bone crest margins (relevant interdental space only) (Fig. 1-b).

Vertical bone defect

The alveolar bone level, which is located more apically than the cement-enamel boundary of the two adjacent teeth, is observed. In this type of bone loss, the related area shows angular destruction characteristics in contrast to horizontal bone destruction [12–14].

At the diagnosis stage, an imaginary line was drawn between the cementum-enamel boundary of the two adjacent teeth and over the alveolar bone crest line. Areas with significant angulation between these two lines were labeled as vertical bone defects. These labels showed a specific triangular geometry characteristic rather than a square, rectangular, or trapezoidal appearance. The destruction characteristic was not symmetrical and radiopacity of the alveolar bone and lamina dura had disappeared in the label area. Very small bone angles due to changes in the tooth axis were not considered as vertical bone defects.

The interdental spaces of all teeth were examined separately in the labeling stage. Interdental bone loss areas with the aforementioned vertical bone defect features were labeled by continuous segmentation technique, starting from the cement-enamel boundary of two adjacent teeth to the alveolar bone crest margins (only the relevant interdental space) (Fig. 1-b).

Furcation defects

It is the bone destruction seen at the root junctions of multi-rooted teeth [15]. In most cases, the alveolar bone line was visible in the teeth region and furcation areas. Therefore, the related diagnosis was made in a standardized way by dentists without any difficulty in the decision phase. Furcation defects were labeled along the radiolucent image and traced to the defect border. (Fig. 1-b).

After labeling was completed, the data set was rechecked by the researchers. At this stage, instead of reviewing the boundaries of segmentation, the bone resorption form and geometry of the labeled region were quickly reviewed to avoid standardization deficiencies that may arise in horizontal-vertical bone loss decisions. In addition, consensus was reached in cases where dentists were conflicted about labeling some initial furcation defects. Also, the final evaluation made by 4 dentists together allowed the forgotten and overlooked labels to be noticed and completed before artificial intelligence training. Labels with no consensus on diagnosis were removed and not used in algorithm training and result metrics. In addition, in some radiographs, no labeling was made in areas that could not be diagnosed as horizontal or vertical bone loss due to partial superimposition or other specific reasons.

Observers agreed on the diagnosis of 1121 panoramic radiography images. These radiographic images included 2251 total alveolar bone loss, 25303 interdental bone loss (21839 horizontals, 3464 verticals), and 2815 furcation defects in total. This dataset was used to develop an AI model that automatically detects periodontal problems that have been labeled.

Intra-observer and inter-observer agreements

For intra-observer evaluation, the researchers labeled alveolar bone destruction and furcation defects with the segmentation method on 10 radiographs with alveolar bone losses. Each researcher repeated this procedure 1 week later. A computer program (Python version 3.6.1, Python Software Foundation) and NumPy library were used and the splicing of the segmented fields was done and Intersection over Union (IoU) values were calculated. It was found that each investigator showed high consistency in the diagnosis and segmentation of these parameters at different times (IoU > 0.91). In addition, kappa values were calculated to evaluate the consistency of the researchers’ decision for horizontal and vertical bone defects in the interdental area. Kappa values showed excellent agreement when the two times were compared in the decision of bone loss characteristics in each investigator (K = 0.81–0.99). Also, the labeling areas of different researchers were examined for evaluation of inter-observer agreement using this data set. The IoU values obtained were found to be between 0.8126 and 0.8737. While the results were 0.85 and above for horizontal bone loss, total alveolar bone loss, and furcation defect, they were lower for vertical bone defect (> 0.81).

Creating a CNN-based AI model

Initially, all images were resized as 1024x512 pixels. Panoramic radiographic images to be used in the data sets created for developing the furcation defect model were cropped to 4 (left mandible, right mandible, left maxilla, and right maxilla). Since furcation defects were only seen in multi-rooted molars, it was aimed to focus more on the relevant areas in this way. The radiography images to be used in other parameter models were not cropped but used as a whole. Images that do not have the related parameter label were excluded from the main data sets of that parameter (Fig. 2, 3). In line with these criteria, 1121 panoramic radiograph images for total alveolar bone destruction, 1120 panoramic radiograph images for horizontal bone destruction, 828 panoramic radiograph images for vertical bone defects, and 1941 cropped panoramic radiograph images for furcation defect were included in the main datasets (Table 1, Fig. 3).

Table 1

Numerical data on the development of deep learning algorithms
Model Name	Number of Training Images	Number of Education Labels	Number of Test Images	Number of Test Labels	Verification Image Count	Number of Validation Tags	Epoch	Learning Rate	Model
Alveolar Bone Loss	935	1878	93	187	93	186	800	0.0001	Unet
Horizontal Bone Loss	934	18232	93	1821	93	1786	800	0.0001	Unet
Vertical Defect	690	2869	69	308	69	287	800	0.00001	Unet
Furcation Defect	1619	2358	161	230	161	227	800	0.00001	Unet

Main datasets were created by combining all panoramic radiography images containing the relevant parameter labels for each parameter. Then these main datasets were split as a training set (80%), validation set (10%), and testing set (10%) randomly for the development of AI systems (Table 1). The labels in the training set were used to create the algorithm of the AI model. This model was tested with the data in the validation set for adjusting the algorithm and questioning the need for further training. The final version of the algorithm, which was developed using the training set and successful trials were carried out with the labels in the validation set, was used in the detection of periodontal bone destruction and pattern on the radiographs in the testing set. The diagnostic results of the system in the testing set were compared with the dentist labels in this dataset by a computer command, and the success metrics of the system were presented. In short, while determining the final success of the model, the dentist diagnoses in the test dataset were accepted as the gold standard and these labels were used as the ground truth dataset. (Fig. 3). These steps were performed separately for each parameter.

The U-Net architecture was used to perform deep learning (Table 1) [23, 30]. U-Net is a type of CNN that can perform semantic segmentation assignments on various images such as radiography images. U-Net contains four block levels with 32, 64, 128, and 256 convolution filters in each block. The working mechanism of this architecture is to convert images to vectors for pixel classifications and then convert these vectors back to images for segmentation. The encoder and decoder paths are used in order and perform model training. There is a maximum pool layer in the paths of encoders and up-convolution layers in the paths of decoders. For the current study, the U-Net architecture steps were followed and the working mechanism was presented in Fig. 4 in detail [23, 30]. In addition, the PyTorch library (v. 3.6.1; Python Software Foundation, Wilmington, DE, USA) and Adaptive Moment Estimation (ADAM) optimizer (for updating the learning rate for each parameter) were used in the model development stages. All training was made with a computer equipped with 16 GB RAM and NVIDIA GeForce GTX 1660 TI graphics card.

Statistical Analysis

The statistical evaluation phase of the study was carried out with the confusion matrix method and ROC analysis. A script was created to perform these evaluations automatically with the help of a computer. A computer program (Python version 3.6.1, Python Software Foundation) and OpenCV and NumPy libraries were used to develop this script. The script used at this stage accepted the labels made by experienced dentists in the testing set as the gold standard. It automatically presented a success metric by comparing the diagnostics of the algorithm in the relevant dataset with the dentists’ diagnoses. While the script was running, it presented results using some calculations below.

Sensitivity, precision, and F1 scores were calculated using the confusion matrix method [23]. Initially, three different calculations were for the confusion matrix method: numbers of true positives (TP, there was a periodontal problem and it was segmented correctly), numbers of false positives (FP, there wasn’t a periodontal problem but it was detected wrongly), and numbers of false negatives (FN, there was a periodontal problem but it wasn’t detected). Then, sensitivity, precision, and F1 scores were determined with the following calculations:

Sensitivity (recall): TP/ (TP + FN)

Precision: TP/ (TP + FP)

F1 score: 2TP/ (2TP + FP + FN)

In addition to the confusion matrix method, receiver operating characteristic (ROC) curve analysis and precision-recall evaluation were also performed to provide more detailed data. The area under the ROC curve (AUC) values are calculated by this analysis. ROC is a probability curve and the area under it, AUC, represents the degree or measure of separability. As the area under the curve increases, the discrimination performance between classes increases. It is known that as the success of the system increases, the AUC grows and approaches the value of 1. In line with this information, interpretations were made and the results were presented. In addition, the labels and system results of the radiographs in the test data set in which the evaluations were carried out were reviewed by experienced dentists for each parameter.

The results were calculated with the confusion matrix method, which demonstrates the success of the developed AI model, were presented in Table 2. When the diagnostic performance metrics were evaluated, sensitivity, precision, F1 score, and accuracy values, respectively, were 1, 0.995, 0.997, 0. 994 for total alveolar bone loss; 0.947, 0.939, 0.943, 0.892 for horizontal bone losses; 0.558, 0.846, 0.673, 0.506 for vertical bone defects and 0.892, 0.933, 0.912, 0.837 for furcation defects at 50% IoU. AUC values were 0.951 for total alveolar bone loss, 0.910 for horizontal bone losses, 0.733 for vertical bone defects, and 0.868 for furcation defects (Table 2). In Fig. 5, The input and output images of the radiographs of some patients in the AI system were shown as an example.

Table 2

The results obtained with the confusion matrix method and ROC analysis
Model Name	Found Correct For IoU Thresold: 50%	Found Wrong for IoU Thresold: 50%	Not Found for IoU Thresold: 50%	Sensitivity	Precision	F1 Score	Acurracy	AUC
Alveolar Bone Loss	185	1	0	1	0.995	0.997	0.994	0.951
Horizontal Bone Loss	108	7	6	0.947	0.939	0.943	0.892	0.910
Vertical Defect	115	21	91	0.558	0.846	0.673	0.506	0.733
Furcation Defect	181	13	22	0.892	0.933	0.912	0.837	0.868

Also, ROC curve analysis and precision-recall results for all periodontal parameters were presented in Fig. 6.

In addition, the input and output images of each case in the testing set were reviewed by the dentists who labeled them, and it was analyzed in which situations the system made more mistakes. When the output images of total alveolar bone loss were examined, it was observed that the success of the system was quite good, and there were few errors in the detection of only some deep angular bone loss areas, and furcation areas. When the output images of horizontal bone loss were examined, it was analyzed that the system could not diagnose in some radiographs in large interdental regions caused by tooth deficiency. It was observed that some vertical bone defects were determined as horizontal bone loss in regions with partial superimposition of anatomical structures. The most common error made by the system in detecting vertical bone defects was to perceive very minimal bone angulations as a vertical bone defect. The most common mistake made by the system when identifying furcation defects was perceiving the relevant area as furcation in teeth with close roots and labeling these areas as furcation defects. Another mistake was that he overlooked and failed to detect some of the initial level furcation defects.

In recent years, the use of AI systems in medicine and dentistry in image interpretation has attracted great interest and CNN systems change these areas rapidly [23]. However, the number of studies conducted in the field of periodontology is still limited. The current study aims to automatically evaluate panoramic radiographs by an AI system and to determine the success of the related system in detecting periodontal disease findings on radiographs. The results of this study show that AI systems can be a decision-support mechanism for physicians/specialists in the diagnosis of periodontal disease, one of the most common diseases in the world.

Different architectures can be used in AI-based studies. U-Net architecture, a CNN-based AI architecture, is one of the image segmentation techniques that enable the evaluation of images in the medical field with AI systems and provides a more precise and successful evaluation with fewer training sets [31, 32]. Therefore, in the current study, image processing was performed using the U-Net architecture.

In the literature, there were many studies in which AI algorithms were used in the processing and interpretation of patient data, which played a role in diagnosis and treatment planning, such as dental radiographic images [19, 33–35], intraoral photographs [36–39] and pathology images [40]. In addition, these systems were also used in situations such as pre-examination evaluation and risk estimation [41]. Also, it was seen that some of the studies aiming to determine periodontal disease from 2D dental radiographs using AI systems were performed on periapical radiographs [20, 27, 42] and some on panoramic radiographs [3, 6, 24, 28, 29].

In one of the studies in the literature, Lee et al. (2018) used 1740 periapical radiographs and evaluated the prediction success of the CNN algorithm they developed. In the results of this study, they reported that the accuracy of determining premolar teeth with periodontal damage was 81% and 76.7% for molars [26]. Similar to this study, Lee et al. (2021) performed another study on 693 periapical radiographs [17]. In this study, it was aimed to find the radiographic bone loss by grouping it according to the severity of destruction (stages I-II-III) [17] High success rates were found in the detection of periodontal disease in the results of these two studies.

On the other hand, Khan et al. (2021) tried to find dental problems such as caries, alveolar bone loss, and inter radicular bone loss on 206 periapical images using different AI architectures (U-Net, X Net, and Seg Net) in their study [27]. Although this study was similar to our study in terms of using the segmentation method in the detection of periodontal bone loss, no evaluation was made in terms of bone destruction characteristics (horizontal, vertical) [27]. Another difference was Khan et al. (2021) performed their studies on periapical radiographs and with fewer data [27]. Although the evaluations were made on different types of radiography images, it could be said that the success rates of our study were higher. We think that the reason for the higher success rates of the AI system developed in our study may be due to the use of a larger data set. Because as the number of data used for training in AI studies increases, the success rates of the model also increase [43].

In another study, Kurt Bayrakdar et al. (2021) used Google Net Inception v3 architecture on a large data set (2276 panoramic radiographs). They evaluated the success of the AI systems they developed in the determination of radiographs of patients with periodontitis/alveolar bone destruction [6]. In this study, 1137 radiographs of patients with bone destruction and panoramic radiographs of 1139 periodontal healthy individuals were used for the development of the AI model. F1 score and accuracy were found as high as 91% in this study [6]. However, no labeling was made, only the classification method was used to train the radiographs in the form of patient/healthy and to find them by the system. In this respect, they are quite different from our study in terms of this study planning. When the literature is reviewed, it is seen that there are different techniques in computer vision such as classification, object detection, and image segmentation [44, 45]. Classification is useful at the yes-no level of deciding whether an image contains an object (pathology, anatomical structure…etc.) or not. Object detection deals with distinguishing between objects (pathology, anatomical structure…etc.) on an image. It does this by simply surrounding the object with a box, without defining its boundaries. In segmentation, images are divided into pixel groupings and it is possible to define an object (pathology, anatomical structure…etc.) from its exact boundaries [44, 45]. The image segmentation technique plays a significant role in identifying a disorder, treatment planning, and routine follow-ups in the medical field because it has the advantage of presenting disorders/pathologies by identifying all their borders [46]. The reason why it is not preferred in some studies is that it takes time and is a tedious process [47]. Based on this information, it can be said that the segmentation technique we used in our study is the most advantageous method and provides the physician with more detailed information for diagnosis and treatment planning.

There were also studies in the literature comparing the diagnosis of physicians with different experiences and AI model predictions. The results obtained in this way are undoubtedly more interpretable and reveal the success of the system more clearly. For example, Krois et al. (2019) compared the evaluation of 6 dentists and the results of AI in a panoramic radiography study with the CNN technique. They reported the accuracy, specificity, and sensitivity rates for AI-system as 81% in the results of their study [24]. Since this study deals with the evaluation of many physicians, it was a more comprehensive and superior study in terms of planning. However, in the present study, periodontal bone losses were not classified as horizontal or vertical. Unlike this study, one of the main purposes of our study was to guide the physician in treatment planning by enabling the system to find defects in the form of vertical/horizontal/furcation.

In another similarly planned study, Kim et al. (2020) compared the evaluations of 5 different clinicians with the performance of AI in their study for the determination of bone resorption sites using 12179 panoramic radiography [28]. They reported that while the average F1 scores of the clinicians’ results in determining bone destruction were 69%, AI showed higher success and the F1 score was 75%. This study reveals the success of AI systems in radiography interpretation and shows promise for the future use of these systems [28]. Also, Chang et al. (2020) tried to perform the determination of bone loss from panoramic radiographs without any evaluation of bone destruction angulation and defect type [3]. In this study, the staging was tried to be made according to the 2017 periodontitis classification by calculating the amount of bone destruction and destruction [3]. Bone destruction geometry is important when staging in the new classification. In fact, with this aspect, evaluations such as vertical-horizontal-furcation defects, as in our study, can provide more accurate and systematic results. This should be taken into account in future studies.

Finally, Jiang et al. (2022) tried to detect periodontal bone destruction in their study using the CNN model using 640 panoramic radiographs. This study is the most similar to our study in the literature [29] because Jiang et al (2022) also determined the defect types of periodontal disease in the form of vertical-horizontal-furcation defects in their study [29]. Although it was similar to our study in this respect, the object detection method was used in this study for labeling. The segmentation method we use is a much more advantageous method since it determines the defective area with its borders. Because it provides more detailed information to the physician to determine the severity of the disease and to plan the treatment in the next process. In addition, in this method, the defect area is processed like a detailed map and provides more advanced diagnostic support visually. On the other hand, when two studies were compared in detail, it was seen that they used multiple observers, and two different AI architectures (U-Net and Yolo-v4), and included the severity of the disease in their assessment. In these respects, it could be said that their study was superior and more detailed than ours.

One of the limitations of the study was that the decision of multiple observers was not separately compared with the predictions of AI. Another limitation was that no measurements were made to determine disease severity in the study and it was aimed to detect only bone loss areas. More extensive research could have been done using calibrated panoramic radiographs, based on direct measurement, or to determine the percentage of root affected. However, It should be noted that in all 2D radiographic imaging techniques, the evaluation of bone craters, lamina dura, and periodontal bone level is limited to the projection geometry and superposition of adjacent anatomical structures [48]. In other words, these radiographs do not provide clear and reliable information for measurement and treatment planning. For this reason, performing AI-based studies with three-dimensional (3D) radiographic imaging techniques such as Cone-beam computed tomography systems (CBCT) will prevent this limitation. Undoubtedly, as the number of studies in this field increases, AI algorithms with stronger decision-support capability and providing much more detailed information will be developed. Despite the limitations, our study is promising for such studies to be carried out in the future. In addition, the performance of the AI model in determining vertical bone defects in our study was weaker than other periodontal parameters. It is a known fact that vertical bone defects are less common than horizontal bone loss, and therefore, the number of labels for vertical bone defects was more limited in our study [14]. We think that the number of labels used in the detection of vertical bone defects with AI in our study was therefore less, and this situation caused the performance of the model to be found to be weaker. Because the most important thing in the success of AI studies is to work with large data sets. This limitation could be overcome, albeit to a limited extent, by using different AI architectures and making different technical plans. For example, in our study, cross-validation techniques could be used during model development and multi-class training could be done. On the other hand, the more important reason for the low success rates for the vertical bone defect parameter also may be the incomplete understanding of the outline of the alveolar bone crest in 1, 2, and 3-walled defects. We think that interobserver agreements were also lower for this parameter due to this limitation, which may be due to the limited view of panoramic radiography. This diagnostic difficulty can be eliminated by using some radiography techniques that provide more detailed images in future studies.

When the literature was examined, it was seen that the use of AI systems in dental radiographs for periodontal status determination has very successful and promising results. To the best of our knowledge, the present study is the first study aimed at detecting bone resorption and its types using deep learning algorithms and segmentation methods. These destruction types play an important role in periodontal treatment planning and the prognosis of teeth. With this aspect, these systems are candidates to be a decision-support mechanism for physicians in radiographic interpretation. The development of these systems with more data sets will increase success rates. There is a need for more comprehensive studies on 2D and 3D radiographs in this regard.

CNN

Convolutional Neural Network

Artificial intelligence

Two-dimensional

IoU

Intersection over Union

ADAM

Adaptive Moment Estimation

Numbers of true positives

numbers of false positives

numbers of false negatives

ROC

Receiver operating characteristic

AUC

area under the ROC curve

Three-dimensional

CBCT

Cone-beam computed tomography systems

Ethics approval and consent to participate: The research protocol was approved by the Eskisehir Osmangazi University Ethics Committee (decision number: 2019-227). All participants gave informed consent. We confirmed that the study was carried out in accordance with the relevant guidelines and regulations of the Helsinki Declaration.

Consent for publication: Additional informed consent was obtained from all individual participants included in the study.

Availability of data and materials: The datasets generated and/or analyzed during the current study are not publicly available due to the conditions specified when receiving approval from the ethics committee but are available from the corresponding author on reasonable request.

Competing interests: The authors declare that there is no conflict of interest.

Funding: This research was supported by the Ondokuz Mayis Research Fund (PYO.DIŞ.1904.18.006).

Authors' contributions: All authors have made substantial contributions to the conception and design of the study. SKB, ISB, MBY, and NS have been involved in data collection and ÖÇ has been involved in data analysis. SKB, ISB, MBY, NS, OK, BCUS, and RJ have been involved in the data interpretation and drafting of the manuscript. SKB, ISB, BK, and KO have been involved in revising it critically and have given final approval for the version to be published. * The all participants in this study gave informed consent (both oral and written).

Acknowledgments: This study has been supported by Eskişehir Osmangazi University Scientific Research Projects Coordination Unit under project number 202045E06.

Authors information:

Sevda Kurt-Bayrakdar. Eskisehir Osmangazi University, Faculty of Dentistry, Department of Periodontology, Eskisehir, Turkey. Research Scholar in Division of Oral and Maxillofacial Radiology, Department of Care Planning and Restorative Sciences, University of Mississippi Medical Center School of Dentistry, Jackson, MS, USA.

İbrahim Şevki Bayrakdar. Eskisehir Osmangazi University, Faculty of Dentistry, Department of Oral and Maxillofacial Radiology, Eskisehir, Turkey. Research Scholar in Division of Oral and Maxillofacial Radiology, Department of Care Planning and Restorative Sciences, University of Mississippi Medical Center School of Dentistry, Jackson, MS, USA.

Muhammed Burak Yavuz. Eskisehir Osmangazi University, Faculty of Dentistry, Department of Periodontology, Eskisehir, Turkey.

Nichal Sali. Eskisehir Osmangazi University, Faculty of Dentistry, Department of Periodontology, Eskisehir, Turkey.

Özer Çelik. Eskisehir Osmangazi University, Faculty of Science, Department of Mathematics and Computer Science, Eskisehir, Turkey.

Oğuz Köse. Recep Tayyip Erdogan University, Faculty of Dentistry, Department of Periodontology, Rize, Turkey.

Bilge Cansu Uzun Saylan. Dokuz Eylül University, Faculty of Dentistry, Department of Periodontology, İzmir, Turkey.

Batuhan Kuleli. Eskisehir Osmangazi University, Faculty of Dentistry, Department of Orthodontics, Eskisehir, Turkey.

Rohan Jagtap. Division of Oral and Maxillofacial Radiology, Department of Care Planning and Restorative Sciences, University of Mississippi Medical Center School of Dentistry, Jackson, MS, USA.

Kaan Orhan.Ankara University, Faculty of Dentistry, Department of Oral and Maxillofacial Radiology, Ankara, Turkey.

Dentino A, Lee S, Mailhot J, Hefti AF. Principles of periodontology. Periodontol 2000 2013;61:16-53.
Bourgeois D, Inquimbert C, Ottolenghi L, Carrouel F. Periodontal Pathogens as Risk Factors of Cardiovascular Diseases, Diabetes, Rheumatoid Arthritis, Cancer, and Chronic Obstructive Pulmonary Disease-Is There Cause for Consideration? Microorganisms 2019;7.
Chang H-J, Lee S-J, Yong T-H, Shin N-Y, Jang B-G, Kim J-E, et al. Deep learning hybrid method to automatically diagnose periodontal bone loss and stage periodontitis. Scientific reports 2020;10:1-8.
Mol A. Imaging methods in periodontology. Periodontol 2000 2004;34:34-48.
Scarfe WC, Azevedo B, Pinheiro LR, Priaminiarti M, Sales MAO. The emerging role of maxillofacial radiology in the diagnosis and management of patients with complex periodontitis. Periodontol 2000 2017;74:116-39.
Kurt Bayrakdar S, Özer Ç, Bayrakdar IS, Orhan K, Bilgir E, Odabaş A, et al. Success of artificial intelligence system in determining alveolar bone loss from dental panoramic radiography images. Cumhuriyet Dental Journal 2020;23:318-24.
Rushton VE, Horner K. The use of panoramic radiology in dental practice. J Dent 1996;24:185-201.
Clerehugh V, Tugnait A. Diagnosis and management of periodontal diseases in children and adolescents. Periodontol 2000 2001;26:146-68.
Ivanauskaite D, Lindh C, Rangne K, Rohlin M. Comparison between Scanora panoramic radiography and bitewing radiography in the assessment of marginal bone tissue. Stomatologija. 2006;8(1):9-15. PMID: 16687909.
de Faria Vasconcelos K, Evangelista KM, Rodrigues CD, Estrela C, de Sousa TO, Silva MA. Detection of periodontal bone loss using cone beam CT and intraoral radiography. Dentomaxillofac Radiol 2012;41:64-9.
Haas LF, Zimmermann GS, De Luca Canto G, Flores-Mir C, Corrêa M. Precision of cone beam CT to assess periodontal bone defects: a systematic review and meta-analysis. Dentomaxillofac Radiol 2018;47:20170084.
Persson RE, Hollender LG, Laurell L, Persson GR. Horizontal alveolar bone loss and vertical bone defects in an adult patient population. J Periodontol 1998;69:348-56.
Gomes-Filho IS, Sarmento VA, de Castro MS, da Costa NP, da Cruz SS, Trindade SC, et al. Radiographic features of periodontal bone defects: evaluation of digitized images. Dentomaxillofac Radiol 2007;36:256-62.
Jayakumar A, Rohini S, Naveen A, Haritha A, Reddy K. Horizontal alveolar bone loss: A periodontal orphan. J Indian Soc Periodontol 2010;14:181-5.
Braun X, Ritter L, Jervøe-Storm PM, Frentzen M. Diagnostic accuracy of CBCT for periodontal lesions. Clin Oral Investig 2014;18:1229-36.
Papapanou PN, Sanz M, Buduneli N, Dietrich T, Feres M, Fine DH, et al. Periodontitis: Consensus report of workgroup 2 of the 2017 World Workshop on the Classification of Periodontal and Peri-Implant Diseases and Conditions. J Periodontol 2018;89 Suppl 1:S173-s82.
Lee CT, Kabir T, Nelson J, Sheng S, Meng HW, Van Dyke TE, et al. Use of the deep learning approach to measure alveolar bone level. J Clin Periodontol 2022;49:260-9.
Sen D, Chakrabarti R, Chatterjee S, Grewal DS, Manrai K. Artificial intelligence and the radiologist: the future in the Armed Forces Medical Services. BMJ Mil Health 2020;166:254-6.
Kurt Bayrakdar S, Orhan K, Bayrakdar IS, Bilgir E, Ezhov M, Gusarev M, et al. A deep learning approach for dental implant planning in cone-beam computed tomography images. BMC Med Imaging 2021;21:86.
Lee JH, Kim DH, Jeong SN, Choi SH. Detection and diagnosis of dental caries using a deep learning-based convolutional neural network algorithm. J Dent 2018;77:106-11.
Fukuda M, Inamoto K, Shibata N, Ariji Y, Yanashita Y, Kutsuna S, et al. Evaluation of an artificial intelligence system for detecting vertical root fracture on panoramic radiography. Oral Radiol 2020;36:337-43.
Bilgir E, Bayrakdar İ, Çelik Ö, Orhan K, Akkoca F, Sağlam H, et al. An artifıcial ıntelligence approach to automatic tooth detection and numbering in panoramic radiographs. BMC Med Imaging 2021;21:124.
Bayrakdar IS, Orhan K, Çelik Ö, Bilgir E, Sağlam H, Kaplan FA, et al. A U-Net Approach to Apical Lesion Segmentation on Panoramic Radiographs. Biomed Res Int 2022;2022:7035367.
Krois J, Ekert T, Meinhold L, Golla T, Kharbot B, Wittemeier A, et al. Deep Learning for the Radiographic Detection of Periodontal Bone Loss. Sci Rep 2019;9:8495.
Thanathornwong B, Suebnukarn S. Automatic detection of periodontal compromised teeth in digital panoramic radiographs using faster regional convolutional neural networks. Imaging Sci Dent 2020;50:169-74.
Lee JH, Kim DH, Jeong SN, Choi SH. Diagnosis and prediction of periodontally compromised teeth using a deep learning-based convolutional neural network algorithm. J Periodontal Implant Sci 2018;48:114-23.
Khan HA, Haider MA, Ansari HA, Ishaq H, Kiyani A, Sohail K, et al. Automated feature detection in dental periapical radiographs by using deep learning. Oral Surg Oral Med Oral Pathol Oral Radiol 2021;131:711-20.
Kim J, Lee HS, Song IS, Jung KH. DeNTNet: Deep Neural Transfer Network for the detection of periodontal bone loss using panoramic dental radiographs. Sci Rep 2019;9:17615.
Jiang L, Chen D, Cao Z, Wu F, Zhu H, Zhu F. A two-stage deep learning architecture for radiographic staging of periodontal bone loss. BMC Oral Health 2022;22:106.
Ronneberger O, Fischer P, Brox T. U-net: Convolutional networks for biomedical image segmentation. International Conference on Medical image computing and computer-assisted intervention: Springer; 2015. p. 234-41.
Zhou Z, Siddiquee MMR, Tajbakhsh N, Liang J. UNet++: A Nested U-Net Architecture for Medical Image Segmentation. Deep Learn Med Image Anal Multimodal Learn Clin Decis Support (2018) 2018;11045:3-11.
Du G, Cao X, Liang J, Chen X, Zhan Y. Medical image segmentation based on u-net: A review. Journal of Imaging Science and Technology 2020;64:1-12.
Lee S, Kim D, Jeong HG. Detecting 17 fine-grained dental anomalies from panoramic dental radiography using artificial intelligence. Sci Rep 2022;12:5172.
Kılıc MC, Bayrakdar IS, Çelik Ö, Bilgir E, Orhan K, Aydın OB, et al. Artificial intelligence system for automatic deciduous tooth detection and numbering in panoramic radiographs. Dentomaxillofac Radiol 2021;50:20200172.
Pauwels R, Brasil DM, Yamasaki MC, Jacobs R, Bosmans H, Freitas DQ, et al. Artificial intelligence for detection of periapical lesions on intraoral radiographs: Comparison between convolutional neural networks and human observers. Oral Surg Oral Med Oral Pathol Oral Radiol 2021;131:610-6.
Zhang X, Liang Y, Li W, Liu C, Gu D, Sun W, et al. Development and evaluation of deep learning for screening dental caries from oral photographs. Oral Dis 2022;28:173-81.
Takahashi T, Nozaki K, Gonda T, Mameno T, Ikebe K. Deep learning-based detection of dental prostheses and restorations. Sci Rep 2021;11:1960.
Fu Q, Chen Y, Li Z, Jing Q, Hu C, Liu H, et al. A deep learning algorithm for detection of oral cavity squamous cell carcinoma from photographic images: A retrospective study. EClinicalMedicine 2020;27:100558.
Li W, Liang Y, Zhang X, Liu C, He L, Miao L, et al. A deep learning approach to automatic gingivitis screening based on classification and localization in RGB photos. Sci Rep 2021;11:16831.
Zehra T, Shaikh A, Shams M. Dawn of artificial intelligence - Enable digital Pathology in Pakistan: A paradigm shift. J Pak Med Assoc 2021;71:2683-4.
Alhazmi A, Alhazmi Y, Makrami A, Masmali A, Salawi N, Masmali K, et al. Application of artificial intelligence and machine learning for prediction of oral cancer risk. J Oral Pathol Med 2021;50:444-50.
Alotaibi G, Awawdeh M, Farook FF, Aljohani M, Aldhafiri RM, Aldhoayan M. Artificial intelligence (AI) diagnostic tools: utilizing a convolutional neural network (CNN) to assess periodontal bone level radiographically-a retrospective study. BMC Oral Health 2022;22:399.
Surya L. An exploratory study of AI and Big Data, and it's future in the United States. International Journal of Creative Research Thoughts (IJCRT), ISSN 2015;2320-882.
Khan AI, Al-Habsi S. Machine learning in computer vision. Procedia Computer Science, 2020;167, 1444-1451.
Wu H, Liu Q, Liu X. A review on deep learning approaches to image classification and object segmentation. Computers, Materials & Continua, 2019;60(2).
Dhruv B, Mittal N, Modi M. Artificial intelligence optimized image segmentation techniques for renal cyst detection. J Med Eng Technol. 2022;46(5):415-423.
Leite AF, Gerven AV, Willems H, Beznik T, Lahoud P, Gaêta-Araujo H, Vranckx M, Jacobs R. Artificial intelligence-driven novel tool for tooth detection and segmentation on panoramic radiographs. Clin Oral Investig. 2021;25(4):2257-2267
Acar B, Kamburoğlu K. Use of cone beam computed tomography in periodontology. World J Radiol. 2014; 28:6(5):139-47.

No competing interests reported.

Download PDF

Journal Publication

published 31 Jan, 2024

Read the published version in BMC Oral Health →

Editorial decision: Revision requested
17 Dec, 2023
Reviews received at journal
02 Dec, 2023
Reviewers agreed at journal
19 Nov, 2023
Reviewers agreed at journal
18 Nov, 2023
Reviewers agreed at journal
18 Nov, 2023
Reviewers invited by journal
18 Nov, 2023
Editor assigned by journal
14 Nov, 2023
Editor invited by journal
09 Nov, 2023
Submission checks completed at journal
09 Nov, 2023
First submitted to journal
09 Nov, 2023

You are reading this latest preprint version

Detection of Periodontal Bone Loss Types and Furcation Defects from Panoramic Radiographs Using Deep Learning Algorithm: A Retrospective Study

Status:

Journal Publication

Version 1

Abstract

Background

Methods

Results

Conclusions

Figures

Background

Methods

Radiographic Image Selection

Determination and labeling of periodontal problems

Intra-observer and inter-observer agreements

Creating a CNN-based AI model

Statistical Analysis

Results

Discussion

Conclusions

Abbreviations

Declarations

References

Additional Declarations

Status:

Journal Publication

Version 1