Open Access
Add to BibSonomy
Abstract
We propose a method to automatically segment the periodontal structures of the tooth enamel and the alveolar bone using convolutional neural network (CNN) and to measure quantitatively and automatically the alveolar bone level (ABL) by detecting the cemento-enamel junction and the alveolar bone crest in optical coherence tomography (OCT) images. The tooth enamel and the alveolar bone regions were automatically segmented using U-Net, Dense-UNet, and U2-Net, and the ABL was quantitatively measured as the distance between the cemento-enamel junction and the alveolar bone crest using image processing. The mean distance difference (MDD) measured by our suggested method ranged from 0.19 to 0.22 mm for the alveolar bone crest (ABC) and from 0.18 to 0.32 mm for the cemento-enamel junction (CEJ). All CNN models showed the mean absolute error (MAE) of less than 0.25 mm in the x and y coordinates and greater than 90% successful detection rate (SDR) at 0.5 mm for both the ABC and the CEJ. The CNN models showed high segmentation accuracies in the tooth enamel and the alveolar bone regions, and the ABL measurements at the incisors by detected results from CNN predictions demonstrated high correlation and reliability with the ground truth in OCT images.
© 2022 Optica Publishing Group under the terms of the Optica Open Access Publishing Agreement
1. Introduction
Alveolar bone is a structure that surrounds and supports teeth. Periodontitis refers to a condition that leads to destruction of the alveolar bone and is irreversible, unlike gingivitis, where inflammation is confined to the gingiva and is reversible [1]. Therefore, accurate evaluation of the alveolar bone condition and early diagnosis of the alveolar bone loss (ABL) are very important. Currently, panoramic or periapical radiographic images are mainly used to diagnose periodontal conditions. However, radiographic magnification results in limitations in accurate measurement of bone loss and underestimation of bone loss [2]. Exact evaluation of the superimposed buccal or lingual surfaces also suffers as panoramic or periapical radiographic images are two-dimensional [3]. Although three-dimensional cone beam computed tomography (CBCT) could show the buccal or lingual surfaces in greater detail, it is difficult to use as a routine diagnostic tool due to its high-dose radiation [4].
Optical coherence tomography (OCT) is a promising diagnostic technique that can provide a real-time cross-sectional image of biological tissue [5]. It utilizes a near-infrared laser and has advantages of being non-invasive, non-destructive, and irradiation-free. It has been commonly used in the field of ophthalmology, and its applicability has been studied in various areas of medicine and dentistry [6]. Previous studies on periodontal application of OCT reported that calculus [7], periodontal pockets [5,8,9], and peri-implant bone defects [10] could be observed and quantitatively measured in OCT images. In addition, it was also reported that the detailed human periodontal profile, including epithelium, connective tissue, and alveolar bone, could be identified in OCT images [6]. Recently, a study was also performed on measuring the alveolar bone level in porcine jaws [11]. The ABL was measured as the distance between the alveolar bone crest (ABC) and the cemento-enamel junction (CEJ) in ultrasound images [12]. However, there have been no studies for measuring the alveolar bone level of human periodontal tissue in OCT images.
Recently, deep learning-based methods have been used extensively to solve complex problems in medical and dental imaging [13]. A deep convolutional neural network (CNN), a type of deep learning (DL), is the most commonly used method for organ segmentation [14,15] as well as classification [16,17] and detection [18,19] of organs. Various attempts have been made to determine specific characteristics of target regions intended for detection and classification [20]. The CNNs show promising results in many OCT image processing tasks including border discrimination and structure segmentation [21]. The DL applications indicate potential utility for evaluating OCT images for narrow anterior chamber angle and angle-closure glaucoma based on slit-lamp image analysis, categorization, and detection in glaucoma treatment [22]. Other applications have included image quality improvement [23], identification of ocular disease biomarkers [24], and keratoplasty post-surgery screening [21]. Although DL approaches have shown considerable promise in OCT image analysis, their application in OCT images remains in an early stage [24–26].
Automatic and reliable segmentation of the human periodontal tissue is essential to accurate measurement of ABL in the periodontal OCT images. However, the periodontal structures show poor contrast, unclear borders of the alveolar bones, and structural deformations in OCT images, and manual segmentation of the periodontal structures in OCT images is time- and labor-consuming. There have been no studies to measure the alveolar bone level quantitatively and automatically in OCT images. Therefore, the purpose of this study was to automatically segment the periodontal structures of the tooth enamel and the alveolar bone using CNN and to measure quantitatively and automatically the ABL by detecting the CEJ and the ABC in OCT images.
2. Methods
2.1 Data acquisition and preprocessing
We acquired OCT images from 11 patients (six female and five male patients, ages from 25 to 36) who visited Seoul National University Dental Hospital. The study was performed in accordance with the Declaration of Helsinki. The OCT images were obtained from the mandibular and maxillary incisors using a swept source OCT system (SSOCT) (Oz-tec Co., Ltd., Daegu, Korea) (Fig. 1). The center wavelength of the OCT source was 1,310 ± 10 nm with 50 kHz sweep frequency, and output power was 16 mW on average. The system was capable of capturing 500 B-scan images per second, and its axial ($\Delta z$) and lateral ($\Delta x$) resolutions were 17.161 $\mu $m, and 10.03 $\mu $m, respectively, in air.
Fig. 1. (a) OCT image acquisition from the incisor of a patient (b) using a sweep-source OCT system consisting of workstation (W), source (S), probe (P), and head rest (H).
The patient head was fixed with a head rest (Anatomical Concepts, Scotland, UK), and the patient wore a silicone mouth bar (JN Pacific, Anyang, Korea) to keep the mouth open. The laser source from the OCT system emitted on the patient tooth surface was exactly aligned with the longitudinal axis of the teeth with reference to the direction of the mesial and distal margins of the tooth crown. During the process, a customized Visible Detector Card (Edmund Optics Korea, Seoul, Korea) attached on a piece of the acrylic body (215 mm x 23 mm) was used to check the exact location and alignment of the laser source on the tooth surface.
We selected a total of 500 B-scan images showing periodontal structures from all obtained images. Images obtained from teeth with large anatomical variations and prosthetic treatments were excluded. The original B-scan images of 756 × 768 pixels were cropped to 600 × 700 pixels to remove unnecessary noisy areas, which retained an area large enough to encompass all anatomical regions for experiment. We obtained OCT images from 11 patients, and the dataset were was separated into 7 (400 images), 2 (50 images), and 2 (50 images) patients for training, validation, and test sets, respectively. The number of images acquired from each patient was between 20 and 30 images. Finally, the training images were flipped horizontally, and increased 2-fold as a result.
One periodontist with seven years of experience annotated the 500 images manually for segmenting the alveolar bone and tooth enamel regions using software (Labelbox, CA, USA). Kakizaki et al. observed the periodontal profile, and described the appearance of the tooth and periodontal structures in detail in OCT images [6], and the segmentation of the structures was performed manually based on this. The authors confirmed that the manual segmentation was performed exactly according to the description by comparing the original image with the labeled image.
2.2 Calibration of the OCT image
The accurate axial resolution of the SSOCT images could be acquired or computed from the real interference spectrum. Generally, the axial resolution of the image could be approximated as ($\Delta z \approx \frac{{\lambda _2^c}}{{\Delta \lambda }}$) for a wide range of spectral shapes [27]. The sweep source of the OCT system was characterized by a continuous increase in optical frequency from ${\lambda _{start}}$ to ${\lambda _{end}}$, where the sweep range could be expressed as $\Delta \lambda $ ($\Delta \lambda = |{{\lambda_{end}} - {\lambda_{start}}} |$) [27].
In order to obtain the real dimensions, it was necessary to calibrate the OCT image using the refractive index of the appropriate anatomical region as the amount of backscattered light in dense tissues decreased exponentially with depth, resulting in compression of the axial image. In this study, we used 1.41 as the refractive index of the gingival tissue for calibration, a value used in other studies (Eq. (1) and (2)) [9,28,29]. Then, the real distance was calculated from the measurements by converting the pixel distance (w, h) into the actual distance (Height, Width, Distance) using calibration (Eq. (3)).
2.3 Overall procedure
The developed method consisted of two stages for measuring the ABL (Fig. 2). In the segmentation stage, the periodontal structures of the tooth enamel and the alveolar bone were automatically segmented in OCT images using the U-Net based CNN models. After training, the models produced segmentation maps of the tooth enamel and the alveolar bone regions. In the detection and measurement stage, the CEJ and the ABC were detected automatically, and the ABL was determined as the distance between CEJ and ABC using image processing (Fig. 3). Given the segmentation maps, we extracted the contours of the enamel of the tooth and the alveolar bone. Then, we calculated the ABL as the closest distance between the CEJ and the ABC pixels. Finally, the real distance of the ABL between the CEJ and the ABC was determined using the calibration explained above.
Fig. 2. The overall procedure for measuring the alveolar bone level (ABL) consisting of segmentation and measurement stages. At the segmentation stage, U-Net based networks were used to segment bone and enamel regions automatically in OCT images. At the detection and measurement stage, the ABL of the distance between the cemento-enamel junction (CEJ) and the alveolar bone crest (ABC) was calculated from the segmentation map produced by U-Net based networks.
Fig. 3. (a) An OCT image of periodontal tissues and surroundings of dentin (D), enamel (E), junctional epithelium (JE), gingival epithelium (GE), alveolar bone (AB), and connective tissue (CT) and an image of (b) the alveolar bone level (ABL) as the distance between the cemento-enamel junction (CEJ) and the alveolar bone crest (ABC).
2.4 Segmentation of periodontal structures using CNN models
We used three U-Net-based CNNs of U-Net, Dense-UNet, and U2-Net to segment the periodontal structures of the alveolar bone and the tooth enamel regions in OCT images (Fig. 4). The U-Net, which has a U-shape structure, is one of the popular deep networks for medical image segmentation [30]. It consists of an encoder part to capture context of input images and a symmetric decoder part to recover image resolution. The encoder consists of five levels of 3 × 3 convolution layers and max-pooling layers, while the decoder part has the same numbers of convolutional layers and up-sampling layers. The U-Net has approximately 7.7 million trainable parameters. The Dense-UNet [31] has a similar structure to U-Net, where densely connected blocks are used in the encoder and decoder parts for more efficient image segmentation. The encoder is composed of five densely connected blocks with transition blocks of stride 2. The decoder has the same number of densely connected blocks with up-sampling layers for recovering image resolution. The Dense-UNet has approximately 15.4 million trainable parameters.
Fig. 4. The U2-Net architecture consisting of a residual U-block (RSU) and side output layers for deep supervision. The En_n and De_n are encoder and decoder stages at level n based on the RSU block, respectively. The S(n)side is a side output layer at level n for deep supervision. The Sfuse is a fusing layer consisting of a 1 × 1 convolution layer and Softmax activation function to generate the final segmentation map for alveolar bone and tooth enamel regions.
The U2-Net architecture consists of a residual U-block (RSU) and side output layers for deep supervision [32]. It has a stacking U-shape structure like the encoder and decoder, where each stage consists of residual U-block to capture richer local and global information. Also, a deep supervision strategy is adopted by fusing predicted segmentation maps from each stage of decoder to learn rich hierarchical representations. The U2-Net has approximately 19.7 million trainable parameters.
We modified the input resolution of the networks from 512 × 512 pixels to 600 × 700 pixels, which provided sufficient resolution to predict the enamel of the tooth and the alveolar bone regions. By adopting a multi-class segmentation approach, networks output each segmentation map for the tooth enamel and the alveolar bone regions of the same size as the input. The networks were trained using the loss function of the Dice similarity coefficient (DSC) defined as $DSC({G,P} )= 1 - \left( {2\mathop \sum \limits_i^n {G_i}{P_i}} \right)/\left( {\mathop \sum \limits_i^n {G_i} + \mathop \sum \limits_i^n {P_i}} \right)$, where n was the number of pixels, G was the ground truth, and P was the prediction results [33] from a total of 300 epochs with a mini-batch size of 32. An adaptive moment estimation solver was used to optimize the network with a learning rate of 0.0001 and momentum of 0.9. This process was implemented with Python3 based on Keras with a Tensorflow backend using a single NVIDIA Titan RTX GPU 24G.
2.5 Measurement of alveolar bone level
The contours of the tooth enamel and the alveolar bone regions segmented by the CNN predictions were extracted as a sequence of pixels at the border between the background and the object regions using the Suzuki & Be algorithm [34]. The process of the CEJ and ABC detections was simply implemented by detecting the closest pixels on the contours of two segmented areas. The one of pixels on the contour of the tooth enamel region closest to the alveolar bone was detected as the CEJ and the counterpart pixel on that of the alveolar bone towards the enamel region as ABC (Fig. 3). We calculated the ABL as the closest pixel distance between the CEJ and the ABC and determined the actual distance of the ABL between the CEJ and the ABC using the calibration explained above.
2.6 Performance evaluation of the CNN models
We evaluated the segmentation performance of the CNN models using the test dataset not used for training. For the performance, we used five evaluation metrics of Jaccard index (JI, TP/(TP + FN + FP)), Dice similarity coefficient (DSC, F1-score, 2×TP/(2×TP + FN + FP)), recall (RC, TP/(TP + FN)), precision (PR, TP/(TP + FP)), and Hausdorff distance (HD) [35], where true positive (TP) was the number of pixels for which the model correctly predicted the positive class, false positive (FP) was the number of pixels for which the model incorrectly predicted the positive class, and false negative (FN) was the number of pixels for which the model incorrectly predicted the negative class. Hausdorff distance (HD) was defined as ($\max \left\{ {\mathop {\max }\limits_{p \in {P_i}} [{Dist({p,{G_i}} )} ],\mathop {\max }\limits_{g \in {P_i}} [{Dist({p,{P_i}} )} ]} \right\})$, the Euclidean distance between the furthest pixels on the contours of two segmented areas [35]. ${\textrm{P}_\textrm{i}}$ was the ${\textrm{I}_{th}}$ predicted image, and ${\textrm{G}_i}$ was the matched ground truth, where the smaller was HD, the greater was the similarity.
To evaluate the detection accuracy of the CEJ and the ABC, we used mean distance difference (MDD), mean absolute error (MAE), successful detection rate (SDR), and root mean square error (RMSE). The MDD was defined as $\left( {\mathop \sum \limits_{i = 1}^n {R_i}} \right)/n$, where R indicated the Euclidean distance between detected points from ground truth and deep learning predictions for the CEJ and the ABC, and n was the number of data points. The MAE was a measure of errors between the x and y coordinates of detected points from ground truth and deep learning predictions for the CEJ and the ABC. The SDR was the percentage of successfully detected distances within the range of 0.1 mm, 0.3 mm, 0.5, mm 0.7 mm, and 1.0 mm for the CEJ and the ABC. The RMSE was used to measure the difference of the ABL between ground truth and detected results from CNN predictions. The RMSE was defined as $\sqrt {\frac{1}{n}\mathop \sum \limits_{i = 1}^n {{({ABL_g^i - ABL_p^i} )}^2}} $, where n indicated the number of data points. The $AB{L_g}$ and $AB{L_p}$ were ground truth and detected results from CNN predictions, respectively.
Furthermore, we compared ABLs at the incisors between the ground truth and deep learning predictions. We analyzed the correlation and reliability between the ground truth and the predicted ABLs using Pearson correlation coefficients (PCC) and intraclass correlation coefficients (ICC), respectively, using SPSS (ver. 26, SPSS Inc., Chicago, IL, USA). The ANOVA test for HD and MDD and the student’s t-test for the ABL were also performed. The p-value of 0.05 was used as the significance level.
3. Results
The segmentation and detection performance of U-Net, Dense-UNet, and U2-Net were evaluated for 50 OCT images not used for training. Table 1 shows the quantitative comparison of segmentation performance of U-Net, Dense-UNet, and U2-Net. The U2-Net showed higher segmentation performance for JI, recall, and HD than Dense-UNet and U-Net for the alveolar bone region and higher performance of the JI, DSC, precision, and HD than Dense-UNet and U-Net for the tooth enamel region. There were significant differences between HDs by U-Net and other models according to ANOVA (p > 0.05) (Table 1). The overall segmentation performance of U2-Net was superior to those of U-Net and Dense-UNet in the tooth enamel and alveolar bone regions.
Table 1. Segmentation performances for the alveolar bone and tooth enamel regions in OCT images by CNN models (JI, Jaccard index; DSC, Dice similarity coefficient; PR, Precision; RC, Recall; HD, Hausdorff distance (mm)) (*: significant difference between HDs by CNN models (p < 0.05)).
Table 2 shows detection accuracies for the ABC and the CEJ as MDD, MAE, and SDR between the ground truth and CNN predictions. The MDD by CNN models ranged from 0.19 to 0.22 mm for ABC and from 0.18 to 0.32 mm for CEJ. All CNN models exhibited the MAE of less than 0.25 mm in the x and y coordinates and greater than 90% SDR at 0.5 mm for both the ABC and the CEJ (Tables 2). There were significant differences between MDDs and MAE along with y-direction by U-Net and other model predictions for the CEJ according to ANOVA (p > 0.05) (Table 2). The CNN models produced high detection accuracy for the CEJ compared with the clinical tolerance limit of 0.5 mm based on direct probing measurement [36]. The U2-Net model showed the highest SDR in detecting the ABC and the CEJ within the clinical tolerance limit of 0.5 mm.
Table 2. Detection accuracies for the alveolar bone crest and cemento-enamel junction as mean distance difference (MDD) (mm), mean absolute error (MAE) (mm), and successful detection rate (SDR) (%) between the ground truth and detected points from CNN predictions (*: significant difference between MDDs by CNN predictions (p < 0.05)).
The ABL measurements at the mandibular incisors showed no significant difference between the ground truth and predictions by U2-Net (p > 0.05), while others showed a significant difference between the ground truth and CNN predictions (p < 0.05) according to the student’s t-test (Table 3). There were no significant differences among the RMSEs by the CNN models at both incisors according to ANOVA (p > 0.05).
Table 3. Alveolar bone level (ABL) (mm) measurements and the root mean square error (RMSE) (mm) between the ground truth and detected results from CNN predictions at the incisors (*: significant difference between the ground truth and detected results from CNN predictions (p < 0.05), †: measured at both incisors).
Table 4 shows the PCC and ICC of ABL measurements at the incisors between the ground truth and CNN predictions. The PCC values were 0.697, 0.813, and 0.767 by U-Net, Dense-UNet, and U2-Net, respectively (p < 0.05), and indicated strong correlations of ABL measurements between the ground truth and the CNN predictions. The highest PCC was between Dense-UNet and the ground truth. The ICC values were 0.820, 0.896, and 0.868 by U-Net, Dense-UNet, and U2-Net, respectively (p < 0.05), and indicated excellent reliability of ABL measurements between the ground truth and the CNN predictions. The ICC between the Dense-UNet and the ground truth showed the highest reliability.
Table 4. The Pearson correlation coefficients (PCCs) and intraclass correlation coefficients (ICCs) of ABL measurements between the ground truth and detected results from CNN predictions at the incisors (*: significant correlation between the ground truth and detected results from CNN predictions (p < 0.05)).
The qualitative results in Fig. 5 show that the U2-Net is more accurate in the segmentation of the tooth enamel and the alveolar bone areas than Dense-UNet and U-Net, with more true positives, fewer false positives, and fewer false negatives. As a result, the U2-Net predicted the border features of the alveolar bone crest and the sharp CEJ more accurately and showed the best segmentation ability to describe the CEJ and the ABC (Fig. 5).
Fig. 5. The first column shows the original OCT image; the second column shows the ground truth (red) manually labeled by a periodontist; and the third, fourth, and fifth columns show segmentations of the alveolar bone and tooth enamel regions by U2-Net, Dense-UNet, and U-Net, respectively. The TP (yellow), FP (blue), and FN (red) segmentations of the alveolar bone and tooth enamel regions are indicated by the CNN models. The dotted lines indicate the ABL between the ground truth CEJ and ABC, while the solid lines show the ABL between the predicted CEJ and ABC by the CNN models. The first four rows indicate measurements at the maxillary incisors, and the rest at the mandibular incisors.
4. Discussion
In this study, the tooth enamel and alveolar bone regions were automatically segmented using CNN models, and the CEJ and the ABC were automatically detected using image processing in OCT images. The CEJ as a reference point enabled quantitative measurement of the alveolar bone level in OCT images. As a result, the ABL as the distance from the CEJ to the ABC was quantitatively and automatically measured in human intraoral OCT images. In previous studies, the alveolar bone region could be observed, but the CEJ was not considered for measurement of the ABL in OCT images [8]. However, it is clinically important to consider the CEJ as a reference point for quantitative measurement of the ABL. This is the first study to measure quantitatively and automatically the ABL in human OCT images by applying deep learning.
Accurate assessment of the ABL is important for diagnosis and progression of periodontitis. Generally, when the ABL is greater than 2 mm, it is considered alveolar bone loss [37,38]. In addition, it is important to quantitatively measure the ABL in crown lengthening procedures. If the ABL as the distance from the CEJ to the ABC is greater than 2 mm, only the gingiva need to be treated, but if it is less than 2 mm, alveolar bone reduction also should be performed [39]. Therefore, accurate evaluation of the ABL is necessary for proper treatment planning.
Panoramic, bitewing, or periapical radiographs have been mainly used to evaluate the alveolar bone level. However, because these are two-dimensional images, the buccal or lingual bones overlap the teeth and cannot be observed properly, and only the proximal bones can be evaluated [40]. Also, the underestimation of bone loss was reported to range from 13 to 32% in panorama, 11 to 23% in bitewing, and 9 to 20% in periapical radiographs [2], complicating diagnosis of quantitative bone loss, especially in the case of incipient bone loss. Panoramic images have the additional disadvantage of being distorted and lacking in detail [41]. Though CBCT or CT can provide cross-sectional images of intraoral tissue, due to concerns of excessive radiation exposure, alternative diagnostic tools have been suggested. Ultrasound is considered a promising diagnostic tool, and many studies have been conducted on its clinical application. Recently, there was a study on segmenting alveolar bone and locating alveolar bone crest using machine learning in ultrasound images [42]. On the other hand, there are several advantageous features of OCT as a promising alternative diagnostic tool, such as higher resolution than radiographic or ultrasound images [43], and OCT avoids the radiation exposure inherent to radiographic imaging and the hassle of using gel.
There have been studies on the automatic segmentation of teeth and surrounding tissues in OCT images using artificial intelligence [44,45]. Lai et al. experimentally verified that teeth, gingival, and alveolar bones could be segmented by applying intensity quantization for boundary identification in noisy maps using a deep network [46]. Wang et al. measured the volume of the gingiva by segmenting the teeth and the gingiva in a 3D OCT image [44]. The advantage of automatic segmentation compared to manual delineation was that it was time- and labor-efficient and could reduce variation caused by subjective judgment between raters [42]. The authors created an image analysis pipeline that incorporated a deep learning-based segmentation model in OCT images [47–49]. The segmentation model was created to find lesions in individual B-scans automatically [50]. Several studies have measured the alveolar bone level in panoramic or periapical radiographs using deep learning [41,51,52]. However, there were no studies to measure the alveolar bone level based on automatic segmentation of the tooth enamel and alveolar bone in OCT images using deep learning.
The major finding in this study is that the CNN models accurately segmented the periodontal structures of the tooth enamel and the alveolar bone to measure quantitatively and automatically the ABL by detecting the CEJ and the ABC in OCT images. In our experiments, the CNN models showed high segmentation accuracies in both the tooth enamel and the alveolar bone regions, where U2-Net outperformed U-Net and Dense-UNet in the ABC and the CEJ segmentation by preserving sharp boundary details of those. In this study, we applied three U-Net based CNNs of U-Net, Dense-UNet, and U2-Net to segment the periodontal structures of the alveolar bone and the tooth enamel regions in OCT images. The U-Net with U-shaped encoder-decoder architecture has been widely used in biomedical image synthesis, segmentation, and denoising [30,31,53]. Particularly, it exhibited high performance on semantic segmentation by precisely predicting semantic categories on the entire pixel data of the image including a broad range of information of objects in medical image analysis [54,55]. The U2-Net can acquire more local and global information from both shallow and deep layers due to the layered residual U-structured blocks [32]. The residual U-block employed in U2-Net allowed for extraction of intra-stage multi-scale features without reducing feature map resolution. Diverse receptive fields and richer multi-scale contextual characteristics considerably increased segmentation performance of the U2-Net especially for finding edge information [32]. By inherent characteristics of network architecture, the U2-Net accurately segmented both the tooth enamel and alveolar bone regions including boundary details of those by simultaneously learning global and local features of the alveolar bone and border of the sharp CEJ.
We compared our method with several studies that have been performed to measure the distance from the CEJ to the ABC in CBCT images. Wang et al. reported that the mean distance was 1.8 ± 0.7 mm for the maxillary central incisors and 1.9 ± 0.6 mm for the lateral incisors in a CBCT image [56]. Lee et al. reported the mean distance as 2.03 ± 0.61 mm for central incisors and 2.46 ± 0.65 mm for lateral incisors [57]. El Nahass et al. reported a mean distance of 2.10 ± 0.85 mm for maxillary central incisors and 2.09 ± 0.72 mm for lateral incisors [58]. In our results, the alveolar bone level (ABL) at incisors was 2.087 ± 0.568 mm and 2.005 ± 0.557 mm by Dense-UNet and U2-Net predictions, respectively. In comparison with the previous works using CBCT images, our method achieved comparable performance with those of them. As the distance from CEJ to ABC increased with age [56,59], the present subjects in their 20s and 30s might have caused the difference of the ABL measurements. The use of modalities other than OCT to measure the ABL also might have increased the difference in results. Nonetheless, the ABL by CNN predictions demonstrated high correlation and reliability with the ground truth in OCT images.
We developed a method to automatically segment the periodontal structures of the tooth enamel and the alveolar bone using deep learning and to quantitatively measure the ABL by automatically detecting the CEJ and the ABC in OCT images. The method has a number of advantages in that it is time- and labor-efficient and can minimize the subjective error in segmentation and detection of periodontal tissue not clearly delineated in OCT images.
The limitations of this study are as follows. First, as the OCT probe was not mobile, the ABL was measured at only the buccal surface of the anterior teeth. Second, the alveolar bone level was measured at only the periodontal tissue of the healthy subjects. We will collect more dataset including healthy and diseased subjects and evaluate the segmentation performance of the models using them. Third, as the laser source of the OCT equipment on the tooth surface was not visible, it might not have been completely aligned with the longitudinal axis of the tooth even if it was checked using a visible laser detector. In future studies, we will apply this method to the buccal or lingual surfaces of the periodontal tissue of subjects with and without periodontitis using an improved OCT system with a mobile probe.
5. Conclusion
In this study, we applied the CNN models to automatically segment the periodontal structures of the tooth enamel and the alveolar bone and to quantitatively measure the ABL by automatically detecting the CEJ and the ABC in OCT images. The CNN models showed high segmentation accuracies in the tooth enamel and alveolar bone regions, and the ABL measured by detection results from CNN predictions demonstrated high correlation and reliability with the ground truth in OCT images. The proposed method has the potential to be utilized in periodontitis diagnosis or other clinical periodontal procedures.
Funding
Seoul National University (860-20210105); Korea Medical Device Development Fund (1711137883, KMDF_PR_20200901_0011); Korea Medical Device Development Fund (1711138289, RS-2020-KD00014).
Disclosures
The authors have no conflicts of interest to report
Data availability
Data underlying the results presented in this paper are not publicly available but can be obtained from the authors upon reasonable request.
References
1. B. L. Pihlstrom, B. S. Michalowicz, and N. W. Johnson, “Periodontal diseases,” Lancet 366(9499), 1809–1820 (2005). [CrossRef]
2. L. Akesson, J. Hakansson, and M. Rohlin, “Comparison of panoramic and intraoral radiography and pocket probing for the measurement of the marginal bone level,” J Clin. Periodontol. 19(5), 326–332 (1992). [CrossRef]
3. L. Ds and B. S. Latief, “Evaluation of Impacted Mandibular Third Molar Position in Relation to Mandibular Canal on Panoramic Radiography compared to Cone-Beam Computed Tomography,” Journal of International Dental & Medical Research 12 (2019).
4. A. Suomalainen, E. Pakbaznejad Esmaeili, and S. Robinson, “Dentomaxillofacial imaging with panoramic views and cone beam CT,” Insights Imaging 6(1), 1–16 (2015). [CrossRef]
5. C. C. Mota, L. O. Fernandes, R. Cimoes, and A. S. Gomes, “Non-Invasive Periodontal Probing Through Fourier-Domain Optical Coherence Tomography,” J Periodontol. 86(9), 1087–1094 (2015). [CrossRef]
6. S. Kakizaki, A. Aoki, M. Tsubokawa, T. Lin, K. Mizutani, G. Koshy, A. Sadr, S. Oda, Y. Sumi, and Y. Izumi, “Observation and determination of periodontal tissue profile using optical coherence tomography,” J Periodontal. Res. 53(2), 188–199 (2018). [CrossRef]
7. Y. S. Hsieh, Y. C. Ho, S. Y. Lee, C. W. Lu, C. P. Jiang, C. C. Chuang, C. Y. Wang, and C. W. Sun, “Subgingival calculus imaging based on swept-source optical coherence tomography,” J. Biomed. Opt. 16(7), 071409 (2011). [CrossRef]
8. S. H. Kim, S. R. Kang, H. J. Park, J. M. Kim, W. J. Yi, and T. I. Kim, “Improved accuracy in periodontal pocket depth measurement using optical coherence tomography,” J. Periodontal Implant Sci. 47(1), 13–19 (2017). [CrossRef]
9. L. O. Fernandes, C. Mota, L. S. A. de Melo, M. U. S. da Costa Soares, D. da Silva Feitosa, and A. S. L. Gomes, “In vivo assessment of periodontal structures and measurement of gingival sulcus with Optical Coherence Tomography: a pilot study,” J Biophotonics 10(6-7), 862–869 (2017). [CrossRef]
10. S. Kim, S. R. Kang, H. J. Park, B. Kim, T. I. Kim, and W. J. Yi, “Quantitative measurement of peri-implant bone defects using optical coherence tomography,” J. Periodontal Implant Sci. 48(2), 84–91 (2018). [CrossRef]
11. A. F. Zuluaga, R. Daher, and I. Krejci, “Toward optical monitoring of alveolar bone level, assessing the structure of periodontal tissues with optical coherence tomography: an ex-vivo animal pilot study,” in Photonic Therapeutics and Diagnostics in Dentistry, Head and Neck Surgery, and Otolaryngology (International Society for Optics and Photonics, 2021), p. 116270G.
12. K. T. Nguyen, C. Pacheco-Pereira, N. R. Kaipatur, J. Cheung, P. W. Major, and L. H. Le, “Comparison of ultrasound imaging and cone-beam computed tomography for examination of the alveolar bone level: A systematic review,” PLoS One 13, e0200596 (2018). [CrossRef]
13. J. Schmidhuber, “Deep learning in neural networks: An overview,” Neural networks 61, 85–117 (2015). [CrossRef]
14. M. Kallenberg, K. Petersen, M. Nielsen, A. Y. Ng, P. Diao, C. Igel, C. M. Vachon, K. Holland, R. R. Winkel, and N. Karssemeijer, “Unsupervised deep learning applied to breast density segmentation and mammographic risk scoring,” IEEE Trans. Med. Imaging 35(5), 1322–1331 (2016). [CrossRef]
15. X. Zhao, Y. Wu, G. Song, Z. Li, Y. Zhang, and Y. Fan, “A deep learning model integrating FCNNs and CRFs for brain tumor segmentation,” Med. Image Anal. 43, 98–111 (2018). [CrossRef]
16. A. Esteva, B. Kuprel, R. A. Novoa, J. Ko, S. M. Swetter, H. M. Blau, and S. Thrun, “Dermatologist-level classification of skin cancer with deep neural networks,” Nature 542(7639), 115–118 (2017). [CrossRef]
17. X. W. Gao, R. Hui, and Z. Tian, “Classification of CT brain images based on deep learning networks,” Computer methods and programs in biomedicine 138, 49–56 (2017). [CrossRef]
18. A. Teramoto, H. Fujita, O. Yamamuro, and T. Tamaki, “Automated detection of pulmonary nodules in PET/CT images: Ensemble false-positive reduction using a convolutional neural network technique,” Med. Phys. 43(6Part1), 2821–2827 (2016). [CrossRef]
19. A. Y. Hannun, P. Rajpurkar, M. Haghpanahi, G. H. Tison, C. Bourn, M. P. Turakhia, and A. Y. Ng, “Cardiologist-level arrhythmia detection and classification in ambulatory electrocardiograms using a deep neural network,” Nat. Med. 25(1), 65–69 (2019). [CrossRef]
20. A. Krizhevsky, I. Sutskever, and G. E. Hinton, “Imagenet classification with deep convolutional neural networks,” Advances in neural information processing systems 25 (2012).
21. M. Treder, J. L. Lauermann, M. Alnawaiseh, and N. Eter, “Using deep learning in automated detection of graft detachment in Descemet membrane endothelial keratoplasty: a pilot study,” Cornea 38(2), 157–161 (2019). [CrossRef]
22. D. S. J. Ting, V. H. Foo, L. W. Y. Yang, J. T. Sia, M. Ang, H. Lin, J. Chodosh, J. S. Mehta, and D. S. W. Ting, “Artificial intelligence for anterior segment diseases: Emerging applications in ophthalmology,” Br. J. Ophthalmol. 105(2), 158–168 (2021). [CrossRef]
23. J. Ouyang, T. S. Mathai, K. Lathrop, and J. Galeotti, “Accurate tissue interface segmentation via adversarial pre-segmentation of anterior segment OCT images,” Biomed. Opt. Express 10(10), 5291–5324 (2019). [CrossRef]
24. T. Schlegl, P. Seeböck, S. M. Waldstein, U. Schmidt-Erfurth, and G. Langs, “Unsupervised anomaly detection with generative adversarial networks to guide marker discovery,” in International conference on information processing in medical imaging (Springer, 2017), pp. 146–157.
25. M. Pekala, N. Joshi, T. Liu, N. M. Bressler, D. Cabrera DeBuc, and P. Burlina, “Oct segmentation via deep learning: A review of recent work,” in Asian Conference on Computer Vision (Springer, 2018), pp. 316–322.
26. J. Kugelman, D. Alonso-Caneiro, S. A. Read, J. Hamwood, S. J. Vincent, F. K. Chen, and M. J. Collins, “Automatic choroidal segmentation in OCT images using supervised deep learning methods,” Sci. Rep. 9(1), 13298 (2019). [CrossRef]
27. T. Klein and R. Huber, “High-speed OCT light sources and systems,” Biomed. Opt. Express 8(2), 828–859 (2017). [CrossRef]
28. H. Cho, J. Lee, J. Song, M. Jeon, and J. Kim, “Quantitative Depth Measurement of the Human Gingival Sulcus Based on Interferometry System with Detection Algorithm,” International Journal of Mechanical Engineering and Robotics Research 9 (2020).
29. Z. Meng, X. S. Yao, H. Yao, Y. Liang, T. Liu, Y. Li, G. Wang, and S. Lan, “Measurement of the refractive index of human teeth by optical coherence tomography,” J. Biomed. Opt. 14(3), 034010 (2009). [CrossRef]
30. O. Ronneberger, P. Fischer, and T. Brox, “U-net: Convolutional networks for biomedical image segmentation,” in International Conference on Medical image computing and computer-assisted intervention (Springer, 2015), pp. 234–241.
31. S. Cai, Y. Tian, H. Lui, H. Zeng, Y. Wu, and G. Chen, “Dense-UNet: a novel multiphoton in vivo cellular image segmentation model based on a convolutional neural network,” Quant Imaging Med. Surg. 10(6), 1275–1285 (2020). [CrossRef]
32. X. Qin, Z. Zhang, C. Huang, M. Dehghan, O. R. Zaiane, and M. Jagersand, “U2-Net: Going deeper with nested U-structure for salient object detection,” Pattern Recognition 106, 107404 (2020). [CrossRef]
33. J. Ma, J. Chen, M. Ng, R. Huang, Y. Li, C. Li, X. Yang, and A. L. Martel, “Loss odyssey in medical image segmentation,” Med. Image Anal. 71, 102035 (2021). [CrossRef]
34. S. Suzuki, “Topological structural analysis of digitized binary images by border following,” Computer vision, graphics, and image processing 30(1), 32–46 (1985). [CrossRef]
35. D. P. Huttenlocher, G. A. Klanderman, and W. J. Rucklidge, “Comparing images using the Hausdorff distance,” IEEE Trans. Pattern Anal. Machine Intell. 15(9), 850–863 (1993). [CrossRef]
36. A. Badersten, R. Nilvéaus, and J. Egelberg, “Reproducibility of probing attachment level measurements,” J. Clin. Periodontol. 11(7), 475–485 (1984). [CrossRef]
37. M. F. Helmi, H. Huang, J. M. Goodson, H. Hasturk, M. Tavares, and Z. S. Natto, “Prevalence of periodontitis and alveolar bone loss in a patient population at Harvard School of Dental Medicine,” BMC Oral Health 19(1), 254 (2019). [CrossRef]
38. X. Zhang, Y. Li, Z. Ge, H. Zhao, L. Miao, and Y. Pan, “The dimension and morphology of alveolar bone at maxillary anterior teeth in periodontitis: a retrospective analysis-using CBCT,” Int J Oral Sci 12(1), 4 (2020). [CrossRef]
39. H. Nowzari, S. Molayem, C. H. K. Chiu, and S. K. Rich, “Cone beam computed tomographic measurement of maxillary central incisors to determine prevalence of facial alveolar bone width≥ 2 mm,” Clinical Implant Dentistry and Related Research 14(4), 595–602 (2012). [CrossRef]
40. E. H. L. Nascimento, A. C. C. Oenning, B. B. Freire, H. Gaêta-Araujo, F. Haiter-Neto, and D. Q. Freitas, “Comparison of panoramic radiography and cone beam CT in the assessment of juxta-apical radiolucency,” Dentomaxillofacial Radiology 47(1), 20170198 (2018). [CrossRef]
41. J. H. Lee, D. H. Kim, S. N. Jeong, and S. H. Choi, “Diagnosis and prediction of periodontally compromised teeth using a deep learning-based convolutional neural network algorithm,” J. Periodontal Implant Sci. 48(2), 114–123 (2018). [CrossRef]
42. K. C. T. Nguyen, D. Q. Duong, F. T. Almeida, P. W. Major, N. R. Kaipatur, T. T. Pham, E. H. M. Lou, M. Noga, K. Punithakumar, and L. H. Le, “Alveolar Bone Segmentation in Intraoral Ultrasonographs with Machine Learning,” J. Dent. Res. 99(9), 1054–1061 (2020). [CrossRef]
43. R. A. Erdelyi, V. F. Duma, C. Sinescu, G. M. Dobre, A. Bradu, and A. Podoleanu, “Dental Diagnosis and Treatment Assessments: Between X-rays Radiography and Optical Coherence Tomography,” Materials 13(21), 4825 (2020). [CrossRef]
44. G. Wang, N. M. Le, X. Hu, Y. Cheng, S. L. Jacques, H. Subhash, and R. K. Wang, “Semi-automated registration and segmentation for gingival tissue volume measurement on 3D OCT images,” Biomed. Opt. Express 11(8), 4536–4547 (2020). [CrossRef]
45. N. K. Ravichandran, H. T. Lakshmikantha, H.-S. Park, M. Jeon, and J. Kim, “Micron-scale human enamel layer characterization after orthodontic bracket debonding by intensity-based layer segmentation in optical coherence tomography images,” Sci. Rep. 11(1), 1–15 (2021). [CrossRef]
46. Y. C. Lai, C. H. Chiu, Z. Q. Cai, J. Y. Lin, C. Y. Yao, D. Y. Lyu, S. Y. Lee, K. W. Chen, and I. Y. Chen, “OCT-Based Periodontal Inspection Framework,” Sensors 19(24), 5496 (2019). [CrossRef]
47. T. T. Khaing, T. Okamoto, C. Ye, M. A. Mannan, H. Yokouchi, K. Nakano, P. Aimmanee, S. S. Makhanov, and H. Haneishi, “ChoroidNET: A Dense Dilated U-Net Model for Choroid Layer and Vessel Segmentation in Optical Coherence Tomography Images,” IEEE Access 9, 150951–150965 (2021). [CrossRef]
48. N. Siddique, S. Paheding, C. P. Elkin, and V. Devabhaktuni, “U-net and its variants for medical image segmentation: A review of theory and applications,” IEEE Access 9, 82031–82057 (2021). [CrossRef]
49. B. M. Velden, M. Veta, J. Pluim, M. Alberti, and F. G. Heslinga, “Radial U-Net: Improving DMEK Graft Detachment Segmentation in Radial AS-OCT Scans,” in International Workshop on Ophthalmic Medical Image Analysis (Springer, 2021), pp. 72–81.
50. B. Hassan, S. Qin, R. Ahmed, T. Hassan, A. H. Taguri, S. Hashmi, and N. Werghi, “Deep learning based joint segmentation and characterization of multi-class retinal fluid lesions on OCT scans for clinical use in anti-VEGF therapy,” Comput. Biol. Med. 136, 104727 (2021). [CrossRef]
51. J. Kim, H. S. Lee, I. S. Song, and K. H. Jung, “DeNTNet: Deep Neural Transfer Network for the detection of periodontal bone loss using panoramic dental radiographs,” Sci. Rep. 9(1), 17615 (2019). [CrossRef]
52. J. Krois, T. Ekert, L. Meinhold, T. Golla, B. Kharbot, A. Wittemeier, C. Dorfer, and F. Schwendicke, “Deep Learning for the Radiographic Detection of Periodontal Bone Loss,” Sci. Rep. 9(1), 8495 (2019). [CrossRef]
53. Z. Zhou, M. M. Rahman Siddiquee, N. Tajbakhsh, and J. Liang, “Unet++: A nested u-net architecture for medical image segmentation,” in Deep Learning in Medical Image Analysis and Multimodal Learning for Clinical Decision Support (Springer, 2018), pp. 3–11.
54. O. Oktay, J. Schlemper, L. L. Folgoc, M. Lee, M. Heinrich, K. Misawa, K. Mori, S. McDonagh, N. Y. Hammerla, and B. Kainz, “Attention u-net: Learning where to look for the pancreas,” arXiv preprint arXiv:1804.03999 (2018).
55. C. Li, Y. Tan, W. Chen, X. Luo, Y. He, Y. Gao, and F. Li, “ANU-Net: Attention-based Nested U-Net to exploit full resolution features for medical image segmentation,” Computers & Graphics 90, 11–20 (2020). [CrossRef]
56. H. M. Wang, J. W. Shen, M. F. Yu, X. Y. Chen, Q. H. Jiang, and F. M. He, “Analysis of facial bone wall dimensions and sagittal root position in the maxillary esthetic zone: a retrospective study using cone beam computed tomography,” Int J Oral Maxillofac Implants 29(5), 1123–1129 (2014). [CrossRef]
57. S. L. Lee, H. J. Kim, M. K. Son, and C. H. Chung, “Anthropometric analysis of maxillary anterior buccal bone of Korean adults using cone-beam CT,” J Adv. Prosthodont. 2(3), 92–96 (2010). [CrossRef]
58. H. El Nahass and S. N. Naim, “Analysis of the dimensions of the labial bone wall in the anterior maxilla: a cone-beam computed tomography study,” Clin. Oral Impl. Res. 26(4), e57–e61 (2015). [CrossRef]
59. A. Srebrzynska-Witek, R. Koszowski, I. Rozylo-Kalinowska, and M. Piskorz, “CBCT for estimation of the cemento-enamel junction and crestal bone of anterior teeth,” Open Med. 15(1), 774–781 (2020). [CrossRef]
