Disease activation maps for subgingival dental calculus identification based on intelligent dental optical coherence tomography

During the treatment of periodontitis, removal of dental calculus is essential but still tricky despite developments of several imaging modalities. In this research, we propose a novel approach to provide an intuitive guidance, to automatically detect the present of subgingival calculus, and to identify the site of the lesion in optical coherence tomography images based on convolutional neural network model and the class activation maps technique. Our result shows good visualizations of dental calculus both in B-scan and in volumetric view. We believe the improvement can break down the technical barrier and finally facilitate the prevalence of dental optical coherence tomography


| INTRODUCTION
Periodontitis is a chronic infection of periodontal tissues and is one of the major causes of tooth loss. Periodontitis not only deteriorates oral health but is also a risk factor of systemic diseases [1][2][3][4][5] according to research proposed by American Academy of Periodontology [6]. What is worse, periodontitis is one of the most prevailing diseases of human beings. A nationally representative study investigated the epidemiology of periodontitis during 2009 to 2014 and had shown that 42% of dentate US adults over 30 years old suffered from periodontitis [7]. The ultimate goal of periodontitis prevention and treatment is to control the growth of biofilm and to restore the supporting structures of the teeth [8]. Dental plaque, the primary etiological cause of periodontitis, is formed by the accumulation of bacterial matrix. Although dental plaque can be removed simply by brushing and flossing in general cleaning of the oral cavity, once it hardens and forms dental calculus by precipitation [9], the rough surface of dental calculus leads to more dental plaque accumulation. At first, dental plaque causes gingival inflammation and bleeding, which is not obvious in the early stage and easily ignored. Gradually, bacteria destroy gingival tissue and lead to both gingival and bone destruction, which is the primary factor of tooth loss in adults. Therefore, it is crucial and necessary to remove dental calculus completely in the treatment of periodontitis. Dental calculus is classified into supragingival calculus and subgingival calculus according to its forming position. Previous studies have shown that subgingival calculus is closely related to the progression of periodontitis [10][11][12]. The research evidence also suggests that subgingival calculus is the determinant factor in the periodontal loss of attachment [13] and periodontitis [14]. A successful periodontitis treatment relays on the effective removal of dental plaque and subgingival dental calculus [15].
Nowadays, dentists usually use ultrasonic scaling and periodontal curette to remove dental calculus. A complete examination of periodontitis includes recognition of gingival inflammation and extent of teeth mobility, periodontal probing using probe, and radiographic examination [16]. However, periodontal probing relies on clinicians' experiences, thus having poor reproducibility [17], and this procedure also causes patient discomforts. Radiograph is restricted by the resolution power and the imaging direction, which makes the images on the buccaland lingual-side overlapped [18], so radiograph is not reliable to be utilized in the early detection of dental calculus. There are various modalities developed to overcome this problem. Fiberoptic endoscopy visualizes the subgingival root surface in real-time so as to monitor residual calculus in a minimally invasive manner [19], but the incremental effect is still questionable [20]. Spectro-optical technology makes use of the characteristic spectral signature of subgingival calculus and achieves an acceptable accuracy; however, this method is restricted by either its technological limits or incomplete scanning process and thus still needs to be investigated [21]. Autofluorescence-based technology can differentiate calculus from the healthy tooth surface owing to several distinct fluorescence bands from calculus specimens with high accuracy but only under in vitro condition [22]. In addition to the detection-only systems, laser-based technology further combines the calculus-removal ability; nonetheless the use of photoablation in calculus removal is still on the debate among clinicians due to its cause of loss of root substance [23,24]. According to Sherman's research [25], even after currently the most efficient method, ultrasonic scaling, subgingival curettage and root planning performed by clinically-experienced dentists, 57% of the surface of the extracted teeth were still identified with residual dental calculus. The falsenegative rate and false-positive rate of subgingival dental calculus examination by dental probing are up to 77.4 and 11.8% respectively. The success of treatments much depends on the visibility of subgingival dental calculus the modalities provide [26]; however, none of the above can provide dentists real-time images throughout the process of teeth scaling.Optical coherence tomography (OCT) is a high-resolution tomographic optical imaging technique and has been well-developed in clinical ophthalmology and oncology. OCT can provide the micrometer resolution and an up to 0.8 mm gingiva penetration depth, which is also ideal for dental examinations. In 1998, OCT was firstly applied to oral imaging of soft tissue and hard tissue [27]. Studies of OCT on evaluations of dental caries, cracked tooth, dental calculus and dental fillings were then carried out afterward [28][29][30][31][32][33]. Our previous study proved that OCT is capable of imaging subgingival dental calculus at the presence of gingiva [34]. With the depth-resolved and real-time imaging ability, OCT is suitable for subgingival dental calculus detection. However, as other novel methods may encounter [35], since the usage habits and the requirement of pretraining for image interpretation, dental OCT has not yet thrived in spite of the development of various kinds of optical probe design [36]. The barrier from bench research to clinical implementation much restricts the development of dental OCT.
The problem draws us to its underlyingly profound insight into translational medicine. All diagnostic medical imaging techniques encounter the same question: What does a medical image really mean to us? Traditionally, medical images are interpreted and diagnosed by well-trained clinicians or technicians, but how does an image exactly link to the concept of a disease? If the concept of disease resides in specific patterns or substances in the medical image, we can possibly visualize and digitize a "disease." This is a crucial issue when it comes to the era of digital technology. Then the problem next is: how? Artificial intelligence sheds new light on this gap of bench-bedside study. With the help of deep learning algorithms, computer-aided detection (CAD) is now a trend in the medical imaging, and disease visualization becomes possible. We think the concept can especially work out for dental calculus such a concrete lesion.
Recently, Chia-Yen Lee employed machine learning: K-means, Markov random field (MRF) and other designed feature extraction methods to delineate the region of supragingival dental calculus [37], showing the feasibility of machine learning in calculus detection. In the same year, Masaki Tsubokawa also comprehensively investigated the clinical applicability of OCT on examination of subgingival dental calculus [38] and made the suspicious subgingival dental calculus more visible by simply applying two-tone processing based on the brightness value. Nonetheless, there is still no subgingival dental calculus detection been reported with artificial intelligence using OCT. Here we utilize the combination of convolutional neural network (CNN) and class activation map (CAM). With convolution layers in model architecture, CNN can see the spatial patterns, which differ from one category to another, and finally makes the classification of images possible. Previous researches have shown the undeniable diagnosing power of deep learning recognizing lesions in OCT images in ophthalmology [39,40], breast tissue [41][42][43][44][45] and other human tissue [46,47]. However, what a model truly saw was still unknown until the investigation of the receptive field of convolution layers either by deconvolution or by masking techniques. In our study, we make use of CAM technique to indicate the discriminative image regions used by a CNN model for categorization. In other words, the detected image regions are highly likely to be the location of lesions. The same diagnosing strategy has been implemented on OCT images of ophthalmological diseases [48] like glaucoma [49,50] and age-related macular degenerationa [51][52][53] and is proved effective. We apply this technique for subgingival dental calculus detection and visualize the location of lesions. Threedimensional views of pure subgingival dental calculus are derived from OCT volumetric data through CAM, showing the potential of intuitive surgical guidance, and we call this technique "disease activation map (DAM)." We hope that with the help of our intelligent identification system, dentists can easily examine subgingival dental calculus from CAD image in real-time with high detection power.

| METHODS
Extracted human teeth with and without dental calculus were confirmed by an experienced dentist. Both bare teeth and subgingival samples were prepared sequentially with an established procedure and imaged by the OCT system, and the two phases of samples were used for ground truth validation and performance testing respectively. The measured data were divided into training, validation, and testing group for deep learning classification. The receiver operating characteristic (ROC) curves were provided for model performance evaluations. Finally, we applied CAM for dental calculus visualizations both in B-scans and in volumetric views and further created DAM and CAD images.

| Sample preparation
The study was approved by the Institutional Review Board (IRB) of Veteran General Hospital in Taipei (IRB no. 2014-04-013C). All samples including teeth with dental calculus and healthy ones were put in thymol saturated solution for preservation. We prepared two phases of samples, bare teeth and subgingival samples, for the designed experiments as shown in Figure 1. We used a periodontal curette to remove the residual soft tissue from teeth; the sample of the first phase preparation was complete and imaged by the OCT system to detect and record the location of dental calculus. Afterward, we referred to Xiaojun Zha's method [54], using silicone (anti-bacteria silicone, INITIAL TECHNOLOGY INC.) as the artificial gingiva to cover the teeth samples, and prepared the subgingival samples for the second experiment. The silicone gel was placed in a syringe and squirted onto the teeth samples in the prevention of the formation of air bubbles. We attached artificial skin (DuoDERM Extra Thin CGF Dressing, ConvaTec Inc.) onto the silicone gel F I G U R E 1 An example of (A) the bare tooth and (B) the subgingival samples. The dental calculus is circled by the yellowish dash line. The scale bar represents 1 cm to avoid strong reflection from the surface of silicone and to modify the thickness of the silicone in a range of 0.5 to 0.9 mm. The samples of the second phase were then ready for OCT measurement when the silicone became dry.

| Experimental setup
Our OCT system shown in Figure 2 is a single-mode fiber based Mach Zehnder interferometer. A high speed swept source laser (HSL-20-50, Santec Corp.) with an average output power of 15 mW was employed as our light source. The wavelength of the laser is centered at 1.31 μm with a full width at half maximum (FWHM) of 100 nm, leading to a theoretical longitudinal resolution of 8 μm. The A-line scanning rate is 50 kHz and triggered the data acquisition by the reflection from the fiber Bragg grating (FBG-SMF-1266-80-0.2-A-(2)60F/E, L = 1M, TATSUTA Electric Wire Cable Co., Ltd) detected by amplified photodetector (PDA05CF2, Thorlabs, Inc.). A laser built-in k-trigger was used for the calibration of kspace linearity.
The light from HSL was firstly divided by C1 coupler, and the 1% of the light was transmitted into the FBG, which was designed to cause an 80% reflection of the incoming light at the wavelength of 1266.0 nm. This reflected light was detected by the APD as the signal of A-trigger. On the other hand, the remaining 99% of the light was separated by C2 coupler and transmitted into two polarization insensitive optical circulators (PICIR-1214-12-L-05-NE, OF-Link Communications Co., Ltd), Cir1 and Cir2, with 20:80 ratio, and these two beams propagated toward the reference arm and the sample arm respectively. In the reference arm, the beam was collimated by a fiber collimator (F260APC-C, Thorlabs, Inc.) with an achromatic lens (AC254-030-C-ML, Thorlabs, Inc.) and reflected by a gold-coated mirror. As for the sample arm, the beam passed through galvanometers and samples in addition to the optical elements (FC, L2) identical to those in the reference arm, and it leaded to an approximate 18 μm lateral resolution in air. The beams from the reference arm and the sample arm were transmitted by the circulators and formed interference in C3 coupler. A balanced photodetector (PDB480-AC, Thorlabs, Inc.) was used to acquire interference signal with less contamination by noninterference signal, achieving 91.58 dB system sensitivity. Afterward, the captured electric signal was filtered by a high-pass filter (ZFHP-0R23-S+, Mini-Circuits International Inc.) and a low-pass filter (BLP-90+, Mini-Circuits International Inc.) with a designed bandwidth (0.23-81MHz) for signal conditioning. Eventually, by the waveform digitizer (ATS9350, Alazar Technologies Inc.), the interference signals were sampled, and the trigger signals were detected for synchronization.The function generator (AFG-2225, Good Will Instrument Co., Ltd), which controlled the two galvanometers (GVSM002, Thorlabs, Inc.), and the waveform digitizer were controlled by our LabVIEW program, achieving the two-dimensional scanning. The scanning range was set as 5 mm in both axes, and each C-scan consists of 1000 Â 1000 A-scans. We measured each sample volumetrically in four different scanning directions (vertical, horizontal, 45 , and 135 ) to mimic the real situation in clinical practice.

| Data processing
The acquired images were imported into our deep learning models for differentiation of dental calculus. First, since the distribution of the dental calculus is not even, each B-scans was labeled individually as normal or calculus according to the doctor's instruction. We divided the data into the training data, validation data and testing data according to which samples the data have been measured from to ensure the true prediction power in clinical practice. To differentiate dental calculus, CNNs based on VGG16 were implemented as our deep learning models, and the models for the bare teeth and the subgingival samples were trained separately.The CNNs used in our research were configured from Spyder using the TensorFlow application program interface for Python. All training was done on the high-performance windowsbased computer with 16.0 GB RAM, Intel(R) Core(TM) i5-7500 CPU at 3.40 GHz, and an NVIDIA GeForce GTX1660 GPU. A B-scan consisted of 400 (axial) Â 1000 (lateral) pixels originally, and we cropped and compressed OCT images into 103 Â 142 pixels as the input of the deep learning models. The model configuration is shown in Figure 3. The training was done in a batch size of 16 B-scans, and the RMSprop optimizer was used because of the small batches used. Binary cross-entropy was used to evaluate the performance of optimization, and the training process stopped when the difference between training data and validation data continually decreased over 10 times. The training results were presented in the form of ROC curves.
Finally, we visualized the lesion locations in B-scans through applying the CAM technique. Since the value of a CAM image itself contained the information of the detection confidence, which might be related to the size of lesions in the B-scan, by binarizing the CAM image according to the maximal value of itself and then multiplying it by the binarized intensity mask of its original Bscan image, we could produce an image that displayed dental calculus only, showing the possible site of dental calculus; we called it disease activation map (DAM) image. The major difference between CAM and DAM is the ability of disease identification. CAM is a technical tool designed for neural network architecture investigators and was proposed for visualization of the areas that a model focused on. Therefore CAM image is displayed in a pseudo-color scale throughout the whole image and has no ability of disease identification. On the other hand, DAM is designed for clinical use and provides the visualization of a disease location. The identified region was marked in a single color, which will be red in the following results, so as to provide clear and intuitive identification of a disease.

| RESULTS
Here we present the training results of the bare teeth experiment and the subgingival phantom experiment respectively. The visible image, the DAM projection image of the bare tooth, and the DAM projection image of the subgingival phantom from the same tooth were F I G U R E 3 The deep learning models based on VGG16 architecture used in this study F I G U R E 4 ROC curve of the CNN model trained in bare tooth experiment. The area under ROC curve (AUROC) is 0.973 compared, showing good consistency and good detection power. We collected 55 teeth in total, and 26 of them were diagnosed with dental calculus. All of the samples were prepared in the first phase, but only eight healthy teeth and eight teeth with dental calculus were prepared in the second phase because of the manufacturing difficulty.
Consequentially, there were some differences in the process of the two model training, and we are going to describe them in detail in the following sections.

| Bare teeth experiment
For the model training in the bare teeth experiment, we evenly selected 40 B-scans in one volumetric scan to compress the amount of data. Therefore, one sample contributed 160 B-scans, and there were totally 8800 frames selected. We divided the data into training data, validation data, and testing data with the ratio of 8:2:5. Both categories were distributed equally between splits.
The accuracy of the proposed method on the training and validation dataset achieved 99.17%, and the testing accuracy was also up to 95.06%. Figure 4 shows the receiver operating characteristics (ROC) curve of the testing dataset, and the area under ROC curve (AUROC) was 0.973. We show an example of B-scan, merged CAM, and CAD image from one dental calculus sample in Figure 5. As other research has reported, dental calculus appears as an irregular rough lump in B-scans, and the CAM can perfectly indicated the location of the dental calculus. By thresholding the value of CAM and modifying them according to intensity images, we can generate DAM and clearly identify the site of dental calculus. After combining the DAM with intensity image, we can produce the CAD image to assist dentists during removing dental calculus.
We can also stack the images to display the volumetric detection images. Figure 6 shows an example of the volumetric view from one sample with dental calculus. It is hard to observe the dental calculus barely from the C-scan image; however, with the help of our F I G U R E 5 An example of the result of the bare tooth experiment. Here shows (A) the B-scan image of dental calculus site, (B) merged with its corresponding CAM image and (C) the CAD image. The red area represents the dental calculus detected by our CNN model, the socalled DAM image F I G U R E 6 The (A) C-scan, (B) the volumetric DAM and (C) the volumetric CAD image example from a bare tooth. The x-axis represents the fast scanning direction, the y-axis represents the slow scanning direction and the z-axis represents the axial depth direction method, dental calculus can be easily identified and marked in color.

| Subgingival phantom experiment
For the model training in the subgingival phantom experiment, we made the full use of all B-scans, and therefore there were 32000 B-scans (four normal control and four with dental calculus) for training, 8000 B-scans (one normal control and one with dental calculus) for validation data, and 8000 B-scans (one normal and one with dental calculus) for testing. Another two volumetric data were prepared for volumetric view visualizations.
In this experiment, we obtained 94.97% accuracy in the training and validation dataset, and the testing accuracy also achieved 94.92%. The whole prediction process took 2.91 seconds to categorize a volumetric data (1000 frames). Figure 7 shows the ROC curve of the testing F I G U R E 7 ROC curve of the CNN model trained in subgingival phantom experiment. The AUROC is 0.997 F I G U R E 8 An example of the result of the subgingival phantom experiment. Here shows (A) the B-scan image of dental calculus site, (B) merged with its corresponding CAM image and (C) the CAD image. The red area represents the subgingival dental calculus detected by our CNN model F I G U R E 9 The (A) C-scan, (B) the volumetric DAM and (C) the volumetric CAD image example from a subgingival phantom. The xaxis represents the fast scanning direction, the y-axis represents the slow scanning direction, and the z-axis represents the axial depth direction data, and the AUROC was up to 0.997. Likewise, we show an example of B-scan, merged CAM, and CAD images from one subgingival dental calculus sample in Figure 8. The subgingival dental calculus beneath the silicone phantom can be easily seen in the B-scan image, and the merged CAM image also showed an overall good consistency with the site of dental calculus except at the marginal region, where CNN model usually performs worse. The CAD image can help us directly observe the underlying subgingival dental calculus. In consideration of clinical practice, a higher sensitivity is preferred since dental calculus should be completely removed to stop periodontitis from deteriorating while the false-positive occurrence is tolerable. Therefore, our model gave 100% sensitivity and 90.02% specificity and eventually yielded 95.01% accuracy, which is still good enough for dental applications.
Although we can obviously see the subgingival dental calculus in OCT B-scan image, we can barely recognize the suspicious subgingival dental calculus in OCT C-scan image shown in Figure 9A. With the help of DAM, the volumetric CAD image shown in Figure 9C successfully identified the site of the subgingival dental calculus from both the artificial gingiva and the tooth. It spent 313 seconds in total to produce a volumetric CAM/DAM/ CAD image (1000 frames) if without saving the images.

| Comparison
To verify the reliability of dental calculus detection, we compared the dentist's examination result with the DAM projection images generated by the two models. Figure 10 shows the comparison of an example. We imaged the sample at the site of dental calculus before and after the dental calculus was fully covered with the artificial gingiva. The size of the calculus was around 1 Â 3 mm and was covered in the scanning range of OCT C-scan. The dentist's examination result was circled by dashed green lines, in which the dental calculus was diagnosed. On the other hand, we marked the detected calculus in red for both the bare tooth model and the subgingival phantom model. The results showed that they had overall good consistency with the dentist's examination despite the fact that a few portions were mismatched due to the aforementioned border effect.

| DISCUSSIONS
The disease visualization is restricted by the intrinsic limitations of the algorithm of CAM and performs worse at the border of frames. In this case, we suggest enlarging the imaging area and using only the central part for accurate calculus localization. In the production of the DAM image, a constant threshold of the CAM value would lead to either a blurred detection at the periphery or a narrower detection at the center. By tuning the threshold of CAM according to the detection confidence, we successively delineate boundary of the lesion, but the processing procedure should be further investigated in the future.
In the present study, we demonstrate the feasibility of intelligent dental OCT for the calculus examination, and our proposed method can distinguish dental calculus below gingiva with high detection power. We think the choice of the wavelength at 1.31 μm is ideal for dental applications because of both the penetration depth and the resolution. Although there might be some difficulty for the application of dental OCT on the account of F I G U R E 1 0 The visible image of (A) an example. We show the corresponding detected area of DAM projection images generated by (B) the bare tooth model and (C) the subgingival phantom model. The green dashed lines represent the site with dental calculus examined by clinical dentists while the red area represent the detected calculus thicker gingiva, the current research result should be enough for many of the clinical cases [55]. Here, we select silicone as our artificial gingiva to preliminarily verify our idea of applying DAM, but there are still some differences from real human gingiva. For example, the layered structure of periodontal tissue is not mimicked by silicone. The attenuation and scattering coefficients of artificial gingiva are nearly the same whereas the true optical properties vary from person to person especially when gingiva is inflamed [56]. Nonetheless, the layered structure of real human gingiva is still distinctive from the characteristics of subgingival dental calculus, and thus we can positively expect that our model can work well in the real situation.
Most tissue classification tasks were primarily considered achieved by image segmentation. However, image segmentation requests prelabeling, which introduces the observer's bias at the very first step. Especially for dental calculus, the strong reflection of light scattering causes the halation, which makes the labeling more uncertain. In contrast, our proposed deep learning process design makes the use of CAM to automatically find the location of suspicious lesion sites. In comparison to other image segmentation approaches, we do not need to manually and pixelwisely label the samples but obtain more objective and rapid results. We think it is a suitable and clever way to facilitate deep learning not only in the dental calculus detection but also in other OCT applications.
From the aspect of personalized medicine in periodontology, it is crucial to directly deliver a tailored treatment and to reduce trial-and-error procedures during the removal of subgingival dental calculus [57]. In the present study, we demonstrated the ability of intelligent dental OCT for the identification of dental calculus, the biomarker of periodontitis and we believe this will facilitate the development of precision medicine in dentistry. In the future, we will manage to carry out the experiment in the clinical setting and test the accuracy as well as the reproducibility of our method. Unlike other techniques, our method can automatically detect the subgingival dental calculus and visualize them either in two-dimensional images or in volumetric views. We think this would help the clinicians to get used to this novel instrument to perform calculus removal and greatly improve the current clinical dilemma.

| CONCLUSIONS
In this study, we realize the idea of disease visualization, and clinicians can easily recognize the location of subgingival dental calculus from DAM and CAD images either two-dimensionally or volumetrically. To our knowledge, this is the first international quantitative assessment research of sub-gingival dental calculus. We also demonstrate the promising outcomes of our method with both high sensitivity and high specificity, and the detection process is thoroughly automatically done by the deep learning algorithm, which is more objective and time-saving. Eventually, we bridge the technical gap between an OCT technician and a clinician, providing a possible solution of intuitive assistance for dentists during the subgingival dental calculus removal.