A support vector machine-based algorithm to identify bisphosphonate-related osteonecrosis throughout the mandibular bone by using cone beam computerized tomography images

Abstract

Objective:

This study aimed to develop an algorithm to distinguish the patients with bisphosphonate-related osteonecrosis of the jaws (BRONJ) from healthy controls using CBCT images by evaluating both trabecular and cortical bone changes through the whole body of the mandibular bone.

Methods:

Patient data set was created from axial CBCT images of 7 BRONJ patients (28 slices) and 8 healthy controls (27 slices). The healthy bone of healthy controls, bone sclerosis of BRONJ patients, bone necrosis of BRONJ patients, and normal appearing bone of BRONJ patients (NBP) were labeled on CBCT images by three maxillofacial radiologists. Proposed algorithm had preparation and background cancellation, mandibular bone segmentation and centerline determination, spatial transformation of gray values, and classification steps.

Results:

Significant differences between the statistical moments (mean, variance, skewness, kurtosis, standard error, median, mode and coefficient of variance) of healthy and diseased (bone sclerosis and necrosis) groups were observed (p = 0.000, p < 0.05). Also, variations were noted between healthy controls and NBP of BRONJ patients (p = 0.000, p < 0.05).

The statistical moments were utilized to develop the algorithm which has resulted with accuracy of 0.999, sensitivity of 0.998, specificity of 0.998, precision of 1, recall of 0.998, AUC of 1, and F1 score of 0.999 in identification of BRONJ patients from healthy ones.

Conclusion:

The proposed algorithm differentiated the mandibular bones of the healthy and the BRONJ patients with high accuracy in the present test sample.

Introduction

Bisphosphonate-related osteonecrosis of the jaws (BRONJ) is a well-defined complication of bisphosphonate therapy. ^1,2 It is the end result of accumulation of nitrogen containing bisphosphonates in jawbones which show direct inflammatory and/or necrotic effects on soft tissues that are boosted by bacterial lipopolysaccharides. ^2,3 The BRONJ staging system recommended by American Association of Oral and Maxillofacial Surgery ^1,4 is mainly based on clinical manifestations and presents subtle radiographic changes until Stage 3. ⁴ However, radiologic evaluation is essential for the determination of the extent of BRONJ lesion and outcome of its’ treatment. It has been stated that 50% of Stage 0 BRONJ patients who present with non-specific complaints but with bone abnormalities may progress to clinical bone exposure. ⁴ Therefore, the radiographic assessment can play a pivotal role in prompt and correct diagnosis and management of the patients with BRONJ.

Conventional radiography and CBCT scans have been widely used as imaging modalities to provide precise evaluation of osseous changes such as periosteal reaction, sclerotic lesions, lucency and sequester formations. There are several reports on the quantitative evaluation of osseous changes of bisphosphonate medicated patients by CBCT. ^5–11 Patient- and medication-dependent parameters lead to great variability of the radiological findings of BRONJ on radiographic images. ^6,9,11,12 In addition to these parameters, technical shortcomings of CBCT play an important role. Inaccuracies of gray values on CBCT images due to the variabilities in the axial plane leading to cupping and doming artifacts and beam hardening, unevenness between axial slices because of X-ray divergence and differing mass per slice, and high image noise have significant effects on the correct interpretation of images. ^7,13 Additionally, the device characteristics, imaging parameters and patients positioning have been presented as the other factors that limit the use of pixel intensity values in CBCT images to assess the bone density. ^14,15 The acknowledged problems in gray value/pixel intensity and/or fractal dimension/texture analysis investigations which were performed on CBCT images have necessitated the development of further algorithms to establish the bone quality, because to the authors’ knowledge, no distinct clinicoradiological systematization of BRONJ has been stated yet. Moreover, bone quality assessment methods on CBCT images of patients mostly used single or multiple region of interest (ROI) and have inferred a decision about the bone. ^7,9,11,16 However, mechanical properties and architecture of trabecular bone may vary depending on its physiological function and mechanical loading on the skeleton. ^17–19 Thus, instead of examining fractions of the bone, a methodology which uses an algorithm to assess the whole body of the bone would hypothetically provide more information about the bone quality.

Within this context, artificial intelligence (AI) which is machine-based cognitive system that mimics human brain to perform tasks such as learning and problem-solving based on computer algorithms presents new solutions in health sciences. ²⁰ Machine learning (ML) and classification algorithms have not only contributed to the value of diagnostic imaging by increasing the image quality, reducing the image acquisition duration and improving medical communication, but also aided the clinicians to improve disease and pathology diagnosis, and treatment planning. ^7,21–25 Several approaches such as shallow or deep neural networks, fuzzy logic-based systems, expert systems, optimization algorithm-enhanced classifiers, support vector machine, k-means neighbor classifiers can be used for diagnostic purposes in ML. ²⁰ Among software learning algorithms used for intelligent learning approaches, supervised learning is applied to learn to detect normal and abnormal characteristics by analyzing handlabeled images via implementing a data set for training. These programs either classify the data as disease or no disease, or present the output variable as a real or continuous value. ²⁰

Up to date, no specific imaging features of BRONJ have been decisively established for radiological diagnosis of the patients. ²⁶ Previous studies on imaging of BRONJ have reported contradictory results, some stating that initial bone change would be osteosclerosis, ^12,27 whereas others revealed that sclerotic alterations may be increased being parallel to the severity of BRONJ, ^7,28 and transformation of sclerotic areas to osteolytic regions may be observed, as well. ¹¹ Osteolysis and sequestrum formation have been considered as the radiographic features of advanced stages (Stage 3). ^12,26 Some authors ^2,27,28 also reported that alterations may be observed at the cortical margins due to the changing bone mineral density. All these criteria are observer-dependent, subjective subtle changes ⁷ which may inevitably lead to variations among the decisions of the radiologists. In this context, AI appears as an adjunct to assist the decision-making process.

Therefore, the aim of this study which was implemented on CBCT images of BRONJ patients was to develop an algorithm in order to distinguish BRONJ patients from healthy controls radiographically by evaluating sclerotic and/or lytic changes both in trabecular and cortical bone through the whole body of the mandibular bone.

Methods and materials

Data set

Test group and healthy controls

A total of seven individuals with a history of bisphosphate use for their bone metabolic diseases and oncologic treatment who were referred to the Department of Oral and Maxillofacial Radiology, Faculty of Dentistry, Ege University for CBCT evaluation for diagnosis and management of various dental problems were enrolled into the present study. The clinical evaluation and radiographic examination of the CBCT images of these patients were consistent with those of Stage 0 BRONJ ^2,4 and thus, these bisphosphonate users constituted the test group. On the other hand, the control group consisted of eight systemically healthy individuals who were referred to the same facility for the same purpose. The study protocol was approved by the Ethics Committee of the Medical Faculty of Ege University (Approval #20–9.1T/4) and was conducted according to the Declaration of Helsinki on experimentation involving humans. The patients’ sex, age, racial group, type of occlusion and skeletal pattern were not considered for enrollment to the study.

The inclusion criteria were set as:

• The patients’ consent to be included into the study.

• Absence of any contraindication for CBCT imaging,

• For the bisphosphonate users, definite clinical and radiological diagnosis of Stage 0 BRONJ according to the American Association of Oral and Maxillofacial Surgeons (AAOMS). ⁴

American Association of Oral and Maxillofacial Surgeons has declared the radiological findings of the Stage 0 BRONJ patients as ⁴

• Alveolar bone loss or resorption without chronic periodontal disease ⁴

• Trabecular pattern changes such as dense bone and unremodeled bone in alveolar sockets. ⁴

• Osteosclerotic areas within the alveolar bone and/or the surrounding basilar bone ⁴

• Periodontal ligament thickening and periodontal ligament space narrowing. ^1,2,4

• Thickening of the lamina dura ^2,4

• Inferior alveolar canal narrowing ²

Radiographic technique

CBCT examinations were performed using the Kodak 9000 3D (Kodak Carestream Health, Trophy, France) system and the imaging parameters were 10 mA and 70 kVp with 2.5 mm Al equivalent filtration. CBCT acquisition of each patient was completed after a single 360° rotation with 10.8 s scan time, and a volume with a spatial resolution of 76 µm (isotropic voxel) was reconstructed. All images were taken by the same operator and the Digital Imaging and Communications in Medicine (DICOM) files of the CBCT images were saved to a portable hard disk for evaluation.

Data set construction and labeling of data set

Data set was created from CBCT images of 7 BRONJ patients with total 28 slice images and eight healthy patients with total 27 slice images. All images which were provided in DICOM format were opened and displayed for the labeling step in MATLAB Software, as shown in the study flowchart (Figure 1).

Figure 1.

Flowchart of the study presenting the steps in algorithm development using CBCT images of BRONJ patients and healthy controls. BRONJ, bisphosphonate-related osteonecrosis of the jaw.

The axial CBCT images were selected as the working environment since this plane provided the whole image of the mandibular bone. In BRONJ patients, the margin of the cortical bone towards the inner trabecular bone was obscure when compared to the healthy patients. This finding was validated on each CBCT image by assessment of the gray values and healthy controls, and different outputs were calculated (Figure 2).

Figure 2.

The statistical characteristics of the gray values of lines resulted with different bone profiles in BRONJ patients and healthy controls. BRONJ, bisphosphonate-related osteonecrosis of the jaw.

Thus, the cortical margin alterations which are the consequence of the changing bone mineral density ^27,28 were considered the basic concept of the algorithm to distinguish the healthy and diseased bone.

In the next phase of the study to create the data set, a panel of three oral and maxillofacial radiologists with >15 years of experience in CBCT imaging selected four regions on the CBCT images of the BRONJ patients and controls with consensus decision making: (1) healthy bone of healthy controls, (2) bone sclerosis of the BRONJ patients, (3) bone necrosis of the BRONJ patients, (4) normal appearing bone of the BRONJ patients.

The alveolar bone was considered “normal” when regular trabecular pattern with optimal density was observed and radiopaque lingual and buccal cortical bone margins were clearly differentiated from the alveolar bone. The bone with generalized or local density increase was grouped as “sclerotic”, and in such cases, the trabecular bone would not be easily distinguished from the cortical bone. On the contrary, the areas of trabecular rarefaction or decreased density were considered as “necrotic” areas, and these sites would be easily distinguished from the cortical bone. This categorization was utilized both in BRONJ and healthy control groups. Thus, a “gold-standard” was achieved and was used in order to test the correspondence of the results of the algorithm with those of the standard (Figure 3).

Figure 3.

The bone necrosis and/or bone sclerosis on the CBCT images of the BRONJ patients were labeled as “sclerotic” and/or “necrotic”. Also, healthy appearing areas were also marked on the images as “healthy appearing areas”. The mandibular bones of the patients without any bone disease were labeled as “healthy area” on the CBCT images of the control group. BRONJ, bisphosphonate-related osteonecrosis of the jaw.

Algorithm design and implementation

The proposed algorithm had to go through four major stages: (1) preparation and background cancellation of DICOM file, (2) mandibular bone segmentation and centerline determination, (3) spatial transformation of gray values, (4) classification.

Preparation and background cancellation of DICOM files

The background noise and the area outside of the field of view (FOV) of CBCT device were equalized by thresholding. The thresholding level was selected as the median gray value of the background noise in the image. After median thresholding of CBCT images, a mean gray value-based thresholding step was applied for further discrimination of the background and mandibular bone (Figure 4).

Figure 4.

Background Cancellation of DICOM Images (a-Original DICOM, b-Processes DICOM, c-Mandibular Bone Segmentation, d-Trabecular Bone Path). DICOM, Digital Imaging and Communications in Medicine.

Mandibular bone segmentation and trabecularbone path

In order to define the image regions corresponding to trabecular bone, the mandibular bone should be carefully selected. For this purpose, an adaptive threshold-based segmentation was used and bone was segmented on every CBCT image. ²⁹ After thresholding, morphological manipulations (closing and area opening) were applied to get the largest connected area in the binary image. This largest area corresponded to the mandibular bone region(Figure 4c).

Following the extraction of the bone region, trabecular bone path which is an imaginary line that lies along the centerline of the bone was defined to flatten the image. Skeleton strength map (SSM) of the image was generated by Euclidian distance transform (EDT) and was used to mark trabecular bone path on the bone. Trabecular bone path points were calculated by thresholding the amplitude of directional gradient of EDT of the image (Figure 4d).

Before drawing lines perpendicular to the trabecular bone path, a curve fitting step was added to the algorithm to smooth the noisy nature of the extracted trabecular bone path. A third-order polynomial was selected because of the parabolic shape of mandibular arch, and was fitted to the resulted trabecular bone path points. First derivative of this third-order polynomial was used to find the slope of the perpendicular lines to the polynomial curve as the first derivative of a function defines the slopes of tangent lines attached to the curve. After the calculation of the slopes, all points on the perpendicular lines were calculated by trigonometrical relationships on a selected line length of 150 pixels (Equation 1) where x and y are the co-ordinates of points on the perpendicular line, α slope of the third-order polynomial, r and c are the co-ordinates of points on the third-order polynomial.

$$\begin{gathered} {\displaystyle x\left(i\right)=r\left(i\right)+l\ast -\sin \alpha \\ y\left(i\right)=c\left(i\right)+l\ast \cos \alpha } \end{gathered}$$ (1)

After the calculation of these points, curved path of mandibular arch was flattened into orthogonal geometry. This step can be considered as “straightening of a curve” as it is shown in Figure 5.

Figure 5.

Flattened mandibular arch with perpendicular lines to trabecular bone path

Feature extraction

Statistical features of spatial gray value data were used as the input of classification algorithm. The mean number of perpendicular lines on the trabecular bone path per image slice was 316.8727 (± 57.78079). After the perpendicular lines were drawn on each point on the centerline of the mandibular bone, the statistical moments (the mean, variance, skewness, kurtosis, standard error, median, mode and coefficient of variance) of the gray values along the line were calculated. During the analyses, each point on the perpendicular line was considered as an individual data point. Statistical features were recorded in a separate database file for corresponding 17,117 data points (the healthy bone in controls: 7853 data, bone sclerosis of BRONJ patients: 3245 data, bone necrosis of BRONJ patients: 2812 data, normal appearing bone of the BRONJ patients (NBP: 3207). Data set was constructed from the samples of both patients and controls. The descriptive statistical analyses were performed and statistical moments were determined. Statistical analyses were performed by using Independent Sample Test for pair-wise comparisons, with SPSS v. 25.0 statistical software (SPSS Inc., Chicago, IL). In all tests, p was considered as 0.05. After the validation of the data set by Independent Sample Tests, ANOVA was performed for the determination of the most significant features that have the most impact on the classification process. Importance scores of features were calculated by the negative logarithm of p values of ANOVA.

Classification

Support vector machines (SVMs) algorithm was used for the classification of data according to the statistical moments. Data set (the healthy bone in controls: 6057 data, bone sclerosis of BRONJ patients: 3245 data, bone necrosis of BRONJ patients: 2812 data)with 12,114 points was separated into three parts, named train, validation and test with ratios of 70, 15 and 15%, respectively. As a result of the statistical comparison of control and NBP groups, a significant statistical difference was detected. Thus, normal appearing bone of the BRONJ patients (NBP) was considered as different than control group and was not included in the training data set (Figure 6).

Figure 6.

The patients whose images included the bone necrosis and/or bone sclerosis were considered as “diseased”, but the images without any bone necrosis and/or bone sclerosis were stated as “healthy”. MRONJ, medication-related osteonecrosis of the jaw.

The diagnostic ability of the designed algorithm was determined by dividing the whole data either as “diseased” or “healthy”. The diseased group was created by joining the bone necrosis and bone sclerosis groups, whereas the healthy area was considered as the healthy group.

SVM classification algorithm with Gaussian kernelwas trained. For training and implementation of SVM, MATLAB r2020b (MathWorks inc., MA) was utilized.

Algorithm evaluation

To evaluate the performance of the algorithm, true positive (TP), false positive (FP), true negative (TN), false negative (FN), accuracy, specificity, precision, recall, F-measure, and the area under curve ROC (AUC) values were established. ²⁹

True positive-TP indicated a diseased mandibular bone that was classified correctly as diseased. False negative-FN represented incorrect classification of diseased mandibular bone as healthy.

Incorrect prediction of healthy mandibular bone as diseased bone was considered false positive-FP. Correct classification of a healthy mandibular bone as healthy bone represented true negative-TN. ²⁹ The evaluation indices were defined as:

Accuracy was the proportion of the mandibular bones that were correctly labeled among the total number of mandibular bones [Accuracy = (TP + TN)/(TP + TN + FP + FN)].

Specificity was defined as the proportion of the mandibular bones that were correctly identified [Specificity = TN/(TN + FP)].

Precision was the proportion of the mandibles that were correctly predicted as diseased (TP) among those labeled as diseased (TP+FP) [Precision = TP/(TP+FP)].

Recall (Sensitivity) was the proportion of the diseased mandibles (TP+FN) that were correctly identified (TP) [Recall = TP/(TP+FN)]

F-measure is the value that combines both precision and recall [F-measure = (2 x Precision x Recall)/ (Precision+Recall)].

Plotting the TP rate against the fFP rate at different thresholds created the receiver operating characteristic (ROC) curve that presented the diagnostic ability of a binary classifier system. The AUC measured the ability of a classifier to distinguish classes, and was used as a summary of the ROC curve. Since the AUC ranged from 0 to 1, the higher AUC value revealed better performance of the model at distinguishing between the positive and negative classes.

Results

Relevance of the statistical moments in identification of the groups

The statistical moments (the mean, variance, skewness, kurtosis, standard error, median, mode and coefficient of variance) of data groups which were separated into four groups as healthy bone in controls, bone sclerosis of BRONJ patients, bone necrosis of BRONJ patients, and NBP were assessed.The statistical analyses revealed that there was a considerable difference between means of healthy and NBP groups, and the largest coefficient of variation was calculated within the NBP group (Table 1). Thus, Independent Sample Test was applied in order to establish the significance of the differences between these two groups, and the results showed that there was a significant difference between healthy and NBP groups, which indicated that the mean, variance, skewness, kurtosis, standard error, median, mode and coefficient of variance of gray values could be utilized to discriminate the healthy bones and normal appearing bone of the BRONJ patients (p = 0.000, p < 0.05) (Table 1).

Table 1.

Descriptive statistics of datasets

	Mean	Variance	Skewness	Kurtosis	Standard error	Median	Mode	Coefficient of variance
Healthy	137.5	1311.5	1.1724	3.304	2.516	122.83	106.25	0.2610
Bone necrosis in BRONJ	117.76	1160.51	1.6666	5.279	2.6958	107.0	91.47	0.28344
Bone sclerosis in BRONJ	131.4	1510.10	1.0526	2.899	3.1297	113.70	98.64	0.29379
Normal appearing bone in BRONJ	115.3	1522.85	1.4758	4.345	3.1332	103.13	82.96	0.33422

BRONJ, bisphosphonate-related osteonecrosis of the jaw.

The data of bone sclerosis and bone necrosis groups were combined and reorganized as “diseased”. When the differences between healthy and the diseased groups (including the bone sclerosis and bone necrosis groups) were investigated, the results disclosed statistical differences between all groups (p < 0.05). The mean, variance, skewness, kurtosis, standard error, median, mode and coefficient of variance belonging to healthy and diseased groups were statistically distinct from each other (p = 0.000, p < 0.05).

Classification of the groups as “healthy” and “diseased” by the algorithm

The data of both the healthy and the diseased groups were entered into the novel algorithm.

Three different SVM classifiers were trained with different numbers of feature vector size. The results of using four features (mean, variance, skewness, kurtosis) indicated the accuracy of 0.898, sensitivity of 0.906, specificity of 0.889, AUC of 0.910 and F1 of 0.899 in differentiating the healthy and the diseased bone (Table 2).

Table 2.

Performances of three SVM classifier

Case	TP	FP	TN	FN	Accuracy	Sensitivity	Specificity	Precision	Recall	AUC	F1
Four features	827	100	803	86	0.898 (0.884–0.912)	0.906 (0.887–0.925)	0.889 (0.869–0.910)	0.892 (0.872–0.912)	0.906 (0.887–0.925)	0.910	0.899
Three features (MIF)a	846	35	913	22	0.969 (0.961–0.977)	0.975 (0.964–0.985)	0.963 (0.975–0.952)	0.960 (0.947–0.973)	0.975 (0.964–0.985)	0.975	0.967
Eight features	945	0	870	1	0.999 (0.998–1)	0.998 (0.997–1)	1 (1)	1 (1)	1 (0.997–1)	1	0.999

AUC, area under the curve; SVM, support vector machine.

The comparison of the importance scores showed that the three features (mean, median and mode) possessed the highest impact on the classification between the healthy and the diseased bone (Figure 7).

^aMost important features: mean, median, mode), TP, true positive; FP, false positive; TN, true negative; FN, false negative.

Figure 7.

Importance scores of features revealing that mean, median and mode had the highest impact on the classification of the diseased and healthy bones. CoV, coefficient of variation.

Using these three features, accuracy with 0.969, sensitivity with 0.975, specificity with 0.963, AUC with 0.995 and F1 with 0.967 were achieved and this SVM with three features showed better results with respect to that of other four features (Table 2).

On the other hand, using eight features (mean, variance, skewness, kurtosis, standard error, median, mode and coefficient of variance) to discriminate the healthy and the diseased bone had resulted with higher accuracy with 0.999, sensitivity with 0.998, specificity with 0.998, precision with 1, recall with 0.998, AUC with 1, andF1 score with 0.999 (Table 2).

Discussion

The present study presents a novel algorithm that detects the bone density value changes throughout the whole mandibular bone including both the cortical and trabecular areas, rather than selecting ROIs on the images to analyze. The cortical margin alterations in BRONJ patients which were previously reported as the result of the changing bone mineral density ^27,28 were also noted in most of the BRONJ patients who were enrolled into the present study. Even though this feature was not stated among the radiological findings of the BRONJ patients, it was observed that the transition between the mandibular cortical and trabecular bone was not as distinct as it was in healthy people in some areas of the BRONJ patients. Thus, in order to include this finding into the analyses, both inner and outer cortical bones were added to the trabecular areas to be evaluated and the whole body of the mandible was assessed as the region of interest. With this approach, it was also possible to eliminate the probability of bias in selection of the ROI to measure the bone, and the variations through the bone structure due to physiological/mechanical forces that influence the bone construction.

The bone area covered in FOV of CBCT images and location of the bone change considerably due to the nature of anatomical and functional dimensional differences of mandibular bone individually. Considering that this change has a deteriorating effect in the subsequent image processing steps, an initial stage was developed for the background elimination and regularization of CBCT images prior to application of the algorithm.

In the present study, the novel algorithm differentiated the mandibular bones of the healthy and the BRONJ patients with high accuracy (F1 = 0.999). The other unexpected finding was that the healthy bone and the normal appearing bone of BRONJ patients presented different gray value distribution even though they appear similar on the radiographs. The limited capability of human eye to detect subtle changes on the radiographs ³⁰ may be the reason of the radiographic similarity of the mandibular bones of healthy people and the normal appearing mandibular bones of the BRONJ patients, while they presented significantly distinct statistical moments of gray values. Although the parameters such as the technical variations of CBCT devices and the exposure conditions may contribute to the gray value changes of the bone assessed on radiographic images, the possibility of detecting actual bone changes with software may be recognized, and close dental follow-up of those patients may be recommended.

In dental literature, depending upon the task and the radiographic images utilized, the convolutional neural network (CNN) presented AUC values ranging between 76.7% ³¹ and 0.99, ³² accuracy between 0.75 ³³ and 0.970, ³² and F1 value between 0.78 ²² and 0.93. ^{21–24,31–36} Regarding the estimation of the bone mass and detection of the bone changes, the performance of machine learning on three perpendicular planar CBCT images revealed an excellent correlation (>0.81) between manual and automatic measurements, and automatic method distinguished low bone mass in the mandibular cortex. ³⁵ Lee et al had successive studies investigating the efficacy of CNN in osteoporosis diagnosis on panoramic radiographs. ^36,37 The CNN system had high agreement with oral and maxillofacial radiologists in detecting osteoporosis with 99.91 AUC measure. ³⁶ Later, they have reported that transfer learning and fine-tuning had enhanced the efficacy of CNN for screening osteoporosis in panoramic images and showed an overall AUC of 0.858. ³⁷

For BRONJ patients, CNN provided exciting results. Similar to the present study, Guggenberger et al and Zhou et al have used axial images of the CBCT to detect the sclerotic and/or lytic areas of BRONJ. ^7,11 Guggenberger et al reported that quantitative measurements and volume-based evaluations on CBCT images have not performed high diagnostic accuracy as the qualitative image parameters and ROI-based measurements. The AUC for mean bone density value of ROI-based measurements was 0.83, with a sensitivity of 83% and specificity of 77% for diagnosis of BRONJ. ⁷ This AUC value was lower than ours, which was calculated as 1.0. In our study, the whole body of the mandible was assessed as the ROI, and both inner and outer cortical bones were added to the trabecular areas to be evaluated. With this approach, it was also possible to eliminate the probability of bias in selection of the ROI to measure the bone, and the variations through the bone structure due to physiological/mechanical forces that influence the bone construction. On the other hand, Guggenberger et al (2014) have divided each mandible into 10 different regions as condylar process (head and neck), coronoid process, ascending ramus, posterior mandible and anterior mandible. They have manually selected 10 mm² ROIs within the center of each segmented area on the CBCT images.The results revealed that mean bone density value increased from the condylar process to the anterior hemimandible in healthy individuals, because of the variations through the bone structure due to physiological/mechanical forces that influence the bone. ⁷ Further, considering that posterior mandible contained 82% of the lesions, the authors calculated the AUC values of these areas separately. AUC from ROC analyses increased considerably with improved sensitivity and specificity, ⁷ and became closer to our AUC result. The cause of this alteration may be the similarity of the areas to be analyzed in both studies, even though Guggenberger et al have not included the cortical bones of the mandible. ⁷

In ML, artificial neural network, logistic regression, decision tree, random forest and SVM are the methods that may be used for classification of data. ³⁸ Among those, SVM algorithm which is also utilized in the present study, uses the features of the test samples to arrange them in the appropriate locations in an imaginary high-dimensional space which is also created by SVM a priori. Then, a hyperplane is used to separate the samples in order to provide the data classification. ³⁸

In the literature, studies have utilized either CBCT images or conventional intra- and extraoral radiographs for diagnosing bone alterations and thus, the outputs varied consequently due to the characteristics of radiographic techniques. It’s stated that for timely and adequate detection of radiographic changes of the bones of BRONJ patients, 3D radiographic imaging shall be practiced. ³⁹ Recently, the clinicians presented higher AUC value of 0.88 with CBCT than that of panoramic radiography (AUC: 0.562), and superiority of CBCT imaging to panoramic radiography for detection of histologically confirmed non-vital bone has been affirmed. ⁴⁰ Thus, CBCT images of BRONJ patients and the controls were enrolled into the present study in order to detect the diseased bone regions without exposure and with non-specific symptoms.

The limitations of the study such as the small number of the sample size and the lack of histological diagnoses as the gold-standard shall not be overlooked. Also, a large data set with random selected test subset is used in the present study in order to overcome the probability of overfitting of the model, but overfeeding is still a concern in such study designs. ³² In future studies, other ML diagnostic algorithms (i.e. shallow or deep neural networks, fuzzy logic-based systems, expert systems, optimization algorithm enhanced classifiers, k-means neighbor classifiers) may be implemented in addition to the presented novel algorithm so that the efficacy of the methods can be compared by utilizing the AUC.

In conclusion, a meticulous clinical and radiographic screening of the patients under anti-resorptive therapy needs to be the major strategy to perform an adequate dental preventive program. Using ML to improve the detection of the bone mineral changes within the bone prior to later stages of BRONJ would be an important phase of this procedure, and accordingly aids to maximize the benefits of anti-resorptive treatment by assisting to decrease the complications.