- Email: [email protected]
Computer-aided diagnosis of periapical cyst and keratocystic odontogenic tumor on cone beam computed tomography
Accepted Manuscript
Computer-aided diagnosis of periapical cyst and keratocystic odontogenic tumor on cone beam computed tomography E. Yilmaz , T. Kayikcioglu , S. Kayipmaz PII: DOI: Reference:
S0169-2607(16)30429-1 10.1016/j.cmpb.2017.05.012 COMM 4427
To appear in:
Computer Methods and Programs in Biomedicine
Received date: Revised date: Accepted date:
2 May 2016 15 April 2017 26 May 2017
Please cite this article as: E. Yilmaz , T. Kayikcioglu , S. Kayipmaz , Computer-aided diagnosis of periapical cyst and keratocystic odontogenic tumor on cone beam computed tomography, Computer Methods and Programs in Biomedicine (2017), doi: 10.1016/j.cmpb.2017.05.012
This is a PDF file of an unedited manuscript that has been accepted for publication. As a service to our customers we are providing this early version of the manuscript. The manuscript will undergo copyediting, typesetting, and review of the resulting proof before it is published in its final form. Please note that during the production process errors may be discovered which could affect the content, and all legal disclaimers that apply to the journal pertain.
ACCEPTED MANUSCRIPT
Highlights
• • • •
The database used in this study is consisted of 50 different dental CBCT scans. Dental periapical cyst and keratocystic odontogenic tumor lesions were classified. Six different classifiers were used for the classification experiments. The Support Vector Machine (SVM) achieved the best classification performances.
AC
CE
PT
ED
M
AN US
CR IP T
•
A decision support system for classification of dental lesions on CBCT is proposed.
ACCEPTED MANUSCRIPT
Computer-aided
diagnosis
of
periapical
cyst
and
keratocystic odontogenic tumor on cone beam computed
E. Yilmaz1,*, T. Kayikcioglu2, S. Kayipmaz3
1
Department of Computer Engineering, Karadeniz Technical University, 61080, Trabzon,
AN US
Turkey 2
Department of Electrical and Electronics Engineering, Karadeniz Technical University, 61080,
Trabzon, Turkey 3
CR IP T
tomography
Department of Oral Diagnosis and Radiology, Karadeniz Technical University Faculty of
M
Dentistry, 61080, Trabzon, Turkey
PT
Abstract
ED
*Corresponding author; [email protected]
CE
Background and objectives: In this article, we propose a decision support system for effective classification of dental periapical cyst and keratocystic odontogenic tumor (KCOT) lesions obtained via
AC
cone beam computed tomography (CBCT). CBCT has been effectively used in recent years for diagnosing dental pathologies and determining their boundaries and content. Unlike other imaging techniques, CBCT provides detailed and distinctive information about the pathologies by enabling a three-dimensional (3D) image of the region to be displayed. Methods: We employed 50 CBCT 3D image dataset files as the full dataset of our study. These datasets were identified by experts as periapical cyst and KCOT lesions according to the clinical, radiographic and
ACCEPTED MANUSCRIPT
histopathologic features. Segmentation operations were performed on the CBCT images using viewer software that we developed. Using the tools of this software, we marked the lesional volume of interest and calculated and applied the order statistics and 3D gray-level co-occurrence matrix for each CBCT dataset. A feature vector of the lesional region, including 636 different feature items, was created from those statistics. Six classifiers were used for the classification experiments.
CR IP T
Results: The Support Vector Machine (SVM) classifier achieved the best classification performance with 100% accuracy, and 100% F-score (F1) scores as a result of the experiments in which a ten-fold cross validation method was used with a forward feature selection algorithm. SVM achieved the best classification performance with 96.00% accuracy, and 96.00% F1 scores in the experiments in which a
AN US
split sample validation method was used with a forward feature selection algorithm. SVM additionally achieved the best performance of 94.00% accuracy, and 93.88% F1 in which a leave-one-out (LOOCV) method was used with a forward feature selection algorithm.
Conclusions: Based on the results, we determined that periapical cyst and KCOT lesions can be
M
classified with a high accuracy with the models that we built using the new dataset selected for this study. The studies mentioned in this article, along with the selected 3D dataset, 3D statistics calculated from the
ED
dataset, and performance results of the different classifiers, comprise an important contribution to the
Keywords
PT
field of computer-aided diagnosis of dental apical lesions.
CE
Computer aided diagnosis; Dental apical lesion; Classifier; Cone beam computed tomography; Periapical
AC
cyst and keratocystic odontogenic tumor; Volumetric textural features; Dental image dataset
1. Introduction Oral diagnostic radiology comprises a sub-branch of dentistry relating to information needed for treatment planning for dental diseases. Oral examination may not always be sufficient for the diagnosis of dental disease. Therefore, during the dental examination, radiological imaging methods are utilized to assist in the examination of invisible intrabony regions. Dental images
ACCEPTED MANUSCRIPT
can be obtained via periapical radiography, panoramic radiography, and cone beam computed tomography (CBCT) [1]. Dental regions containing one or more tooth roots can be viewed using periapical radiographs. Panoramic radiographs are used to obtain a two-dimensional (2D) panoramic view of the upper and lower jaw. Images are acquired with the help of conventional X-rays in both
CR IP T
periapical and panoramic devices [2].
Dental CBCT technology has been used since the late 1990s. Three-dimensional (3D) images of the neck and chin area can be obtained using this technology [3]. CBCT devices are useful in implant planning [4], teeth segmentation [5], jaw tissue segmentation [6], detection of
AN US
facial traumas of the jaw [7], and obtaining 3D images of lesions and other pathologies [8].
Pathological lesions around the tooth root can be identified based on their size and location using volumetric images obtained with these devices. The data obtained by CBCT can sometimes indicate a lesion or abnormal anatomical findings. Radiology experts typically
M
examine those images along axial, sagittal, and coronal planes and can thereby readily make determinations about the findings [9,10].
ED
Detected abnormalities are often expressed as endodontic or periapical lesions [11,12]. During radiological examinations, it is important for the doctor to understand the characteristics
PT
of detected lesions in a timely and appropriate manner. According to current studies, volumetric
CE
measurements on a CBCT scan are an accurate indicator of the volume of the actual periapical lesion in the jawbone [13,14] . As a result, the planned treatment can be effective in reversing
AC
the destruction of the pathology, which supports the importance of CBCT and other radiological examination methods [2]. Lesions detected as a result of scanning are evaluated with consideration of some of their
properties, such as location, periphery, and internal structures in the surrounding area [3]. Some lesions have the potential to develop into cancer. Early diagnosis of such lesions is crucial to preventing disease progression [1].
ACCEPTED MANUSCRIPT
In recent years, many studies have been performed in the areas of medical imaging and signal processing. A majority of these studies address the classification of medical information. In each study, a specific method is preferred depending on the characteristic of the disease of interest. Classification methods are used to obtain meaningful results from medical information. Feature extraction processes are applied to that medical information for classification
CR IP T
experiments. In accordance with the content of the medical information, various features are used in each study. The power spectral density technique, for example, may be useful for extracting features from the obtained signal data in research conducted on a brain computer interface [15]. To establish a good classifier in a study performed on a retinal image database,
AN US
features such as shape, texture, and intensity may be suitable for determining the image quality [16].
Lesion classification is a sub-problem in the field of computer vision. Dental structures located in the head and neck area do not have a specific form. Therefore, research conducted on
M
dental images in the field of computer vision is extremely challenging. In the present study, we employ textural features to address the formal diversity observed in dental images. Unlike
ED
research in medical image processing in other disciplines, few published studies exist that address computer-aided detection and classification of apical lesions [17–21].
PT
In this article, we discuss classification of periapical cyst and keratocystic odontogenic tumor
CE
(KCOT) lesions observed in 50 different dental CBCT datasets. These datasets were identified by experts as periapical cysts and KCOT lesions according to the clinical, radiographic, and
AC
histopathologic features. KCOT was previously referred to as the odontogenic keratocyst (OKC). It was reclassified as the keratocystic odontogenic tumor by the World Health Organization (WHO) in 2005 [22,23]. The KCOT lesion is known for its aggressive behavior and high recurrence rate [24,25]. It is therefore necessary to especially focus on differentiating KCOT from cystic lesions. The aim of this study was thus to develop a CBCT classifier method for exact differentiation of KCOT from other cystic lesions.
ACCEPTED MANUSCRIPT
2. Materials and methods Unlike previous research, we used a new 3D dental image dataset. Accordingly, we further extracted a new feature vector derived from this dataset. In our experiments, lesions were classified into two groups using ten-fold cross validation, split sample validation and leave-one-out cross validation (LOOCV) methods. Fig.1 presents a flow diagram of the
CR IP T
operations performed to identify the most effective model for correctly distinguishing periapical cyst and KCOT lesion [22] images obtained via CBCT. 2.1. Database
The database used in this study consisted of a dental CBCT dataset obtained from 50 patients
AN US
between the years of 2013 and 2016 who attended Karadeniz Technical University, Faculty of Dentistry, Department of Oral Diagnosis and Radiology Clinic for routine controls. The images were obtained with a KODAK K9500 Trophy CBCT device. Data were recorded in 3D dataset
M
files using the Digital Imaging and Communications in Medicine (DICOM) standard. Medical image information comprising the 3D dataset file was converted to the appropriate data range in Hounsfield
scale
[26].
The
dataset
of
our
study
is
accessible
at
ED
the
http://www.dentalimagedataset.com/database/dentalcbct001. By using these anonymized data,
PT
any researcher can repeat the test that we describe herein.
CE
2.2. Volume of interest detection and segmentation Dental images were obtained using the KODAK 9500 CBCT imaging unit during routine
AC
controls of patients. The 3D dataset file scan is typically examined through imaging software bundled with the device. However, existing software does not have the necessary tools to mark the volume of interest (VOI) in dental images. Furthermore, the software does not contain a suitable tool for feature extraction. We therefore developed software using MATLAB R2014a. The developed software contains both a VOI marker and feature extraction tools. In addition to the DICOM viewer, VOI marking
ACCEPTED MANUSCRIPT
and feature extracting can be performed using this individual software. Furthermore, new features can be added to it for future studies. Adaptation of current segmentation methods for marking VOI on dental images is the continued focus of our research. In this study, we employed the manual segmentation method for VOI marking. Approximately 127 sections per lesion, and a total of 6397 axial slides were
CR IP T
manually marked. We accepted these manual segmented regions as the ground truth. A radiology expert marked the region of interest (ROI) on a selected 2D plane corresponding to the VOI with the lesion. This step was repeated for other sections on selected axial, sagittal, or coronal planes until the whole VOI segmentation was completed.
AN US
The Hounsfield values of each voxel within the segmented VOI were stored in a 3D matrix. The dimensions of the matrix were equivalent to a cube that surrounds the lesion (bounding box) in 3D space. Stored information in the 3D matrix was used for feature extraction of the detected lesion. Voxels containing lesion information in the matrix were represented as original
M
Hounsfield values. Voxels outside the lesion and voxels surrounding the lesion are represented
2.3. Feature extraction
ED
as a -1000 equivalent of air (see Fig. 2).
PT
Order statistics (median, standard deviation, skewness, kurtosis, entropy) and 3D Haralick features were calculated from the pixel intensities on each CBCT dataset to be used for binary
CE
classification experiments.
AC
2.3.1. Order statistics
An image histogram was used for calculating the order statistics [27]. The calculation of the normalized histogram from voxel values is shown as follows:
ACCEPTED MANUSCRIPT
P( Di )
N 1M 1 L 1
( f ( x, y, z) i) / NML
(1)
x 0 y 0 z 0
where Di represents discrete voxel values used for calculation of the histogram, and P(Di) represents the probability density value obtained from the number of repetitions at the histogram intensities. N, M, and L represent the numbers of voxels located along the x, y, and z directions
CR IP T
of the VOI. The mean, standard deviation, skewness, kurtosis, and entropy were calculated using probability density values P(Di), as shown in Equations (2) to (6), respectively.
H 1
Di P( Di )
AN US
i 0
(2)
H 1
( Di ) 2 P( Di )
(3)
i 0
H 1
1
3
( Di )3 P( Di )
PT
ED
4
H 1
1
4
(4)
i 0
M
3
( Di ) 4 P( Di ) 3
(5)
i 0
H 1
e P( Di ) log2 P( Di )
(6)
i 0
CE
The voxel values in the 3D dataset were sorted in ascending order in a generated list, Lv. The
AC
median was calculated as the middle value of Lv. 2.3.2. 3D Haralick feature co-occurrence matrix Textural information is often utilized in the analysis of medical images. Textural information is composed of interconnected pixel clusters, which contain varying levels of color information repeated in the image. To extract meaningful features from texture, the regional information is obtained by calculating variations in the image.
ACCEPTED MANUSCRIPT
The gray-level co-occurrence matrix (GLCM) proposed by Haralick [28] is generally used to obtain the textural features. GLCM is used in a texture analysis process to obtain the spatial dependency of gray-level values in the image. This matrix is generally preferred for calculation of certain statistics of the repetitive gray-level values on 2D image information. GLCM is a square matrix that demonstrates all the rows and columns of image gray-level brightness values.
CR IP T
Therefore, the GLCM size depends on the number of gray-level brightness values in the image. In 2D images, spatial dependencies of gray levels along the x and y axes are obtained to form a GLCM. In addition, spatial dependencies of gray levels along the z axis are taken into account to form a similar matrix in volumetric images. Therefore, the 3D GLCM matrix of a
AN US
volumetric image may include more information than a 2D GLCM matrix. Equation (7) represents the calculation of a 3D GLCM as an adapted version of the 2DGLCM matrix.
Md (i, j )
Wz dz Wx dx Wy dy
G( x, y, x) i AND G( x dx, y dy, z dz) j
z 1
x 1
y 1
(7)
M
if T rue 1, else False 0
ED
In Equation (7), i represents the i-th column and j represents the j-th row of the 3D GLCM. The relationship between the pixel pairs is represented as d. G(x,y,z) and G(x+dx,y+dy,z+dz)
PT
denote gray-level values at the (x,y,z) and (x+dx,y+dy,z+dz) coordinates on the (Wx×Wy×Wz)
CE
sliding window. Arguments x, y, z represent the position of the sliding window in the volumetric image information.
AC
GLCM is a matrix of n × n dimensions, and n refers to the number of gray levels in the image. To more rapidly calculate the matrix, gray levels are quantized into specific intervals to reduce the number of gray levels. In addition to the 2D GLCM method, 26 different interconnection directions are considered to calculate the relationship between the pixels in the 3D GLCM [29]. As shown in Fig. 3, 13 of those 26 interconnection directions are unique. In our study, we used 12 different features obtained from each 3D GLCM. Those features were respectively determined as energy, entropy, correlation, contrast, variance, sum mean,
ACCEPTED MANUSCRIPT
inertia, cluster shade, cluster tendency, homogeneity, maximum probability, and inverse variance [30]. For the creation of the 3D GLCM, 13 different directions and four different distance metrics were used to calculate the correlation between pixels. In Equation (7), argument d containing the distance metric is replaced with (1;2;4;8). As a result of these calculations, 12 different
CR IP T
feature information items were extracted from the 52 GLCM. Finally, we obtained 624 different 3D textural features from the 3D GLCM. 2.3.3. Feature vector
To generate the feature vector, the order statistics of each dataset were calculated in two ways.
AN US
Initially, the order statistics were calculated from the dataset that represented the lesion within the boundaries of the VOI. After that, the order statistics were calculated using the original Hounsfield values within the bounding box surrounding the segmented lesional VOI. Finally, 12
M
features were extracted from the order statistics and 624 features were extracted from the 3D GLCM of the Hounsfield values within the same bounding box. As a result, the feature vector
ED
consisted of 636 different values, including the order statistics and 3D textural feature information.
PT
2.4. Classification methods
CE
We conducted experiments with six different classifiers to diagnose periapical cyst and KCOT lesions detected with CBCT. The selected classifiers were the k-nearest neighbors (k-NN) [31],
AC
naive Bayes [32], decision trees [33], random forest [34], neural network (NN) [35], and support vector machine (SVM) [36]. 2.5. Experiments We conducted model application experiments for the classification of periapical cyst and KCOT lesions detected via CBCT.
ACCEPTED MANUSCRIPT
2.5.1. Classifier configuration In our study, we compared the performance of classification methods expressed in Section 2.4. We used Rapid Miner [37] data mining software for accomplishing the classification tests. The experiments were repeated with different parameters for each classifier to determine the optimal
CR IP T
parameter combination. Thus, we obtained the best classification results for each classifier. 2.5.2. Performance evaluation
Accuracy and F1 values [38] were calculated in training and testing phases. The resultant performance metrics were used to compare classifier performance. In our binary classification experiments, a confusion matrix was calculated to obtain the performance metrics. Table 1
AN US
represents the confusion matrix for the binary classification, where tp is true-positive, tn is truenegative, fp is false-positive, and fn is false-negative. The most widely used metric for measuring the classification performance is accuracy:
tp tn tp tn fp fn
(8)
ED
M
accuracy
In addition to accuracy, other metrics should be considered to facilitate a proper decision about
PT
the classifier performance. For this purpose, we used the F-score to calculate the relations
F - score
( 2 1)tp ( 2 1)tp 2 fn fp
(9)
AC
CE
between positive labels of data and those given by a classifier.
In binary classification, the F-score is calculated by assuming β = 1, and the resulting value is expressed as an F1 metric.
F1
2tp 2tp fp fn
(10)
ACCEPTED MANUSCRIPT
2.5.3. Model application We conducted our experiments on a 2.40-GHz 8 Core Intel 7 4700HQ CPU device with 16 GB of RAM. We used Rapid Miner for the model application stages. In our first experiment sets, we used features of 25 periapical cyst and 25 KCOT lesions to measure the performance of the classifiers discussed herein. In those tests, a ten-fold cross validation method was used for
CR IP T
verification of the classifier performances. In our second experiment sets, 50 different feature vectors obtained from the CBCT dataset were divided into two sub-groups. For the model application, 25 of these feature vectors were used for training phases; the remaining 25 feature vectors were used for the test phases. Each of the training and test sets included feature vectors
AN US
of periapical cyst and KCOT lesions. Stratified sampling [39] was used to allocate training and test sets with equal variance. In our third experiment sets, we used features of 25 periapical cyst and 25 KCOT lesions as in the second experiment sets. This time LOOCV method was used for verification of the classifier performances.
M
In the first, second, and third experiment sets, the performances of the classifiers were calculated by two methods. In the first method, the entire feature vector consisted of 636
ED
different values, which were used for the model application. The obtained results were used to measure the classifier performances. In the second method, the feature vector size was reduced
PT
to obtain the best feature combination for the classifier. A forward feature selection algorithm
CE
[40] was used to identify the most appropriate feature value combination for the classifier. 3. Results
AC
In the first experiment group, the full 50 feature vector with 636 attributes was used for the model application. Ten-fold cross validation was selected for model validation in the training and testing phases. As a result of the multiple experiments conducted, the SVM classifier achieved the best performance of 94.00% accuracy, 94.00% F1 with the ability to cope with many independent variables. That result was obtained by SVM with an analysis of variance (ANOVA) kernel structure, which performs well in multidimensional regression problems.
ACCEPTED MANUSCRIPT
NN achieved the second best results with 92.00% accuracy, and 91.67% F1 compared to the established models. NN achieved the best classification performance with three hidden layers, 500 training cycles, and 0.9 learning rate parameters. These results were followed by NB with the Laplace correction parameter. The remaining results sorted in the order of best performance were DT, RF and k-NN. (see Fig.4).
CR IP T
For the second experiment group, we used the forward feature selection algorithm to select the appropriate feature combinations for each classifier. We attempted different parameters for each classifier and achieved the best feature combinations by applying the feature selection algorithm. As in the first experiment group, we employed ten-fold cross validation for the
AN US
model validation. The SVM classifier achieved the best classification performance with 100% accuracy,
and
100% F1, in the experiments performed for each classifier. SVM is less prone to overfitting [31] while feature selection methods are also used for enhanced generalization by reducing
properties were considered together.
M
overfitting. SVM achieved this result with the dot kernel when the above two characteristic
ED
NB achieved the second best performance of 98.00% accuracy, 98.00% F1 with Laplace correction parameter. Selecting a subset of relevant features yielded better results with NB. The
PT
NN classifier achieved the third best result with the selected feature subsets. These results were
CE
respectively followed by k-NN, RF and DT performance results (see Fig. 5). In the third experiment group, we split the 50 feature vector into two groups using stratified
AC
sampling. Each 25-element group consisted of periapical cyst and KCOT lesion feature information items. Experiments were conducted with all 636 attributes of the feature vectors without using a feature selection algorithm. We employed split sample validation in the model application processes. First, the 25 feature vectors were used for the training phases in the experiments. Then, the other group consisting of 25 feature vectors was used in the test phases of the model application stages.
ACCEPTED MANUSCRIPT
As a result of the experiments, SVM achieved the best performance with 96.00% accuracy, and 96.00 % F1. SVM achieved the best classification performance with the ANOVA kernel parameter. SVM showed that it benefitted from all 636 elements of the feature vectors. NB achieved the second best performance of 92.00% accuracy, 92.37% F1 with the Laplace correction parameter. NN achieved the third best result with one hidden layer, 500 training
CR IP T
cycles, and 0.9 learning rate parameters. Those results were respectively followed by DT, RF, and k-NN performance results (see Fig. 6).
In the fourth experiment group, we employed split sample validation for the model application processes. We split the 50 feature vector into two groups using stratified sampling,
AN US
as was done in the third experiment group. Similar to the second experiment group, we used the forward feature selection algorithm to select appropriate feature combinations for each classifier. SVM achieved the best classification performance with 96.00% accuracy, and 96.00% F1 scores. Those scores were obtained with the ANOVA kernel parameter applied on
M
the SVM classifier.
RF achieved the second best result with 92.00% accuracy, 92.86% F1. RF achieved this
ED
result with the number of trees equal to ten, the selected criterion as accuracy, and 0.25 confidence level parameters. The third best performance was obtained using both NB and NN
PT
classifiers with 92.00% accuracy, and 92.00% F1 scores. NB achieved this result with Laplace
CE
correction, and fix bandwidth parameters and NN achieved this performance with two hidden layer, 500 training cycles, and 0.9 learning rate parameters. Those results were respectively
AC
followed by k-NN, and DT performance results (see Fig. 7). In the fifth experiment group, the full 50 feature vector with 636 attributes was used for the
model application, as was done in the first and third experiment groups. The LOOCV was selected for model validation in the training and testing phases. As a result of the multiple experiments conducted, the SVM classifier achieved the best results with 88.00% accuracy, and 88.00% F1 compared to the established models. SVM achieved the best classification performance with the dot kernel parameter. NN achieved the second best result of 88.00%
ACCEPTED MANUSCRIPT
accuracy, and 87.50% F1 with three hidden layer, 500 training cycles, and 0.3 learning rate parameters. Those results were respectively followed by DT, RF, NB, and k-NN performance results (see Fig. 8). For the sixth experiment group, we used the forward feature selection algorithm to select the appropriate feature combinations for each classifier. We attempted different parameters for each
CR IP T
classifier and achieved the best feature combinations by applying the feature selection algorithm. As in the fifth experiment group, we employed LOOCV for the model validation. The SVM classifier achieved the best classification performance with 94,00% accuracy, and 93,88% F1, in the experiments performed for each classifier. SVM achieved this result with the
AN US
Epechenikov kernel.
The second best performance was obtained using both NB and NN classifiers with 92.00% accuracy, and 92.00% F1 scores. NB achieved this result with Laplace correction parameter, and NN achieved this performance with two hidden layer, 500 training cycles, and 0.3 learning
M
rate parameters. Those results were respectively followed by RF, DT, and k-NN performance results (see Fig. 9).
ED
In our experiments, we chose the feature selection algorithm to compose a subset from the most appropriate combination of the features [41,42]. Elements in those subsets contained the
PT
most appropriate combination of features for enhancing the performance of the applied model.
CE
According to the results for the six different experimental groups, the feature vector subsets obtained using the forward feature selection algorithm were, in general, positively contributed to
AC
the performance of the classifier. In the experiments in which the ten-fold cross validation method was used, SVM achieved the best performance with the vector consisting of the single selected feature. This selected feature was obtained from 3D GLCM. In the experiments in which the split sample validation method was used, SVM achieved the best performance with a sub-feature vector consisting of two values obtained from 3D GLCM. In the last experiment set in which the leave one out validation method was used, SVM also achieved the best
ACCEPTED MANUSCRIPT
performance
with
a
sub-feature vector consisting of single value obtained from 3D GLCM Fig. 10 depicts the number of feature vector elements selected for each classifier as a result of the experiments performed with the forward feature selection algorithm.
CR IP T
4. Discussion The main objective of this study was to classify periapical cyst and KCOT lesions detected on volumetric images obtained by CBCT. A few studies have been conducted in the field of computer-aided diagnosis of dental apical lesions. Some of these studies were performed on images obtained with panoramic imaging methods. Published studies using CBCT are in early
AN US
stages.
One of these studies involved classification of jaw bone cysts and necrosis via the processing of panoramic radiographs [19]. In that study, the performances of DT, NB, NN, k-NN, and
in DT, NB, and NN classifiers.
M
SVM were compared. The resulting best performance value with 85% accuracy was measured
ED
Another published study addressed non-invasive differential diagnosis of dental periapical lesions obtained using CBCT [18]. In that study, 17 lesions detected in CBCT datasets were
PT
classified as cysts or granulomas. A feature vector was composed of order statistics, which were calculated from lesional region information. The LOOCV method was used for model
CE
validation. The updated version of this study [20] was conducted with 28 CBCT datasets. In both the first and last studies, the LDA-Adaboost classifier combination achieved the best
AC
performance results with an accuracy of 94.1%. Classification of lesions detected in dental panoramic images using SVM was the subject of
another study [43]. The dataset used in that study consisted of 133 different panoramic dental images. In the dataset, 53 of the dental images were identified as cystic lesions; the other 80 images were identified as tumor lesions. The feature vector was comprised of order statistics, GLCM, and gray-level run length matrix (GLRM) values derived from the dataset. As a result
ACCEPTED MANUSCRIPT
of this research, the best classification performance was obtained with an accuracy of 87.18%, and area under curve (AUC) of 0.9444. In a recent article [21], a hybrid method for accurate and fast classification of maxillofacial cysts was proposed. In that study, 96 histopathologically confirmed maxillofacial cysts detected in CBCT datasets were classified in three classes. A feature extraction approach was used for
CR IP T
discriminating those 38 radicular cyst, 36 dentigerous cyst, and 22 keratocystic odontogenic tumor lesions. The proposed framework was based on orthogonalized SPHARM and contourlet features. As a result, classification accuracies of 94.29%, and 96.48% was achieved with SVM and sparse discriminant analysis (SDA) classifiers.
AN US
In the present study we employed a new CBCT dataset that consists of 50 periapical lesion information items. The selected CBCT images were examined by radiologists using software specially prepared for this study. The radiologists marked all planes corresponding to VOI containing the lesion by using the tools available in the software. A 3D dataset of the lesional
M
region was obtained by combining the marked ROIs. A feature vector was formed from order statistics and 3D GLCM information obtained from each lesional VOI. The feature vector
ED
consisted of 636 different feature information items. Classification experiments were performed to distinguish the lesions as either periapical cysts or KCOT. Six different classifiers (k-NN,
PT
NB, DT, RF, NN, and SVM) were used for the model application experiments. Six different
CE
experiment groups were designed for the classification of periapical cyst and KCOT lesions obtained via CBCT.
AC
In the first experiment groups, all 50 feature vectors obtained from the CBCT dataset were used in training and testing. The ten-fold cross validation method was applied for model accuracy in those experiments. Different parameter combinations were used for each classifier to obtain the best classification performance. According to the experiment results, the SVM classifier achieved the best performance with 94.00% accuracy, and 94.00% F1 scores. According to the results, the use of features extracted from segmented CBCT datasets for classification of dental lesions as periapical cysts or KCOT was determined to be appropriate.
ACCEPTED MANUSCRIPT
The classifier performances were increased by reducing the size of the feature vector using the forward feature selection algorithm. The SVM classifier achieved good results in classifying periapical cyst and KCOT lesions obtained via the 50 segmented CBCT dataset. The best performance results were achieved as 100% accuracy, and 100% F1 using the SVM classifier with ten-fold cross validation for model validation and forward feature selection for feature
CR IP T
reduction. SVM achieved the best classification performance with 96.00% accuracy, and 96.00% F1 scores in the experiments in which a split sample validation method was used with a forward feature selection algorithm. SVM additionally achieved the best performance of 94.00% accuracy, and 93.88% F1 when LOOCV method was used with a forward feature
AN US
selection algorithm. 5. Conclusion
Endodontic lesions are amorphous in shape; thus, results of the segmentation implemented
M
by different oral radiologist may vary. Despite this limitation, a bounding box that surrounds the lesion in 3D space was considered for feature extraction of the segmented lesion in this study.
segmentation results.
ED
We are currently developing a hybrid segmentation method that will provide more accurate
PT
In contrast to previous studies, we conducted our experiments with a larger dataset consisting of images obtained via CBCT. We acquired more information for separating the
CE
lesions into the correct classes in terms of the number and diversity of the features extracted from this dataset. We expanded the domain of our study with multiple classifiers used in our
AC
experiments and different parameters that we attempted for each classifier. In sum, SVM achieved the best classification results for classifying dental periapical cysts
and KCOT lesions in volumetric images obtained using CBCT. In future work, we will expand the CBCT dataset with other types of dental pathologies, including histopathological diagnoses, and we will conduct multiclass classification experiments with the expanded dataset.
ACCEPTED MANUSCRIPT
Conflict of interest No conflicts of interest are declared. Acknowledgements The authors would like to thank Dr. Umit Cobanoglu, Dr. Celal Candirli, Dr. Nuray Yilmaz
CR IP T
Altintas, Dr. Yavuz Tolga Korkmaz, Oral Diagnosis and Radiology Department of Karadeniz Technical University for providing the dataset used in this study. The data referenced in this study were obtained and used with the permission of Karadeniz Technical University Clinical Research Ethics Committee.
[1]
AN US
References
S. White, M. Pharoah, Oral radiology: principles and interpretation, 6th ed., Mosby, 2013.
B. Basrani, Endodontic radiology, John Wiley & Sons, 2012.
[3]
D. Sarment, ed., Cone beam computed tomography, John Wiley & Sons, Inc, Chichester,
M
[2]
[4]
ED
UK, 2013. doi:10.1002/9781118769027.
W. De Vos, J. Casselman, G.R.J. Swennen, Cone-beam computerized tomography
PT
(CBCT) imaging of the oral and maxillofacial region: a systematic review of the literature.,
Int.
J.
Oral
Maxillofac.
Surg.
38
(2009)
609–25.
[5]
CE
doi:10.1016/j.ijom.2009.02.028. D.X. Ji, S.H. Ong, K.W.C. Foong, A level-set based approach for anterior teeth
AC
segmentation in cone beam computed tomography images., Comput. Biol. Med. 50 (2014) 116–28. doi:10.1016/j.compbiomed.2014.04.006.
[6]
R. Lloréns, V. Naranjo, F. López, M. Alcañiz, Jaw tissues segmentation in dental 3D CT images using fuzzy-connectedness and morphological processing, Comput. Methods Programs Biomed. 108 (2012) 832–843. doi:10.1016/j.cmpb.2012.05.014.
[7]
W.C. Scarfe, Imaging of maxillofacial trauma: evolutions and emerging revolutions.,
ACCEPTED MANUSCRIPT
Oral Surg. Oral Med. Oral Pathol. Oral Radiol. Endod. 100 (2005) S75-96. doi:10.1016/j.tripleo.2005.05.057. [8]
S. Kayipmaz, O.S. Sezgin, S.T. Saricaoglu, G. Çan, An in vitro comparison of diagnostic abilities of conventional radiography, storage phosphor, and cone beam computed
82. doi:10.1016/j.ejrad.2010.09.011. [9]
CR IP T
tomography to determine occlusal and approximal caries., Eur. J. Radiol. 80 (2011) 478–
W.C. Scarfe, A.G. Farman, P. Sukovic, Clinical applications of cone-beam computed tomography in dental practice., J. Can. Dent. Assoc. 72 (2006) 75–80.
[10]
S. Patel, New dimensions in endodontic imaging: Part 2. Cone beam computed
[11]
AN US
tomography., Int. Endod. J. 42 (2009) 463–75. doi:10.1111/j.1365-2591.2008.01531.x. T. Koivisto, W.R. Bowles, M. Rohrer, Frequency and distribution of radiolucent jaw lesions: A retrospective analysis of 9,723 cases., J. Endod. 38 (2012) 729–32. doi:10.1016/j.joen.2012.02.028.
N.R. Johnson, O.M. Gannon, N.W. Savage, M.D. Batstone, Frequency of odontogenic
M
[12]
cysts and tumors: a systematic review., J. Investig. Clin. Dent. 5 (2014) 9–14.
[13]
ED
doi:10.1111/jicd.12044.
H.M. Pinsky, S. Dyda, R.W. Pinsky, K.A. Misch, D.P. Sarment, Accuracy of three-
PT
dimensional measurements using cone-beam CT., Dentomaxillofac. Radiol. 35 (2006)
[14]
CE
410–6. doi:10.1259/dmfr/20987648. D.B. Prajapati, N.P. Desai, C.K. Modi, Influence of dental metallic artifact from
AC
multislice CT in the assessment of simulated mandibular lesions, in: 2012 Int. Conf. Commun. Syst. Netw. Technol., IEEE, 2012: pp. 198–202. doi:10.1109/CSNT.2012.50.
[15]
O. Aydemir, T. Kayikcioglu, Comparing common machine learning classifiers in lowdimensional feature vectors for brain computer interface applications, Int. J. Innov. Comput. Inf. Control. 9 (2013) 1145–1157.
[16]
U. Sevik, C. Kose, T. Berber, H. Erdol, Identification of suitable fundus images using automated quality assessment methods., J. Biomed. Opt. 19 (2014) 46006.
ACCEPTED MANUSCRIPT
doi:10.1117/1.JBO.19.4.046006. [17]
J.H.S. Simon, R. Enciso, J.-M. Malfaz, R. Roges, M. Bailey-Perry, A. Patel, Differential diagnosis of large periapical lesions using cone-beam computed tomography measurements and biopsy., J. Endod. 32 (2006) 833–7. doi:10.1016/j.joen.2006.03.008.
[18]
A. Flores, S. Rysavy, R. Enciso, K. Okada, Non-invasive differential diagnosis of dental
CR IP T
periapical lesions in cone-beam CT, in: 2009 IEEE Int. Symp. Biomed. Imaging From Nano to Macro, IEEE, 2009: pp. 566–569. doi:10.1109/ISBI.2009.5193110. [19]
J. Mikulka, E. Gescheidtová, M. Kabrda, V. Peřina, Classification of jaw bone cysts and necrosis via the processing of orthopantomograms, Radioengineering. 22 (2013).
K. Okada, S. Rysavy, A. Flores, M.G. Linguraru, Noninvasive differential diagnosis of
AN US
[20]
dental periapical lesions in cone-beam CT scans, Med. Phys. 42 (2015) 1653–1665. doi:10.1118/1.4914418. [21]
F. Abdolali, R.A. Zoroofi, Y. Otake, Y. Sato, Automated classification of maxillofacial
Comput.
Methods
M
cysts in cone beam CT images using contourlet transformation and Spherical Harmonics, Programs
Biomed.
139
(2017)
197–207.
[22]
ED
doi:10.1016/j.cmpb.2016.10.024.
D. Bhargava, A. Deshpande, M.A. Pogrel, Keratocystic odontogenic tumour (KCOT)--a
[23]
CE
0302-9.
PT
cyst to a tumour., Oral Maxillofac. Surg. 16 (2012) 163–70. doi:10.1007/s10006-011-
L. Barnes, W.H. Organization, I.A. for R. on Cancer, Pathology and genetics of head and
AC
neck tumours, IARC, 2005.
[24]
L.A. Gaitán-Cepeda, D. Quezada-Rivera, F. Tenorio-Rocha, E.R. Leyva-Huerta, Reclassification of odontogenic keratocyst as tumour. Impact on the odontogenic tumours prevalence., Oral Dis. 16 (2010) 185–7. doi:10.1111/j.1601-0825.2009.01625.x.
[25]
M.S. Tekkesin, S. Mutlu, V. Olgac, Expressions of bax, bcl-2 and ki-67 in odontogenic keratocysts (keratocystic odontogenic tumor) in comparison with ameloblastomas and radicular cysts, Turkish J. Pathol. 28 (2012) 49. doi:10.5146/tjpath.2012.01097.
ACCEPTED MANUSCRIPT
[26]
G.N. Hounsfield, Computerized transverse axial scanning (tomography): Part I. Description of system. 1973., Br. J. Radiol. 68 (1995) H166-72.
[27]
R.C. Gonzalez, R.E. Woods, Digital image processing (3rd edition), Prentice Hall, 2008.
[28]
R.M. Haralick, K. Shanmugam, I. Dinstein, Textural features for image classification, IEEE Trans. Syst. Man. Cybern. 3 (1973) 610–621. doi:10.1109/TSMC.1973.4309314. E. Ben Othmen, M. Sayadi, F. Fniaech, 3D gray level co-occurrence matrices for
CR IP T
[29]
volumetric texture classification, in: 3rd Int. Conf. Syst. Control, IEEE, 2013: pp. 833– 837. doi:10.1109/ICoSC.2013.6750953. [30]
J. Varjo, Implementing texture analysis software frame for magnetic resonance image
[31]
J. Han, M. Kamber, J. Pei, Data mining: concepts and techniques, 3rd ed., Morgan Kaufmann Publishers Inc., 2011.
[32]
AN US
data in MATLAB, Tampere University of Technology, 2014.
N. Friedman, D. Geiger, M. Goldszmidt, Bayesian network classifiers, Mach. Learn. 29
L. Rokach, Data mining with decision trees: Theory and applications, World scientific, 2007.
[34]
L.
Breiman,
ED
[33]
M
(1997) 131–163. doi:10.1023/A:1007465528199.
Random
forests,
Mach.
Learn.
45
(2001)
5–32.
R.P. Lippmann, Pattern classification using neural networks, IEEE Commun. Mag. 27
CE
[35]
PT
doi:10.1023/A:1010933404324.
(1989) 47–50. doi:10.1109/35.41401. C. Cortes, V. Vapnik, Support-vector networks, Mach. Learn. 20 (1995) 273–297.
AC
[36]
doi:10.1007/BF00994018.
[37]
I. Mierswa, M. Wurst, R. Klinkenberg, M. Scholz, T. Euler, YALE, in: Proc. 12th ACM SIGKDD Int. Conf. Knowl. Discov. Data Min. - KDD ’06, ACM Press, New York, New York, USA, 2006: p. 935. doi:10.1145/1150402.1150531.
[38]
M. Sokolova, G. Lapalme, A systematic analysis of performance measures for classification
tasks,
Inf.
Process.
Manag.
45
(2009)
427–437.
ACCEPTED MANUSCRIPT
doi:10.1016/j.ipm.2009.03.002. [39]
M. Shahrokh Esfahani, E.R. Dougherty, Effect of separate sampling on classification accuracy., Bioinformatics. 30 (2014) 242–50. doi:10.1093/bioinformatics/btt662.
[40]
P. Pudil, J. Novovičová, J. Kittler, Floating search methods in feature selection, Pattern Recognit. Lett. 15 (1994) 1119–1125. doi:10.1016/0167-8655(94)90127-9. I. Guyon, A. Elisseeff, An introduction to variable and feature selection, J. Mach. Learn. Res. 3 (2003) 1157–1182.
[42]
L. Yu, H. Liu, Efficient feature selection via analysis of relevance and redundancy, J. Mach. Learn. Res. 5 (2004) 1205–1224.
I. Nurtanio, Classifying cyst and tumor lesion using support vector machine based on
AN US
[43]
CR IP T
[41]
AC
CE
PT
ED
M
dental panoramic images texture features, Int. J. Comput. Sci. 40 (2013).
ACCEPTED MANUSCRIPT
Table 1. Confusion matrix created for binary classification Data class Classified as Positive Classified as Negative tp
fn
Negative
fp
tn
CR IP T
Positive
PT
ED
M
AN US
Fig. 1. Block diagram representing the outline of study
CE
Fig. 2. Marked lesional region of 3D CBCT data corresponding to a 2D plane (a). 3D
AC
demonstration of a lesional region as a result of the markings on each corresponding 2D plane section (b)
CR IP T
ACCEPTED MANUSCRIPT
AC
CE
PT
ED
M
AN US
Fig. 3. Vector of 13 different directions used in the calculation of the 3D GLCM
Fig. 4. Results of initial experiments of 636 features using ten-fold cross validation
AN US
CR IP T
ACCEPTED MANUSCRIPT
AC
CE
PT
ED
M
Fig. 5. Results of forward feature selection experiments using ten-fold cross validation
Fig. 6. Results of initial experiments of 636 features using split sample validation
AN US
CR IP T
ACCEPTED MANUSCRIPT
AC
CE
PT
ED
M
Fig. 7. Results of forward feature selection experiments using split sample validation
Fig. 8. Results of initial experiments of 636 features using leave-one-out cross validation
AN US
CR IP T
ACCEPTED MANUSCRIPT
AC
CE
PT
ED
M
Fig. 9. Results of forward feature selection experiments using leave-one-out cross validation
Fig. 10. Number of features selected using the forward feature selection algorithm
Computer-aided diagnosis of periapical cyst and keratocystic odontogenic tumor on cone beam computed tomography E. Yilmaz , T. Kayikcioglu , S. Kayipmaz PII: DOI: Reference:
S0169-2607(16)30429-1 10.1016/j.cmpb.2017.05.012 COMM 4427
To appear in:
Computer Methods and Programs in Biomedicine
Received date: Revised date: Accepted date:
2 May 2016 15 April 2017 26 May 2017
Please cite this article as: E. Yilmaz , T. Kayikcioglu , S. Kayipmaz , Computer-aided diagnosis of periapical cyst and keratocystic odontogenic tumor on cone beam computed tomography, Computer Methods and Programs in Biomedicine (2017), doi: 10.1016/j.cmpb.2017.05.012
This is a PDF file of an unedited manuscript that has been accepted for publication. As a service to our customers we are providing this early version of the manuscript. The manuscript will undergo copyediting, typesetting, and review of the resulting proof before it is published in its final form. Please note that during the production process errors may be discovered which could affect the content, and all legal disclaimers that apply to the journal pertain.
ACCEPTED MANUSCRIPT
Highlights
• • • •
The database used in this study is consisted of 50 different dental CBCT scans. Dental periapical cyst and keratocystic odontogenic tumor lesions were classified. Six different classifiers were used for the classification experiments. The Support Vector Machine (SVM) achieved the best classification performances.
AC
CE
PT
ED
M
AN US
CR IP T
•
A decision support system for classification of dental lesions on CBCT is proposed.
ACCEPTED MANUSCRIPT
Computer-aided
diagnosis
of
periapical
cyst
and
keratocystic odontogenic tumor on cone beam computed
E. Yilmaz1,*, T. Kayikcioglu2, S. Kayipmaz3
1
Department of Computer Engineering, Karadeniz Technical University, 61080, Trabzon,
AN US
Turkey 2
Department of Electrical and Electronics Engineering, Karadeniz Technical University, 61080,
Trabzon, Turkey 3
CR IP T
tomography
Department of Oral Diagnosis and Radiology, Karadeniz Technical University Faculty of
M
Dentistry, 61080, Trabzon, Turkey
PT
Abstract
ED
*Corresponding author; [email protected]
CE
Background and objectives: In this article, we propose a decision support system for effective classification of dental periapical cyst and keratocystic odontogenic tumor (KCOT) lesions obtained via
AC
cone beam computed tomography (CBCT). CBCT has been effectively used in recent years for diagnosing dental pathologies and determining their boundaries and content. Unlike other imaging techniques, CBCT provides detailed and distinctive information about the pathologies by enabling a three-dimensional (3D) image of the region to be displayed. Methods: We employed 50 CBCT 3D image dataset files as the full dataset of our study. These datasets were identified by experts as periapical cyst and KCOT lesions according to the clinical, radiographic and
ACCEPTED MANUSCRIPT
histopathologic features. Segmentation operations were performed on the CBCT images using viewer software that we developed. Using the tools of this software, we marked the lesional volume of interest and calculated and applied the order statistics and 3D gray-level co-occurrence matrix for each CBCT dataset. A feature vector of the lesional region, including 636 different feature items, was created from those statistics. Six classifiers were used for the classification experiments.
CR IP T
Results: The Support Vector Machine (SVM) classifier achieved the best classification performance with 100% accuracy, and 100% F-score (F1) scores as a result of the experiments in which a ten-fold cross validation method was used with a forward feature selection algorithm. SVM achieved the best classification performance with 96.00% accuracy, and 96.00% F1 scores in the experiments in which a
AN US
split sample validation method was used with a forward feature selection algorithm. SVM additionally achieved the best performance of 94.00% accuracy, and 93.88% F1 in which a leave-one-out (LOOCV) method was used with a forward feature selection algorithm.
Conclusions: Based on the results, we determined that periapical cyst and KCOT lesions can be
M
classified with a high accuracy with the models that we built using the new dataset selected for this study. The studies mentioned in this article, along with the selected 3D dataset, 3D statistics calculated from the
ED
dataset, and performance results of the different classifiers, comprise an important contribution to the
Keywords
PT
field of computer-aided diagnosis of dental apical lesions.
CE
Computer aided diagnosis; Dental apical lesion; Classifier; Cone beam computed tomography; Periapical
AC
cyst and keratocystic odontogenic tumor; Volumetric textural features; Dental image dataset
1. Introduction Oral diagnostic radiology comprises a sub-branch of dentistry relating to information needed for treatment planning for dental diseases. Oral examination may not always be sufficient for the diagnosis of dental disease. Therefore, during the dental examination, radiological imaging methods are utilized to assist in the examination of invisible intrabony regions. Dental images
ACCEPTED MANUSCRIPT
can be obtained via periapical radiography, panoramic radiography, and cone beam computed tomography (CBCT) [1]. Dental regions containing one or more tooth roots can be viewed using periapical radiographs. Panoramic radiographs are used to obtain a two-dimensional (2D) panoramic view of the upper and lower jaw. Images are acquired with the help of conventional X-rays in both
CR IP T
periapical and panoramic devices [2].
Dental CBCT technology has been used since the late 1990s. Three-dimensional (3D) images of the neck and chin area can be obtained using this technology [3]. CBCT devices are useful in implant planning [4], teeth segmentation [5], jaw tissue segmentation [6], detection of
AN US
facial traumas of the jaw [7], and obtaining 3D images of lesions and other pathologies [8].
Pathological lesions around the tooth root can be identified based on their size and location using volumetric images obtained with these devices. The data obtained by CBCT can sometimes indicate a lesion or abnormal anatomical findings. Radiology experts typically
M
examine those images along axial, sagittal, and coronal planes and can thereby readily make determinations about the findings [9,10].
ED
Detected abnormalities are often expressed as endodontic or periapical lesions [11,12]. During radiological examinations, it is important for the doctor to understand the characteristics
PT
of detected lesions in a timely and appropriate manner. According to current studies, volumetric
CE
measurements on a CBCT scan are an accurate indicator of the volume of the actual periapical lesion in the jawbone [13,14] . As a result, the planned treatment can be effective in reversing
AC
the destruction of the pathology, which supports the importance of CBCT and other radiological examination methods [2]. Lesions detected as a result of scanning are evaluated with consideration of some of their
properties, such as location, periphery, and internal structures in the surrounding area [3]. Some lesions have the potential to develop into cancer. Early diagnosis of such lesions is crucial to preventing disease progression [1].
ACCEPTED MANUSCRIPT
In recent years, many studies have been performed in the areas of medical imaging and signal processing. A majority of these studies address the classification of medical information. In each study, a specific method is preferred depending on the characteristic of the disease of interest. Classification methods are used to obtain meaningful results from medical information. Feature extraction processes are applied to that medical information for classification
CR IP T
experiments. In accordance with the content of the medical information, various features are used in each study. The power spectral density technique, for example, may be useful for extracting features from the obtained signal data in research conducted on a brain computer interface [15]. To establish a good classifier in a study performed on a retinal image database,
AN US
features such as shape, texture, and intensity may be suitable for determining the image quality [16].
Lesion classification is a sub-problem in the field of computer vision. Dental structures located in the head and neck area do not have a specific form. Therefore, research conducted on
M
dental images in the field of computer vision is extremely challenging. In the present study, we employ textural features to address the formal diversity observed in dental images. Unlike
ED
research in medical image processing in other disciplines, few published studies exist that address computer-aided detection and classification of apical lesions [17–21].
PT
In this article, we discuss classification of periapical cyst and keratocystic odontogenic tumor
CE
(KCOT) lesions observed in 50 different dental CBCT datasets. These datasets were identified by experts as periapical cysts and KCOT lesions according to the clinical, radiographic, and
AC
histopathologic features. KCOT was previously referred to as the odontogenic keratocyst (OKC). It was reclassified as the keratocystic odontogenic tumor by the World Health Organization (WHO) in 2005 [22,23]. The KCOT lesion is known for its aggressive behavior and high recurrence rate [24,25]. It is therefore necessary to especially focus on differentiating KCOT from cystic lesions. The aim of this study was thus to develop a CBCT classifier method for exact differentiation of KCOT from other cystic lesions.
ACCEPTED MANUSCRIPT
2. Materials and methods Unlike previous research, we used a new 3D dental image dataset. Accordingly, we further extracted a new feature vector derived from this dataset. In our experiments, lesions were classified into two groups using ten-fold cross validation, split sample validation and leave-one-out cross validation (LOOCV) methods. Fig.1 presents a flow diagram of the
CR IP T
operations performed to identify the most effective model for correctly distinguishing periapical cyst and KCOT lesion [22] images obtained via CBCT. 2.1. Database
The database used in this study consisted of a dental CBCT dataset obtained from 50 patients
AN US
between the years of 2013 and 2016 who attended Karadeniz Technical University, Faculty of Dentistry, Department of Oral Diagnosis and Radiology Clinic for routine controls. The images were obtained with a KODAK K9500 Trophy CBCT device. Data were recorded in 3D dataset
M
files using the Digital Imaging and Communications in Medicine (DICOM) standard. Medical image information comprising the 3D dataset file was converted to the appropriate data range in Hounsfield
scale
[26].
The
dataset
of
our
study
is
accessible
at
ED
the
http://www.dentalimagedataset.com/database/dentalcbct001. By using these anonymized data,
PT
any researcher can repeat the test that we describe herein.
CE
2.2. Volume of interest detection and segmentation Dental images were obtained using the KODAK 9500 CBCT imaging unit during routine
AC
controls of patients. The 3D dataset file scan is typically examined through imaging software bundled with the device. However, existing software does not have the necessary tools to mark the volume of interest (VOI) in dental images. Furthermore, the software does not contain a suitable tool for feature extraction. We therefore developed software using MATLAB R2014a. The developed software contains both a VOI marker and feature extraction tools. In addition to the DICOM viewer, VOI marking
ACCEPTED MANUSCRIPT
and feature extracting can be performed using this individual software. Furthermore, new features can be added to it for future studies. Adaptation of current segmentation methods for marking VOI on dental images is the continued focus of our research. In this study, we employed the manual segmentation method for VOI marking. Approximately 127 sections per lesion, and a total of 6397 axial slides were
CR IP T
manually marked. We accepted these manual segmented regions as the ground truth. A radiology expert marked the region of interest (ROI) on a selected 2D plane corresponding to the VOI with the lesion. This step was repeated for other sections on selected axial, sagittal, or coronal planes until the whole VOI segmentation was completed.
AN US
The Hounsfield values of each voxel within the segmented VOI were stored in a 3D matrix. The dimensions of the matrix were equivalent to a cube that surrounds the lesion (bounding box) in 3D space. Stored information in the 3D matrix was used for feature extraction of the detected lesion. Voxels containing lesion information in the matrix were represented as original
M
Hounsfield values. Voxels outside the lesion and voxels surrounding the lesion are represented
2.3. Feature extraction
ED
as a -1000 equivalent of air (see Fig. 2).
PT
Order statistics (median, standard deviation, skewness, kurtosis, entropy) and 3D Haralick features were calculated from the pixel intensities on each CBCT dataset to be used for binary
CE
classification experiments.
AC
2.3.1. Order statistics
An image histogram was used for calculating the order statistics [27]. The calculation of the normalized histogram from voxel values is shown as follows:
ACCEPTED MANUSCRIPT
P( Di )
N 1M 1 L 1
( f ( x, y, z) i) / NML
(1)
x 0 y 0 z 0
where Di represents discrete voxel values used for calculation of the histogram, and P(Di) represents the probability density value obtained from the number of repetitions at the histogram intensities. N, M, and L represent the numbers of voxels located along the x, y, and z directions
CR IP T
of the VOI. The mean, standard deviation, skewness, kurtosis, and entropy were calculated using probability density values P(Di), as shown in Equations (2) to (6), respectively.
H 1
Di P( Di )
AN US
i 0
(2)
H 1
( Di ) 2 P( Di )
(3)
i 0
H 1
1
3
( Di )3 P( Di )
PT
ED
4
H 1
1
4
(4)
i 0
M
3
( Di ) 4 P( Di ) 3
(5)
i 0
H 1
e P( Di ) log2 P( Di )
(6)
i 0
CE
The voxel values in the 3D dataset were sorted in ascending order in a generated list, Lv. The
AC
median was calculated as the middle value of Lv. 2.3.2. 3D Haralick feature co-occurrence matrix Textural information is often utilized in the analysis of medical images. Textural information is composed of interconnected pixel clusters, which contain varying levels of color information repeated in the image. To extract meaningful features from texture, the regional information is obtained by calculating variations in the image.
ACCEPTED MANUSCRIPT
The gray-level co-occurrence matrix (GLCM) proposed by Haralick [28] is generally used to obtain the textural features. GLCM is used in a texture analysis process to obtain the spatial dependency of gray-level values in the image. This matrix is generally preferred for calculation of certain statistics of the repetitive gray-level values on 2D image information. GLCM is a square matrix that demonstrates all the rows and columns of image gray-level brightness values.
CR IP T
Therefore, the GLCM size depends on the number of gray-level brightness values in the image. In 2D images, spatial dependencies of gray levels along the x and y axes are obtained to form a GLCM. In addition, spatial dependencies of gray levels along the z axis are taken into account to form a similar matrix in volumetric images. Therefore, the 3D GLCM matrix of a
AN US
volumetric image may include more information than a 2D GLCM matrix. Equation (7) represents the calculation of a 3D GLCM as an adapted version of the 2DGLCM matrix.
Md (i, j )
Wz dz Wx dx Wy dy
G( x, y, x) i AND G( x dx, y dy, z dz) j
z 1
x 1
y 1
(7)
M
if T rue 1, else False 0
ED
In Equation (7), i represents the i-th column and j represents the j-th row of the 3D GLCM. The relationship between the pixel pairs is represented as d. G(x,y,z) and G(x+dx,y+dy,z+dz)
PT
denote gray-level values at the (x,y,z) and (x+dx,y+dy,z+dz) coordinates on the (Wx×Wy×Wz)
CE
sliding window. Arguments x, y, z represent the position of the sliding window in the volumetric image information.
AC
GLCM is a matrix of n × n dimensions, and n refers to the number of gray levels in the image. To more rapidly calculate the matrix, gray levels are quantized into specific intervals to reduce the number of gray levels. In addition to the 2D GLCM method, 26 different interconnection directions are considered to calculate the relationship between the pixels in the 3D GLCM [29]. As shown in Fig. 3, 13 of those 26 interconnection directions are unique. In our study, we used 12 different features obtained from each 3D GLCM. Those features were respectively determined as energy, entropy, correlation, contrast, variance, sum mean,
ACCEPTED MANUSCRIPT
inertia, cluster shade, cluster tendency, homogeneity, maximum probability, and inverse variance [30]. For the creation of the 3D GLCM, 13 different directions and four different distance metrics were used to calculate the correlation between pixels. In Equation (7), argument d containing the distance metric is replaced with (1;2;4;8). As a result of these calculations, 12 different
CR IP T
feature information items were extracted from the 52 GLCM. Finally, we obtained 624 different 3D textural features from the 3D GLCM. 2.3.3. Feature vector
To generate the feature vector, the order statistics of each dataset were calculated in two ways.
AN US
Initially, the order statistics were calculated from the dataset that represented the lesion within the boundaries of the VOI. After that, the order statistics were calculated using the original Hounsfield values within the bounding box surrounding the segmented lesional VOI. Finally, 12
M
features were extracted from the order statistics and 624 features were extracted from the 3D GLCM of the Hounsfield values within the same bounding box. As a result, the feature vector
ED
consisted of 636 different values, including the order statistics and 3D textural feature information.
PT
2.4. Classification methods
CE
We conducted experiments with six different classifiers to diagnose periapical cyst and KCOT lesions detected with CBCT. The selected classifiers were the k-nearest neighbors (k-NN) [31],
AC
naive Bayes [32], decision trees [33], random forest [34], neural network (NN) [35], and support vector machine (SVM) [36]. 2.5. Experiments We conducted model application experiments for the classification of periapical cyst and KCOT lesions detected via CBCT.
ACCEPTED MANUSCRIPT
2.5.1. Classifier configuration In our study, we compared the performance of classification methods expressed in Section 2.4. We used Rapid Miner [37] data mining software for accomplishing the classification tests. The experiments were repeated with different parameters for each classifier to determine the optimal
CR IP T
parameter combination. Thus, we obtained the best classification results for each classifier. 2.5.2. Performance evaluation
Accuracy and F1 values [38] were calculated in training and testing phases. The resultant performance metrics were used to compare classifier performance. In our binary classification experiments, a confusion matrix was calculated to obtain the performance metrics. Table 1
AN US
represents the confusion matrix for the binary classification, where tp is true-positive, tn is truenegative, fp is false-positive, and fn is false-negative. The most widely used metric for measuring the classification performance is accuracy:
tp tn tp tn fp fn
(8)
ED
M
accuracy
In addition to accuracy, other metrics should be considered to facilitate a proper decision about
PT
the classifier performance. For this purpose, we used the F-score to calculate the relations
F - score
( 2 1)tp ( 2 1)tp 2 fn fp
(9)
AC
CE
between positive labels of data and those given by a classifier.
In binary classification, the F-score is calculated by assuming β = 1, and the resulting value is expressed as an F1 metric.
F1
2tp 2tp fp fn
(10)
ACCEPTED MANUSCRIPT
2.5.3. Model application We conducted our experiments on a 2.40-GHz 8 Core Intel 7 4700HQ CPU device with 16 GB of RAM. We used Rapid Miner for the model application stages. In our first experiment sets, we used features of 25 periapical cyst and 25 KCOT lesions to measure the performance of the classifiers discussed herein. In those tests, a ten-fold cross validation method was used for
CR IP T
verification of the classifier performances. In our second experiment sets, 50 different feature vectors obtained from the CBCT dataset were divided into two sub-groups. For the model application, 25 of these feature vectors were used for training phases; the remaining 25 feature vectors were used for the test phases. Each of the training and test sets included feature vectors
AN US
of periapical cyst and KCOT lesions. Stratified sampling [39] was used to allocate training and test sets with equal variance. In our third experiment sets, we used features of 25 periapical cyst and 25 KCOT lesions as in the second experiment sets. This time LOOCV method was used for verification of the classifier performances.
M
In the first, second, and third experiment sets, the performances of the classifiers were calculated by two methods. In the first method, the entire feature vector consisted of 636
ED
different values, which were used for the model application. The obtained results were used to measure the classifier performances. In the second method, the feature vector size was reduced
PT
to obtain the best feature combination for the classifier. A forward feature selection algorithm
CE
[40] was used to identify the most appropriate feature value combination for the classifier. 3. Results
AC
In the first experiment group, the full 50 feature vector with 636 attributes was used for the model application. Ten-fold cross validation was selected for model validation in the training and testing phases. As a result of the multiple experiments conducted, the SVM classifier achieved the best performance of 94.00% accuracy, 94.00% F1 with the ability to cope with many independent variables. That result was obtained by SVM with an analysis of variance (ANOVA) kernel structure, which performs well in multidimensional regression problems.
ACCEPTED MANUSCRIPT
NN achieved the second best results with 92.00% accuracy, and 91.67% F1 compared to the established models. NN achieved the best classification performance with three hidden layers, 500 training cycles, and 0.9 learning rate parameters. These results were followed by NB with the Laplace correction parameter. The remaining results sorted in the order of best performance were DT, RF and k-NN. (see Fig.4).
CR IP T
For the second experiment group, we used the forward feature selection algorithm to select the appropriate feature combinations for each classifier. We attempted different parameters for each classifier and achieved the best feature combinations by applying the feature selection algorithm. As in the first experiment group, we employed ten-fold cross validation for the
AN US
model validation. The SVM classifier achieved the best classification performance with 100% accuracy,
and
100% F1, in the experiments performed for each classifier. SVM is less prone to overfitting [31] while feature selection methods are also used for enhanced generalization by reducing
properties were considered together.
M
overfitting. SVM achieved this result with the dot kernel when the above two characteristic
ED
NB achieved the second best performance of 98.00% accuracy, 98.00% F1 with Laplace correction parameter. Selecting a subset of relevant features yielded better results with NB. The
PT
NN classifier achieved the third best result with the selected feature subsets. These results were
CE
respectively followed by k-NN, RF and DT performance results (see Fig. 5). In the third experiment group, we split the 50 feature vector into two groups using stratified
AC
sampling. Each 25-element group consisted of periapical cyst and KCOT lesion feature information items. Experiments were conducted with all 636 attributes of the feature vectors without using a feature selection algorithm. We employed split sample validation in the model application processes. First, the 25 feature vectors were used for the training phases in the experiments. Then, the other group consisting of 25 feature vectors was used in the test phases of the model application stages.
ACCEPTED MANUSCRIPT
As a result of the experiments, SVM achieved the best performance with 96.00% accuracy, and 96.00 % F1. SVM achieved the best classification performance with the ANOVA kernel parameter. SVM showed that it benefitted from all 636 elements of the feature vectors. NB achieved the second best performance of 92.00% accuracy, 92.37% F1 with the Laplace correction parameter. NN achieved the third best result with one hidden layer, 500 training
CR IP T
cycles, and 0.9 learning rate parameters. Those results were respectively followed by DT, RF, and k-NN performance results (see Fig. 6).
In the fourth experiment group, we employed split sample validation for the model application processes. We split the 50 feature vector into two groups using stratified sampling,
AN US
as was done in the third experiment group. Similar to the second experiment group, we used the forward feature selection algorithm to select appropriate feature combinations for each classifier. SVM achieved the best classification performance with 96.00% accuracy, and 96.00% F1 scores. Those scores were obtained with the ANOVA kernel parameter applied on
M
the SVM classifier.
RF achieved the second best result with 92.00% accuracy, 92.86% F1. RF achieved this
ED
result with the number of trees equal to ten, the selected criterion as accuracy, and 0.25 confidence level parameters. The third best performance was obtained using both NB and NN
PT
classifiers with 92.00% accuracy, and 92.00% F1 scores. NB achieved this result with Laplace
CE
correction, and fix bandwidth parameters and NN achieved this performance with two hidden layer, 500 training cycles, and 0.9 learning rate parameters. Those results were respectively
AC
followed by k-NN, and DT performance results (see Fig. 7). In the fifth experiment group, the full 50 feature vector with 636 attributes was used for the
model application, as was done in the first and third experiment groups. The LOOCV was selected for model validation in the training and testing phases. As a result of the multiple experiments conducted, the SVM classifier achieved the best results with 88.00% accuracy, and 88.00% F1 compared to the established models. SVM achieved the best classification performance with the dot kernel parameter. NN achieved the second best result of 88.00%
ACCEPTED MANUSCRIPT
accuracy, and 87.50% F1 with three hidden layer, 500 training cycles, and 0.3 learning rate parameters. Those results were respectively followed by DT, RF, NB, and k-NN performance results (see Fig. 8). For the sixth experiment group, we used the forward feature selection algorithm to select the appropriate feature combinations for each classifier. We attempted different parameters for each
CR IP T
classifier and achieved the best feature combinations by applying the feature selection algorithm. As in the fifth experiment group, we employed LOOCV for the model validation. The SVM classifier achieved the best classification performance with 94,00% accuracy, and 93,88% F1, in the experiments performed for each classifier. SVM achieved this result with the
AN US
Epechenikov kernel.
The second best performance was obtained using both NB and NN classifiers with 92.00% accuracy, and 92.00% F1 scores. NB achieved this result with Laplace correction parameter, and NN achieved this performance with two hidden layer, 500 training cycles, and 0.3 learning
M
rate parameters. Those results were respectively followed by RF, DT, and k-NN performance results (see Fig. 9).
ED
In our experiments, we chose the feature selection algorithm to compose a subset from the most appropriate combination of the features [41,42]. Elements in those subsets contained the
PT
most appropriate combination of features for enhancing the performance of the applied model.
CE
According to the results for the six different experimental groups, the feature vector subsets obtained using the forward feature selection algorithm were, in general, positively contributed to
AC
the performance of the classifier. In the experiments in which the ten-fold cross validation method was used, SVM achieved the best performance with the vector consisting of the single selected feature. This selected feature was obtained from 3D GLCM. In the experiments in which the split sample validation method was used, SVM achieved the best performance with a sub-feature vector consisting of two values obtained from 3D GLCM. In the last experiment set in which the leave one out validation method was used, SVM also achieved the best
ACCEPTED MANUSCRIPT
performance
with
a
sub-feature vector consisting of single value obtained from 3D GLCM Fig. 10 depicts the number of feature vector elements selected for each classifier as a result of the experiments performed with the forward feature selection algorithm.
CR IP T
4. Discussion The main objective of this study was to classify periapical cyst and KCOT lesions detected on volumetric images obtained by CBCT. A few studies have been conducted in the field of computer-aided diagnosis of dental apical lesions. Some of these studies were performed on images obtained with panoramic imaging methods. Published studies using CBCT are in early
AN US
stages.
One of these studies involved classification of jaw bone cysts and necrosis via the processing of panoramic radiographs [19]. In that study, the performances of DT, NB, NN, k-NN, and
in DT, NB, and NN classifiers.
M
SVM were compared. The resulting best performance value with 85% accuracy was measured
ED
Another published study addressed non-invasive differential diagnosis of dental periapical lesions obtained using CBCT [18]. In that study, 17 lesions detected in CBCT datasets were
PT
classified as cysts or granulomas. A feature vector was composed of order statistics, which were calculated from lesional region information. The LOOCV method was used for model
CE
validation. The updated version of this study [20] was conducted with 28 CBCT datasets. In both the first and last studies, the LDA-Adaboost classifier combination achieved the best
AC
performance results with an accuracy of 94.1%. Classification of lesions detected in dental panoramic images using SVM was the subject of
another study [43]. The dataset used in that study consisted of 133 different panoramic dental images. In the dataset, 53 of the dental images were identified as cystic lesions; the other 80 images were identified as tumor lesions. The feature vector was comprised of order statistics, GLCM, and gray-level run length matrix (GLRM) values derived from the dataset. As a result
ACCEPTED MANUSCRIPT
of this research, the best classification performance was obtained with an accuracy of 87.18%, and area under curve (AUC) of 0.9444. In a recent article [21], a hybrid method for accurate and fast classification of maxillofacial cysts was proposed. In that study, 96 histopathologically confirmed maxillofacial cysts detected in CBCT datasets were classified in three classes. A feature extraction approach was used for
CR IP T
discriminating those 38 radicular cyst, 36 dentigerous cyst, and 22 keratocystic odontogenic tumor lesions. The proposed framework was based on orthogonalized SPHARM and contourlet features. As a result, classification accuracies of 94.29%, and 96.48% was achieved with SVM and sparse discriminant analysis (SDA) classifiers.
AN US
In the present study we employed a new CBCT dataset that consists of 50 periapical lesion information items. The selected CBCT images were examined by radiologists using software specially prepared for this study. The radiologists marked all planes corresponding to VOI containing the lesion by using the tools available in the software. A 3D dataset of the lesional
M
region was obtained by combining the marked ROIs. A feature vector was formed from order statistics and 3D GLCM information obtained from each lesional VOI. The feature vector
ED
consisted of 636 different feature information items. Classification experiments were performed to distinguish the lesions as either periapical cysts or KCOT. Six different classifiers (k-NN,
PT
NB, DT, RF, NN, and SVM) were used for the model application experiments. Six different
CE
experiment groups were designed for the classification of periapical cyst and KCOT lesions obtained via CBCT.
AC
In the first experiment groups, all 50 feature vectors obtained from the CBCT dataset were used in training and testing. The ten-fold cross validation method was applied for model accuracy in those experiments. Different parameter combinations were used for each classifier to obtain the best classification performance. According to the experiment results, the SVM classifier achieved the best performance with 94.00% accuracy, and 94.00% F1 scores. According to the results, the use of features extracted from segmented CBCT datasets for classification of dental lesions as periapical cysts or KCOT was determined to be appropriate.
ACCEPTED MANUSCRIPT
The classifier performances were increased by reducing the size of the feature vector using the forward feature selection algorithm. The SVM classifier achieved good results in classifying periapical cyst and KCOT lesions obtained via the 50 segmented CBCT dataset. The best performance results were achieved as 100% accuracy, and 100% F1 using the SVM classifier with ten-fold cross validation for model validation and forward feature selection for feature
CR IP T
reduction. SVM achieved the best classification performance with 96.00% accuracy, and 96.00% F1 scores in the experiments in which a split sample validation method was used with a forward feature selection algorithm. SVM additionally achieved the best performance of 94.00% accuracy, and 93.88% F1 when LOOCV method was used with a forward feature
AN US
selection algorithm. 5. Conclusion
Endodontic lesions are amorphous in shape; thus, results of the segmentation implemented
M
by different oral radiologist may vary. Despite this limitation, a bounding box that surrounds the lesion in 3D space was considered for feature extraction of the segmented lesion in this study.
segmentation results.
ED
We are currently developing a hybrid segmentation method that will provide more accurate
PT
In contrast to previous studies, we conducted our experiments with a larger dataset consisting of images obtained via CBCT. We acquired more information for separating the
CE
lesions into the correct classes in terms of the number and diversity of the features extracted from this dataset. We expanded the domain of our study with multiple classifiers used in our
AC
experiments and different parameters that we attempted for each classifier. In sum, SVM achieved the best classification results for classifying dental periapical cysts
and KCOT lesions in volumetric images obtained using CBCT. In future work, we will expand the CBCT dataset with other types of dental pathologies, including histopathological diagnoses, and we will conduct multiclass classification experiments with the expanded dataset.
ACCEPTED MANUSCRIPT
Conflict of interest No conflicts of interest are declared. Acknowledgements The authors would like to thank Dr. Umit Cobanoglu, Dr. Celal Candirli, Dr. Nuray Yilmaz
CR IP T
Altintas, Dr. Yavuz Tolga Korkmaz, Oral Diagnosis and Radiology Department of Karadeniz Technical University for providing the dataset used in this study. The data referenced in this study were obtained and used with the permission of Karadeniz Technical University Clinical Research Ethics Committee.
[1]
AN US
References
S. White, M. Pharoah, Oral radiology: principles and interpretation, 6th ed., Mosby, 2013.
B. Basrani, Endodontic radiology, John Wiley & Sons, 2012.
[3]
D. Sarment, ed., Cone beam computed tomography, John Wiley & Sons, Inc, Chichester,
M
[2]
[4]
ED
UK, 2013. doi:10.1002/9781118769027.
W. De Vos, J. Casselman, G.R.J. Swennen, Cone-beam computerized tomography
PT
(CBCT) imaging of the oral and maxillofacial region: a systematic review of the literature.,
Int.
J.
Oral
Maxillofac.
Surg.
38
(2009)
609–25.
[5]
CE
doi:10.1016/j.ijom.2009.02.028. D.X. Ji, S.H. Ong, K.W.C. Foong, A level-set based approach for anterior teeth
AC
segmentation in cone beam computed tomography images., Comput. Biol. Med. 50 (2014) 116–28. doi:10.1016/j.compbiomed.2014.04.006.
[6]
R. Lloréns, V. Naranjo, F. López, M. Alcañiz, Jaw tissues segmentation in dental 3D CT images using fuzzy-connectedness and morphological processing, Comput. Methods Programs Biomed. 108 (2012) 832–843. doi:10.1016/j.cmpb.2012.05.014.
[7]
W.C. Scarfe, Imaging of maxillofacial trauma: evolutions and emerging revolutions.,
ACCEPTED MANUSCRIPT
Oral Surg. Oral Med. Oral Pathol. Oral Radiol. Endod. 100 (2005) S75-96. doi:10.1016/j.tripleo.2005.05.057. [8]
S. Kayipmaz, O.S. Sezgin, S.T. Saricaoglu, G. Çan, An in vitro comparison of diagnostic abilities of conventional radiography, storage phosphor, and cone beam computed
82. doi:10.1016/j.ejrad.2010.09.011. [9]
CR IP T
tomography to determine occlusal and approximal caries., Eur. J. Radiol. 80 (2011) 478–
W.C. Scarfe, A.G. Farman, P. Sukovic, Clinical applications of cone-beam computed tomography in dental practice., J. Can. Dent. Assoc. 72 (2006) 75–80.
[10]
S. Patel, New dimensions in endodontic imaging: Part 2. Cone beam computed
[11]
AN US
tomography., Int. Endod. J. 42 (2009) 463–75. doi:10.1111/j.1365-2591.2008.01531.x. T. Koivisto, W.R. Bowles, M. Rohrer, Frequency and distribution of radiolucent jaw lesions: A retrospective analysis of 9,723 cases., J. Endod. 38 (2012) 729–32. doi:10.1016/j.joen.2012.02.028.
N.R. Johnson, O.M. Gannon, N.W. Savage, M.D. Batstone, Frequency of odontogenic
M
[12]
cysts and tumors: a systematic review., J. Investig. Clin. Dent. 5 (2014) 9–14.
[13]
ED
doi:10.1111/jicd.12044.
H.M. Pinsky, S. Dyda, R.W. Pinsky, K.A. Misch, D.P. Sarment, Accuracy of three-
PT
dimensional measurements using cone-beam CT., Dentomaxillofac. Radiol. 35 (2006)
[14]
CE
410–6. doi:10.1259/dmfr/20987648. D.B. Prajapati, N.P. Desai, C.K. Modi, Influence of dental metallic artifact from
AC
multislice CT in the assessment of simulated mandibular lesions, in: 2012 Int. Conf. Commun. Syst. Netw. Technol., IEEE, 2012: pp. 198–202. doi:10.1109/CSNT.2012.50.
[15]
O. Aydemir, T. Kayikcioglu, Comparing common machine learning classifiers in lowdimensional feature vectors for brain computer interface applications, Int. J. Innov. Comput. Inf. Control. 9 (2013) 1145–1157.
[16]
U. Sevik, C. Kose, T. Berber, H. Erdol, Identification of suitable fundus images using automated quality assessment methods., J. Biomed. Opt. 19 (2014) 46006.
ACCEPTED MANUSCRIPT
doi:10.1117/1.JBO.19.4.046006. [17]
J.H.S. Simon, R. Enciso, J.-M. Malfaz, R. Roges, M. Bailey-Perry, A. Patel, Differential diagnosis of large periapical lesions using cone-beam computed tomography measurements and biopsy., J. Endod. 32 (2006) 833–7. doi:10.1016/j.joen.2006.03.008.
[18]
A. Flores, S. Rysavy, R. Enciso, K. Okada, Non-invasive differential diagnosis of dental
CR IP T
periapical lesions in cone-beam CT, in: 2009 IEEE Int. Symp. Biomed. Imaging From Nano to Macro, IEEE, 2009: pp. 566–569. doi:10.1109/ISBI.2009.5193110. [19]
J. Mikulka, E. Gescheidtová, M. Kabrda, V. Peřina, Classification of jaw bone cysts and necrosis via the processing of orthopantomograms, Radioengineering. 22 (2013).
K. Okada, S. Rysavy, A. Flores, M.G. Linguraru, Noninvasive differential diagnosis of
AN US
[20]
dental periapical lesions in cone-beam CT scans, Med. Phys. 42 (2015) 1653–1665. doi:10.1118/1.4914418. [21]
F. Abdolali, R.A. Zoroofi, Y. Otake, Y. Sato, Automated classification of maxillofacial
Comput.
Methods
M
cysts in cone beam CT images using contourlet transformation and Spherical Harmonics, Programs
Biomed.
139
(2017)
197–207.
[22]
ED
doi:10.1016/j.cmpb.2016.10.024.
D. Bhargava, A. Deshpande, M.A. Pogrel, Keratocystic odontogenic tumour (KCOT)--a
[23]
CE
0302-9.
PT
cyst to a tumour., Oral Maxillofac. Surg. 16 (2012) 163–70. doi:10.1007/s10006-011-
L. Barnes, W.H. Organization, I.A. for R. on Cancer, Pathology and genetics of head and
AC
neck tumours, IARC, 2005.
[24]
L.A. Gaitán-Cepeda, D. Quezada-Rivera, F. Tenorio-Rocha, E.R. Leyva-Huerta, Reclassification of odontogenic keratocyst as tumour. Impact on the odontogenic tumours prevalence., Oral Dis. 16 (2010) 185–7. doi:10.1111/j.1601-0825.2009.01625.x.
[25]
M.S. Tekkesin, S. Mutlu, V. Olgac, Expressions of bax, bcl-2 and ki-67 in odontogenic keratocysts (keratocystic odontogenic tumor) in comparison with ameloblastomas and radicular cysts, Turkish J. Pathol. 28 (2012) 49. doi:10.5146/tjpath.2012.01097.
ACCEPTED MANUSCRIPT
[26]
G.N. Hounsfield, Computerized transverse axial scanning (tomography): Part I. Description of system. 1973., Br. J. Radiol. 68 (1995) H166-72.
[27]
R.C. Gonzalez, R.E. Woods, Digital image processing (3rd edition), Prentice Hall, 2008.
[28]
R.M. Haralick, K. Shanmugam, I. Dinstein, Textural features for image classification, IEEE Trans. Syst. Man. Cybern. 3 (1973) 610–621. doi:10.1109/TSMC.1973.4309314. E. Ben Othmen, M. Sayadi, F. Fniaech, 3D gray level co-occurrence matrices for
CR IP T
[29]
volumetric texture classification, in: 3rd Int. Conf. Syst. Control, IEEE, 2013: pp. 833– 837. doi:10.1109/ICoSC.2013.6750953. [30]
J. Varjo, Implementing texture analysis software frame for magnetic resonance image
[31]
J. Han, M. Kamber, J. Pei, Data mining: concepts and techniques, 3rd ed., Morgan Kaufmann Publishers Inc., 2011.
[32]
AN US
data in MATLAB, Tampere University of Technology, 2014.
N. Friedman, D. Geiger, M. Goldszmidt, Bayesian network classifiers, Mach. Learn. 29
L. Rokach, Data mining with decision trees: Theory and applications, World scientific, 2007.
[34]
L.
Breiman,
ED
[33]
M
(1997) 131–163. doi:10.1023/A:1007465528199.
Random
forests,
Mach.
Learn.
45
(2001)
5–32.
R.P. Lippmann, Pattern classification using neural networks, IEEE Commun. Mag. 27
CE
[35]
PT
doi:10.1023/A:1010933404324.
(1989) 47–50. doi:10.1109/35.41401. C. Cortes, V. Vapnik, Support-vector networks, Mach. Learn. 20 (1995) 273–297.
AC
[36]
doi:10.1007/BF00994018.
[37]
I. Mierswa, M. Wurst, R. Klinkenberg, M. Scholz, T. Euler, YALE, in: Proc. 12th ACM SIGKDD Int. Conf. Knowl. Discov. Data Min. - KDD ’06, ACM Press, New York, New York, USA, 2006: p. 935. doi:10.1145/1150402.1150531.
[38]
M. Sokolova, G. Lapalme, A systematic analysis of performance measures for classification
tasks,
Inf.
Process.
Manag.
45
(2009)
427–437.
ACCEPTED MANUSCRIPT
doi:10.1016/j.ipm.2009.03.002. [39]
M. Shahrokh Esfahani, E.R. Dougherty, Effect of separate sampling on classification accuracy., Bioinformatics. 30 (2014) 242–50. doi:10.1093/bioinformatics/btt662.
[40]
P. Pudil, J. Novovičová, J. Kittler, Floating search methods in feature selection, Pattern Recognit. Lett. 15 (1994) 1119–1125. doi:10.1016/0167-8655(94)90127-9. I. Guyon, A. Elisseeff, An introduction to variable and feature selection, J. Mach. Learn. Res. 3 (2003) 1157–1182.
[42]
L. Yu, H. Liu, Efficient feature selection via analysis of relevance and redundancy, J. Mach. Learn. Res. 5 (2004) 1205–1224.
I. Nurtanio, Classifying cyst and tumor lesion using support vector machine based on
AN US
[43]
CR IP T
[41]
AC
CE
PT
ED
M
dental panoramic images texture features, Int. J. Comput. Sci. 40 (2013).
ACCEPTED MANUSCRIPT
Table 1. Confusion matrix created for binary classification Data class Classified as Positive Classified as Negative tp
fn
Negative
fp
tn
CR IP T
Positive
PT
ED
M
AN US
Fig. 1. Block diagram representing the outline of study
CE
Fig. 2. Marked lesional region of 3D CBCT data corresponding to a 2D plane (a). 3D
AC
demonstration of a lesional region as a result of the markings on each corresponding 2D plane section (b)
CR IP T
ACCEPTED MANUSCRIPT
AC
CE
PT
ED
M
AN US
Fig. 3. Vector of 13 different directions used in the calculation of the 3D GLCM
Fig. 4. Results of initial experiments of 636 features using ten-fold cross validation
AN US
CR IP T
ACCEPTED MANUSCRIPT
AC
CE
PT
ED
M
Fig. 5. Results of forward feature selection experiments using ten-fold cross validation
Fig. 6. Results of initial experiments of 636 features using split sample validation
AN US
CR IP T
ACCEPTED MANUSCRIPT
AC
CE
PT
ED
M
Fig. 7. Results of forward feature selection experiments using split sample validation
Fig. 8. Results of initial experiments of 636 features using leave-one-out cross validation
AN US
CR IP T
ACCEPTED MANUSCRIPT
AC
CE
PT
ED
M
Fig. 9. Results of forward feature selection experiments using leave-one-out cross validation
Fig. 10. Number of features selected using the forward feature selection algorithm