Segmentation and Classiﬁcation of Dental Caries in Cone Beam

Tomography Images Using Machine Learning and Image Processing

Luiz Guilherme Kasputis Zanini

1 a

, Izabel Regina Fischer Rubira-Bullen

2 b

and F

atima de Lourdes dos Santos Nunes

3 c

Department of Computer Engineering and Digital Systems, University of S

ao Paulo, S

ao Paulo, Brazil

Department of Surgery, Stomatology Pathology and Radiology, University of S

ao Paulo, Bauru, Brazil

School of Arts, Sciences and Humanities, University of S

ao Paulo, S

ao Paulo, Brazil

Keywords:

Image Processing, Machine Learning, Dental Caries, Segmentation, Classiﬁcation, CBCT, ICDAS.

Abstract:

Dental caries are caused by bacterial action that demineralizes tooth enamel and dentin. It is a serious threat

to oral health and potentially leads to inﬂammation and tooth loss if not adequately treated. Cone Beam

Computed Tomography (CBCT), a three-dimensional (3D) imaging technique used in dental diagnosis and

surgical planning, can potentially contribute to detection of caries. This study aims at developing a com-

putational method to segment and classify caries in CBCT images. The process involves data preparation,

segmentation of caries regions, extraction of relevant features, feature selection, and training machine learning

algorithms. We evaluated our method performance considering different stages of caries severity based on the

International Caries Detection and Assessment System scale. The best results were achieved using a Gaussian

ﬁlter with a multimodal threshold with a convex hull for the region of interest segmentation, feature selection

via Random Forest, and classiﬁcation using a model based on k-nearest neighbors algorithm. We achieved

outcomes with an accuracy of 86.20%, a F1-score of 86.18%, and a sensitivity of 83.35% in multiclass clas-

siﬁcation. These results show that our approach contributes to the early segmentation and classiﬁcation of

dental caries, thereby improving oral health outcomes and treatment planning.

1 INTRODUCTION

Dental caries compose a condition that signiﬁcantly

impacts oral health, which can cause degradation of

enamel and dentin due to the action of bacteria present

in dental plaque. If not treated, this disease can

progress into the interior of the tooth, contacting the

dental pulp, where nerves and blood vessels are lo-

cated, leading to inﬂammation and potential tooth loss

(Rathee and Sapra, 2023).

Cone Beam Computed Tomography (CBCT) is

a technique that allows the acquisition of three-

dimensional (3D) radiographic images. This imaging

modality is widely employed for diagnostic purposes,

ranging from dental root canal treatments to the as-

sessment of tooth demineralization, issues related to

low bone density, and even the planning of surgical

procedures (Setzer et al., 2017). Furthermore, a po-

https://orcid.org/0000-0002-8344-3178

https://orcid.org/0000-0002-5069-9433

https://orcid.org/0000-0003-0040-0752

tential application of CBCT is caries diagnosis, es-

pecially in proximal caries, which is a dental lesion

between adjacent teeth, often hidden from direct view

and better indicated by dental X-rays (Felemban et al.,

2020). As such, it is vital to identify caries lesions

within CBCT images, and once identiﬁed, they should

be incorporated into the ﬁnal CBCT examination re-

ports.

Several studies have explored various methods for

locating and classifying diseases in computed tomog-

raphy images. For example, a study conducted by

(Ezhov et al., 2021) used the U-Net network (Ron-

neberger et al., 2015) to detect dental caries. In (Chen

and Zhang, 2017), a segmentation process was pro-

posed, employing advanced image processing tech-

niques along with a threshold to segment the affected

area. Meanwhile, the study by (Ahmed et al., 2017)

combined unsupervised machine learning techniques,

such as the K-means clustering algorithm, to gener-

ate segmentations. However, these approaches mainly

concentrated on discerning the disease’s presence or

absence. Upon our analysis, we observed few studies

428

Zanini, L., Rubira-Bullen, I. and Nunes, F.

Segmentation and Classiﬁcation of Dental Caries in Cone Beam Tomography Images Using Machine Learning and Image Processing.

DOI: 10.5220/0012365300003657

Paper published under CC license (CC BY-NC-ND 4.0)

In Proceedings of the 17th International Joint Conference on Biomedical Engineering Systems and Technologies (BIOSTEC 2024) - Volume 2, pages 428-435

ISBN: 978-989-758-688-0; ISSN: 2184-4305

conducting classiﬁcation after segmentation, as well

as a notable absence of standardization in lesion clas-

siﬁcation.

Therefore, given this gap, this study aimed to seg-

ment and classify severity levels of lesions. Fig-

ure 1 illustrates the variation in lesions at different

stages, increasing the complexity of diagnosing these

affected areas. Accurately identifying the level of le-

sion severity plays a fundamental role in applying the

appropriate treatment to patients with carious teeth.

Our method segments the affected regions and

subsequently classifying them based on the ICDAS

score. To achieve this purpose, we employed a set of

image processing techniques speciﬁcally adapted for

images acquired through CBCT. The second step in-

volves feature extraction from these regions. Then,

we proceeded with the selection of the most relevant

features, and ﬁnally, we trained machine learning al-

gorithms to perform the classiﬁcation. Therefore, the

contributions of this work:

• A set of image processing techniques developed

speciﬁcally to segment regions affected by caries

in CBCT images;

• Classiﬁcation of the region of interest based on a

standard score used by dentists (ICDAS);

• Extraction of relevant features from the images;

• Evaluation of machine learning algorithms in or-

der to classify different levels of the disease.

These contributions represent signiﬁcant advance-

ments in the development of an effective system for

the segmentation and classiﬁcation of lesioned re-

gions, particularly in images obtained through CBCT.

2 RELATED WORK

In the literature, there are studies that aimed to clas-

sify and segment caries in different imaging modali-

ties. In this section, we highlight approaches involv-

ing feature extraction and techniques similar to those

used in our approach.

The research studies (Jusman et al., 2022a), (Jus-

man et al., 2022b), (Singh and Sehgal, 2017) used

features extracted from x-ray images with the pri-

mary goal of classifying dental cavities. One of the

methods applied the Gray-Level Co-occurrence Ma-

trix (GLCM) to extract texture details, aiming to spot

variations in grayscale shades within the images (Jus-

man et al., 2022a). In (Jusman et al., 2022b), the au-

thors explored the extraction of shape characteristics

using Hu Moments to achieve insights about the tooth

geometry and categorize images with lesions. Fur-

thermore, the study (Singh and Sehgal, 2017) inves-

tigated the classiﬁcation of cavities using the Radon

transformation and Discrete Cosine Transform to ex-

tract patterns about pixel intensities in different direc-

tions.

In a study utilizing periapical images (Geetha

et al., 2020), the authors segmented the affected area

and extracted features from this region, referred to as

texture-based analysis. These features included pa-

rameters such as contrast, correlation, energy, homo-

geneity, mean, entropy, and root mean square. The

study (Sornam and Prabhakaran, 2017) used periapi-

cal images, and GLCM was applied to extract fea-

tures from grayscale images, with a linearly adaptive

optimization using a particle swarm optimization al-

gorithm in conjunction with a neural network to opti-

mize the learning rate parameter. In the study (Datta

et al., 2019), the Particle Swarm Optimization (PSO)

algorithm was employed to divide the image into mul-

tiple regions or segments, enabling the identiﬁcation

of the intersection of two lines, which detects restora-

tions and caries in periapical images.

Panoramic radiographs served as the base for the

studies (Al Kheraif et al., 2019), (Verma et al., 2020),

(Lakshmi and Chitra, 2020). In (Al Kheraif et al.,

2019), a cavity detection method was implemented

using Sobel edge detection and a deep convolutional

neural network (CNN). In (Verma et al., 2020), a

CNN was employed to extract image features, and

a Support Vector Machine (SVM) was used to clas-

sify images as normal or abnormal based on these

features, including Haralick and Hu moments. The

study (Al Kheraif et al., 2019) applied various image

enhancement techniques and compared segmentation

approaches with the hybrid graph cut method.

Many of these studies do not assume the varia-

tion in caries severity levels, limiting themselves to

only identifying the presence or absence of the dis-

ease. Furthermore, these approaches tend to focus

on isolated techniques without integrating informa-

tion about the nuances of gray levels of the image and

tooth shapes. Based on the analyses and discussions

of the presented studies, we propose an approach that

uses image processing segmentation and feature ex-

traction to classify the intensity of the affected area.

3 MATERIAL AND METHODS

As depicted in Figure 2, our approach is split into sev-

eral stages. Firstly, a preprocessing step is responsi-

ble for preparing the initial image by removing noise

and enhancing relevant features. Next, we highlight

Segmentation and Classiﬁcation of Dental Caries in Cone Beam Tomography Images Using Machine Learning and Image Processing

429

Figure 1: Examples of different levels of caries using the score International Caries Detection and Assessment System (IC-

DAS).

CBCT

exam

Stage 1 Stage 3

Pre processing

Segmentation

Feature

Extraction

Dimensionality

reduction

Stage 5

Training Models

Stage 2 Stage 4

0 ICDAS

1 ICDAS

Score

2 ICDAS

3 ICDAS

4 ICDAS

Figure 2: Steps of the proposed approach for image classiﬁcation.

the lesion region segmentation stage, where we iden-

tify the affected area. The third stage performs the

feature extraction, where essential information is ob-

tained from both the segmentation and the original

image. Then, a feature selection stage identiﬁes the

most relevant features. Lastly, machine learning al-

gorithms are trained to classify the caries considering

different severy levels.

3.1 Materials

The dataset used in this study contains 493 images

obtained from CBCT scans, each one classiﬁed by ex-

perts according to the ICDAS. ICDAS is an adopted

system by oral healthcare professionals for assessing

and categorizing the state of dental caries (Gugnani

and Pandit, 2011). This system provides standard-

ization in detecting and recording dental caries le-

sions, assigning a classiﬁcation that represents differ-

ent stages of the disease. Each image has a score clas-

sifying the image according to Figure 1. The score

ranges from zero, indicating the absence of caries, to

1, 2, 3, and 4 values indicating greater severity of le-

sions.

The distribution of images about ICDAS classiﬁ-

cations applied on an image is as follows: 66 images

were classiﬁed as ICDAS 0, indicating the absence

of caries, 32 images received the ICDAS 1 classiﬁ-

cation, 50 images were classiﬁed as ICDAS 2, 151

images were classiﬁed as ICDAS 3, and 194 images

were classiﬁed as ICDAS 4. The image base used

in the experiments of this project was provided by the

Faculty of Dentistry of Bauru of the University of S

Paulo and approved by the Ethics Committee.

3.2 Methods

During the initial image processing phase (Figure 3),

the focus was enhancing visual quality and sharpen-

ing image details. Two techniques were employed:

ﬁrst, a Gaussian ﬁlter was applied to reduce noise

originating from radiation capture and detection. Af-

ter, Gamma correction was used to boost contrast,

thereby enhancing details within the images, making

them easier to analyze and interpret.

The next stage of the process involved the seg-

mentation of dental structures, which begins with an

application of multimodal thresholding aiming at dis-

tinguishing three tooth structures. We search for two

local minima within the image histogram. Utilizing

these minima, we employed the multimodal thresh-

olding method to segment the resulting structures.

Lower values correspond to the background, interme-

diate values are related to dentin and higher values

HEALTHINF 2024 - 17th International Conference on Health Informatics

430

Figure 3: Preprocessing techniques and segmentation techniques to enhance CBCT images.

indicate enamel.

After calculating the segmentation’s contour and

identifying the points covering the region, we applied

the convex hull method. This technique discerns the

intersection of all convex sets within the segmenta-

tion. Subsequently, we merged the original segmenta-

tion with the convex hull area, giving rise to the poten-

tial regions with caries, forming the red area (Figure

3). The carie region is delimited by this procedure in

all the images that compose the CBCT tooth volume.

The feature extraction stage begins with comput-

ing Haralick features, Hu Moments, and shape fea-

tures. Haralick features aim to capture information

about the dental surface’s texture. These features are

achieved by calculating the GLCM from the prepro-

cessed original image, which represents the spatial re-

lationship between pixels with similar gray levels in

the image. The features extracted from this process

include numerical values that describe different as-

pects of the tooth’s texture, such as energy, contrast,

correlation, and homogeneity, among others (L

ofstedt

et al., 2019).

Hu Moments are attributes designed to record

shape or geometry-related information in the image

(Prokop and Reeves, 1992). These parameters are

derived from image segmentation, which can be ex-

tended to the tooth’s structure and are characterized

by their scale, rotation, and reﬂection invariance.

Therefore, in addition to textural information, precise

data about the tooth’s shape can be obtained.

Lastly, shape features obtained after caries seg-

mentation focuses on analyzing the geometry of the

region of interest (caries), containing attributes such

as area, perimeter, circularity, and eccentricity, pro-

viding a description of the overall shape of the region

(Mingqiang et al., 2008).

Data normalization was performed using the

”min-max” technique, which ensures that the features

have the same relative importance during analysis, es-

pecially when the original data has different scales.

The equation 1 represents this procedure, where x is

the value of the feature to be normalized, and x

′

is the

result of this operation.

′

x − min(x)

max(x) − min(x)

(1)

The next step involves the reduction of dimension-

ality in which three distinct methods are applied to

form different sets of features.

The ﬁrst method is Principal Component Anal-

ysis (PCA), which aims to reduce the dimensional-

ity of the dataset by performing feature fusion, trans-

forming it into a simpliﬁed new set (Bro and Smilde,

2014). PCA accomplishes this reduction by identify-

ing linear combinations of the original features that

maximize data variance while preserving relevant in-

formation as much as possible.

The second method adopts the Random Forest

(RF) technique to evaluate the importance of each fea-

ture in order to select the more relevant ones. RF is a

machine learning algorithm capable of estimating the

relevance of variables by minimizing the impurity of

their nodes (Breiman, 2001). At the end of this pro-

cess, each feature receives a relevance score, which

allows selecting of the most informative features for

subsequent analysis.

Finally, the Chi-squared test considers the rela-

tionship between target variables and the extracted

features. This statistical test determines if there is

a statistically signiﬁcant relationship between target

variables and the used features, allowing the identi-

ﬁcation of the most relevant features for the speciﬁc

task.

In the last phase of this process, we evaluate

a variety of inductive algorithms, which encompass

RF, Naive Bayes (NB), K-Nearest Neighbors (KNN),

Support Vector Machine (SVM), logistic regression

(LR), and XGBoost (XG). This variety of algorithms

offers a wide range of techniques, allowing us to

explore different perspectives and approaches. The

Segmentation and Classiﬁcation of Dental Caries in Cone Beam Tomography Images Using Machine Learning and Image Processing

431

multi-class classiﬁcation approach was chosen to en-

hance our analysis, allowing for more precise dif-

ferentiation between various categories within the

dataset. Thus, our results beneﬁt from increased ro-

bustness and accuracy, helping us to identify the most

suitable model for our speciﬁc analytical task, while

considering the unique challenges and requirements

of dental structure image analysis.

3.3 Evaluation Process

The evaluation process involves using different com-

binations of techniques to consider four distinct

datasets. The ﬁrst dataset contains all the original

features. Principal Component Analysis (PCA) is ap-

plied to reduce dimensionality and forms the second

dataset. The third dataset is created by using RF al-

gorithm to select features, and the fourth dataset is

created using the Chi-squared technique to select fea-

tures.

We set the PCA algorithm to retain features that

explain up to 90% of the data variance, which is

a common approach to reduce dimensionality while

preserving most of the information. In the case of

feature selection with the RF algorithm, we chose pa-

rameters to maintain features to represent up to 90%

importance of features that are provided by the al-

gorithm. Finally, the Chi-squared test generates a p-

value, and features with a p-value less than 0.05 were

considered.

We employed a stratiﬁed 5-fold cross-validation

method to assess the model’s performance with ﬁve

distinct metrics. In this process, the dataset is divided

into ﬁve equal segments, with each segment repre-

senting 20% of the total data. The stratiﬁed approach

guarantees that the distribution of ﬁve classes remains

across these ﬁve segments. During each iteration, four

of these stratiﬁed segments (80% of the data) are uti-

lized for training, while the remaining ﬁfth segment,

containing around 20% of the data, takes its turn as

the test set in a rotating technique.

The metrics generated for performance evaluation

include precision, speciﬁcity, recall, F1, and accuracy.

These metrics provide a comprehensive assessment of

how well the model performs the classiﬁcation task,

helping to determine its effectiveness and ability to

handle different scenarios.

To evaluate the best model compared to other

models, the McNemar test was employed to identify

if there is a signiﬁcant difference between the best

results obtained with each approach. The McNemar

test is a statistical tool for determining whether two

different approaches have statistically distinct perfor-

mances, aiding in selecting the strategy (Japkowicz

and Shah, 2015). The calculation of the χ

statistic is

depicted in Equation 2.

(|b − c|− 1)

b + c

(2)

In Equation 2, the variable b indicates how many

times the ﬁrst approach was correct while the second

one was incorrect, whereas variable c represents how

many times the second approach was correct while

the ﬁrst one was incorrect. These values are obtained

by comparing the predictions of each of the models.

The χ

statistic, computed using Equation 2, is then

compared to a chi-squared distribution to derive the p-

value, allowing for the evaluation of the statistical sig-

niﬁcance of the disparities between the approaches.

4 RESULTS AND DISCUSSION

In this section, we conducted a comparative analysis

of the generated data, inductive algorithms, and re-

lated studies and concluded with a ﬁnal discussion.

4.1 Comparison Between Data

In the feature selection process, there was a reduc-

tion of the dataset from 100% of the total features to

72% (using RF), 28% (using Chi-squared), and 25%

(using PCA). It is important to note that PCA is an

unsupervised method in which feature selection does

not consider class labels, unlike Chi-squared and RF,

which are supervised methods.

The performance of the six inductive algorithms

on four different datasets is presented in Table 1. The

evaluation metrics include accuracy, which quantiﬁes

the proportion of correct model predictions, and F1-

score, which combines precision and recall. It is im-

portant to note that none of the classiﬁers performed

best when using the Chi-squared technique among the

analyzed datasets. This suggests that the Chi-square

approach may not be the most appropriate when the

data does not directly represent the relationship be-

tween the features and the target variable, indicating

possible dependencies among the features.

When evaluating the four datasets (Origin, PCA,

Chi, RF), it was identiﬁed that the original set ob-

tained the best results in three of the six algorithms

tested according to Table 1. The feature selection of

the RF algorithm stood out in two of the six induc-

tive algorithms. This pattern suggests that a signiﬁ-

cant portion of the extracted features were essential to

achieving the best results with both the RF algorithm

and the original dataset, considering that RF reduced

the dataset to 72% of its original size. However, the

HEALTHINF 2024 - 17th International Conference on Health Informatics

432

Table 1: Results obtained from the six algorithms varying across four datasets obtained from feature selection with accuracy

and average F1-score metrics.

Data

XG SVM Logistic Regression KNN NB RF

Accuracy F1-Score Accuracy F1-Score Accuracy F1-Score Accuracy F1-Score Accuracy F1-Score Accuracy F1-Score

Origin 81.75% 81.38% 80.94% 80.72% 54.76% 52.58% 80.73% 80.55% 43.01% 42.31% 76.26% 75.34%

PCA 74.43% 73.94% 79.30% 79.42% 46.65% 41.27% 76.67% 76.31% 54.79% 53.23% 68.78% 68.03%

Chi 75.46% 74.54% 75.05% 74.18% 52.72% 47.67% 76.47% 76.03% 45.03% 43.63% 64.49% 62.64%

RF 80.12% 79.63% 81.13% 81.07% 47.23% 40.81% 86.20% 86.18% 49.49% 48.83% 75.86% 74.91%

application of Chi-square and PCA to the distribution

of characteristics resulted in a considerable reduction

in the size of the datasets: 28% with Chi-square and

25% with PCA. This reduction in dataset size nega-

tively impacted performance.

4.2 Comparison Between Inductive

Algorithms

Figure 4 illustrates the variation in the performance

using cross-validation. The performance measure was

obtained through stratiﬁed 5-fold cross-validation. In

this representation of the data, we highlight the previ-

ous best results in Table 1 to the comparison between

the models. Furthermore, we highlight two metrics to

assess the overall performance of the classes.

Figure 4: BoxPlot of accuracy and f1-score for the best 6

algorithms.

The SVM and XG classiﬁers demonstrated simi-

lar performance f1-score 81.75% and 81.07%. How-

ever, SVM revealed the presence of outliers, as evi-

denced in Figure 4. NB, LR, and RF classiﬁers ex-

hibited lower performance. This difference in perfor-

mance can be attributed to the inherent characteristics

of these classiﬁers.

Finally, the KNN algorithm demonstrated the

best performance, standing out as the most effective

choice among the evaluated models. KNN, despite its

simplicity, is a non-parametric classiﬁer, meaning it

makes no speciﬁc assumptions about the data distri-

bution and can manage non-linear relationships. Be-

sides that, KNN does not require complex adaptations

for multiclass classiﬁcation, as it can perform this task

due to cluster choice, making it a robust and effective

option (Jiang et al., 2007).

4.3 Comparison over the Classiﬁcation

Figure 5 shows four metrics over the ﬁve classes of

the problem. Sensitivity, which assesses the ability

to identify positive cases accurately, showed lower

performance when compared to other ICDAS stages,

with rates ranging between 60% and 70% for caries

classiﬁed as ICDAS 1 (Figure 5). These results point

to a signiﬁcant challenge in detecting caries in the

early stages, which are critical for the effectiveness of

preventive measures. It is also relevant to note that the

models demonstrated more satisfactory performance

when classifying more advanced caries stages, such

as ICDAS 3 and ICDAS 4, highlighting a superior

ability to identify caries in these scenarios.

The XG model showed the best precision in clas-

sifying caries as ICDAS 1 (Figure 5). However,

when expanding the analysis to different levels of

caries severity, it was observed that XG achieved re-

sults similar to KNN, maintaining competitive perfor-

mance across different stages of caries.

The F1-score, a metric that balances sensitiv-

ity (recall) and precision, demonstrated that KNN

achieved the best results across different caries stages

(Figure 5). This analysis indicates that, although pre-

cision and sensitivity may vary between classiﬁers

considering different severity levels of caries, F1-

score considers both aspects and provides a balanced

performance assessment.

The analysis between classiﬁers was conducted

using the McNemar test to assess the difference in

performance between the best model (KNN) and the

other models. This test was performed to determine

if, in each test fold, the models showed a statistically

signiﬁcant difference in performance compared to the

best model.

The results showed that, compared to KNN, all

models had a p-value below 0.05, indicating a sig-

niﬁcant difference in performance, except XG. In the

case of XG, only one of the data divisions had a p-

value of 0.103 in the comparison between XG and

KNN, while the other four divisions had p-values be-

low 0.05. This means there is only a difference be-

tween XG and KNN, and in general, KNN outper-

forms all other models and stands out as the best op-

tion.

Segmentation and Classiﬁcation of Dental Caries in Cone Beam Tomography Images Using Machine Learning and Image Processing

433

Figure 5: Comparison of best six algorithms between metrics over the ﬁve classes ICDAS.

4.4 Comparison Studies

Table 2 shows a comparison with other studies. It is

challenging due to notable differences in data, tech-

niques used, and types of images analyzed. Each of

these studies manages speciﬁc characteristics, such as

the number of classes to categorize caries lesions at

different levels. Additionally, some studies focus on

identifying caries regions, while others may not in-

clude this step in their approach. These variations re-

ﬂect the complexity of dental image analysis and the

diversity of approaches available in the literature.

Table 2: Comparison between studies.

Study Accuracy Sensitivity Classes Image

(Jusman et al., 2022b) 96.10% - 4 Bitewing

(Ezhov et al., 2021) - 72.85% 2 CBCT

(Zhu et al., 2022) 93.61% 86.01% 4 Panoramic

(Imak et al., 2022) 99.13% 98.00% 2 Periapical

(Ramana Kumari et al., 2022) 93.67% 94.66% 2 X-ray

When comparing our study to a two-class study,

as presented in Table 2, it is important to note that we

need to adapt our results by combining ICDAS lev-

els 1, 2, 3, and 4 into class 1, If we calculated the

metrics of these binary conditions, the best algorithm

achieves an accuracy of 99.59%. Notably, the KNN

model exhibits a high sensitivity of 98.46% for IC-

DAS 0 (Figure 5), highlighting that in this context,

the primary challenge is distinguishing between vari-

ous levels of caries. The choice of model evaluation

metrics is crucial to align with the speciﬁc study ob-

jectives and ensure the derivation of meaningful and

relevant results.

The study most similar to ours is the research con-

ducted by (Ezhov et al., 2021). In this study, the au-

thors utilized a dataset of 4,398 teeth to train a model

that incorporates a contextual area for caries segmen-

tation. Their model achieved a speciﬁcity of 99.53%

and a sensitivity of 72.85% in binary caries segmenta-

tion. However, this study does not provide a compre-

hensive set of performance metrics and primarily em-

phasizes broader dental applications, such as identi-

fying periodontal bone and periapical lesions. In con-

trast, our model excels in terms of sensitivity within

the broader metrics, achieving an 83.35% sensitivity

in a multi-class problem.

5 CONCLUSION

A limitation of our approach is the limited ability to

classify more than two caries lesions in the same im-

age. The current system was not developed to address

this speciﬁc situation, as classiﬁcation is performed

on individual images. However, this limitation is par-

ticular and our method can be adapted to consider this

topic.

The primary goal of this project is to segment

regions with caries and subsequently classify them,

which was achieved through the application of im-

age processing and machine learning techniques. The

performance metrics of the proposed approach have

shown promising results, achieving an accuracy of

86.20%, an f1-score of 86.18%, and a sensitivity of

83.35% in the context of multiclass classiﬁcation us-

ing ICDAS.

For future work, we intend to compare the current

model to deep learning approaches, and the applica-

tion will be extended to include knowledge transfer

(few-shot learning), enabling an even broader range

of applications and improving the accuracy and effec-

tiveness of the system.

ACKNOWLEDGEMENTS

This work is supported in part by the Brazilian Na-

tional Council for Scientiﬁc and Technological Devel-

opment (CNPq grant number 307710/2022-0), in part

by the Coordenac¸

ao de Aperfeic¸oamento de Pessoal

de N

ıvel Superior - Brasil (CAPES) - Finance Code

001, Brazil, and Ita

u Unibanco S.A. through the PBI

program of the Centro de Ci

encia de Dados (C

of Escola Polit

ecnica at Universidade de S

ao Paulo,

and S

ao Paulo Research Foundation (FAPESP) – Na-

tional Institute of Science and Technology – Medicine

HEALTHINF 2024 - 17th International Conference on Health Informatics

434

Assisted by Scientiﬁc Computing (INCT-MACC) –

grant 2014/50889-7.

REFERENCES

Ahmed, S., Saifuddin, K. M., Ahmed, A. S., Aowlad Hos-

sain, A., and Iqbal, M. T. (2017). Identiﬁcation and

volume estimation of dental caries using CT image.

ICTP, pages 48–51.

Al Kheraif, A. A., Wahba, A. A., and Fouad, H. (2019). De-

tection of dental diseases from radiographic 2d dental

image using hybrid graph-cut technique and convolu-

tional neural network. Measurement, 146:333–342.

Breiman, L. (2001). Random Forests. Machine Learning,

45(1):5–32.

Bro, R. and Smilde, A. K. (2014). Principal component

analysis. Analytical methods, 6(9):2812–2831.

Chen, R. and Zhang, H. (2017). Large-scale 3D Reconstruc-

tion with an R-based Analysis Workﬂow. Proceedings

Big Data Computing, pages 85–93.

Datta, S., Chaki, N., and Modak, B. (2019). A Novel Tech-

nique to Detect Caries Lesion Using Isophote Con-

cepts. IRBM, 40(3):174–182.

Ezhov, M., Gusarev, M., Golitsyna, M., Yates, J. M.,

Kushnerev, E., Tamimi, D., Aksoy, S., Shumilov, E.,

Sanders, A., and Orhan, K. (2021). Clinically appli-

cable artiﬁcial intelligence system for dental diagnosis

with CBCT. Scientiﬁc Reports, 11(1).

Felemban, O. M., Loo, C. Y., and Ramesh, A. (2020). Ac-

curacy of Cone-beam Computed Tomography and Ex-

traoral Bitewings Compared to Intraoral Bitewings in

Detection of Interproximal Caries. The Journal of

Contemporary Dental Practice, 21(12):1361–1367.

Geetha, V., Aprameya, K. S., and Hinduja, D. M. (2020).

Dental caries diagnosis in digital radiographs using

back-propagation neural network. Health Inf Sci Syst,

8(1):8.

Gugnani, N. and Pandit, I. (2011). International caries de-

tection and assessment system (ICDAS): A new con-

cept. International Journal of Clinical Pediatric Den-

tistry, 4(2):93–100.

Imak, A., Celebi, A., Siddique, K., Turkoglu, M., Sengur,

A., and Salam, I. (2022). Dental Caries Detection

Using Score-Based Multi-Input Deep Convolutional

Neural Network. IEEE Access, 10:18320–18329.

Japkowicz, N. and Shah, M. (2015). Performance evalua-

tion in machine learning. Machine Learning in Radi-

ation Oncology: Theory and Applications, pages 41–

56.

Jiang, L., Cai, Z., Wang, D., and Jiang, S. (2007). Survey

of improving k-nearest-neighbor for classiﬁcation. In

Fourth international conference on fuzzy systems and

knowledge discovery (FSKD 2007), volume 1, pages

679–683. IEEE.

Jusman, Y., Widyaningrum, A., and Puspita, S. (2022a). Al-

gorithm of Caries Level Image Classiﬁcation Using

Multilayer Perceptron Based Texture Features. Cy-

berneticsCom, pages 168–173.

Jusman, Y., Widyaningrum, A., Tyassari, W., Puspita, S.,

and Saleh, E. (2022b). Classiﬁcation of Caries X-Ray

Images using Multilayer Perceptron Models Based

Shape Features. ICITDA, pages 1–6.

Lakshmi, M. M. and Chitra, P. (2020). Classiﬁcation of

Dental Cavities from X-ray images using Deep CNN

algorithm. ICOEI, pages 774–779.

ofstedt, T., Brynolfsson, P., Asklund, T., Nyholm, T., and

Garpebring, A. (2019). Gray-level invariant haralick

texture features. PloS one, 14(2):e0212110.

Mingqiang, Y., Kidiyo, K., Joseph, R., et al. (2008). A

survey of shape feature extraction techniques. Pattern

recognition, 15(7):43–90.

Prokop, R. J. and Reeves, A. P. (1992). A survey of

moment-based techniques for unoccluded object rep-

resentation and recognition. CVGIP: Graphical Mod-

els and Image Processing, 54(5):438–460.

Ramana Kumari, A., Nagaraja Rao, S., and Ramana Reddy,

P. (2022). Design of hybrid dental caries segmen-

tation and caries detection with meta-heuristic-based

ResneXt-RNN. Biomedical Signal Processing and

Control, 78:103961.

Rathee, M. and Sapra, A. (2023). Dental Caries. In Stat-

Pearls. StatPearls Publishing, Treasure Island (FL).

Ronneberger, O., Fischer, P., and Brox, T. (2015). U-net:

Convolutional networks for biomedical image seg-

mentation. pages 234–241.

Setzer, F. C., Hinckley, N., Kohli, M. R., and Karabucak,

B. (2017). A Survey of Cone-beam Computed To-

mographic Use among Endodontic Practitioners in the

United States. J Endod, 43(5):699–704.

Singh, P. and Sehgal, P. (2017). Automated caries detection

based on Radon transformation and DCT. ICCCNT,

pages 1–6.

Sornam, M. and Prabhakaran, M. (2017). A new lin-

ear adaptive swarm intelligence approach using back

propagation neural network for dental caries classiﬁ-

cation. ICPCSI, pages 2698–2703.

Verma, D., Puri, S., Prabhu, S., and Smriti, K. (2020).

Anomaly detection in panoramic dental x-rays using

a hybrid Deep Learning and Machine Learning ap-

proach. TENCON, pages 263–268.

Zhu, H., Cao, Z., Lian, L., Ye, G., Gao, H., and Wu,

J. (2022). CariesNet: a deep learning approach for

segmentation of multi-stage caries lesion from oral

panoramic X-ray image. Neural Comput Appl, pages

1–9.

Segmentation and Classiﬁcation of Dental Caries in Cone Beam Tomography Images Using Machine Learning and Image Processing

435