Oral Diseases Recognition Based on Photographic Images

Mazin S. Mohammed

, Salah Zrigui

1,3

and Mounir Zrigui

University of Monastir, Research Laboratory in Algebra,

Numbers Theory and Intelligent sys-tem (RLANTIS), Monstir 5019, Tunisia

University of Al-Mosul, University of the Presidency, Nineveh, Iraq

Lig Laboratory, Grenoble, France

Keywords: Oral Disease Classification, Dental Caries Detection, Dental Images, Deep Learning.

Abstract: Recently, the automation diagnosis process of dental caries plays a critical role in medical applications. This

paper presents a new dataset of photo-graphic images for six different types of oral diseases. The dataset is

gathered and labelled by professional medical operators in the dentistry field. We use the collected dataset to

train a binary classifier to determine whether the region of interests (ROI) needs detection or not inside the

input image. Then, we train a detector to detect and localize the required ROI. Finally, we use the detected

regions to train a CNN network by adopting transfer learning technique to classify various kinds of teeth

diseases. With this model, we obtained an almost 93 % accuracy by modifying and re-training the pre-trained

model VGG19.

1 INTRODUCTION

This increasing global vulnerability to diseases has

left health care systems worldwide strained. To

protect against the spread of disease, hospitals, clinics

and nearly all types of medical facilities had to adhere

to several protective guidelines. This led to a

significant decrease in the number of patients that can

be treated at any given moment. In response,

researchers, more specifically researchers in the field

of artificial intelligence have been innovating and

proposing novel methods and technologies to ensure

safe diagnosis and treatment with minimal direct

contact. One of the most prominent fields for such

innovation is the automated diagnosis of dental

imagery (Araújo et al., 2023).

AI techniques have been used successfully in

various types of disciplines such as nature language

processing (Merhbene, Zouaghi and Zrigui, 2010;

Mahmoud and Zrigui, 2019), computer vision

(Mansouri, Charhad and Zrigui, 2017; Farhani,

Terbeh and Zrigui, 2019; Daood, AL-Saegh and

Mahmood, 2023), speech recognition (Bellagha and

Zrigui, 2020; Slimi et al., 2020; Amari et. Al., 2022),

biometrics, smart home applications (Alhafidh et al.,

2018), medical imaging, healthcare, robotics,

banking & finance, agriculture, military & defence,

marketing & advertising, and even oil discovery &

gas exploration. Lately, computer vision has been

used as an efficient tool in medical applications to

offer an accurate diagnosis and avoid errors in human

judgement. The use of artificial intelligence in

dentistry appears to has a great potential and it is

expected to play a vital role in the future of dental

health-care and oral diseases diagnosis.

Deep learning strategies have achieved remarkable

progress in understanding and analysing dental images.

Some neural network architectures such as

Convolutional Neural Networks (CNNs) lend

themselves naturally to exploit the availability of X-ray

and photographic images dataset to perform teeth

segmentation, classification, numbering, and lesions

detection. With the long waiting time to receive dental

care and the importance of an early diagnosis. We

decided to build a tool that helps the average person get

an early evaluation of his dental state. In this paper we

build a dental care detection and classification system

that can provide an early diagnosis from a simple

picture taken via any smartphone. The system takes as

an input a dataset that is comprised of photographic

images collected from some local clinics with the help

of medical team of specialist dentists. The remainder

of the paper is structured as follows. Section 2

examines prior research works. Section 3 presents the

material data and the proposed methodology. Section 4

showcases the experimental results. Lastly, we present

486

Mohammed, M., Zrigui, S. and Zrigui, M.

Oral Diseases Recognition Based on Photographic Images.

DOI: 10.5220/0012361500003636

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

In Proceedings of the 16th International Conference on Agents and Artiﬁcial Intelligence (ICAART 2024) - Volume 3, pages 486-493

ISBN: 978-989-758-680-4; ISSN: 2184-433X

our conclusions in Section 5.

2 LITERATURE REVIEW

In this section we will review some of the methods

that have been proposed to automate the diagnosis

process of dental diseases. As we mentioned before,

all the previous researches were either based on X-ray

images or photographic images. For example, the

authors in (Liu et al., 2019) proposed intelligent

dental Health-IoT system which was implemented on

smart hardware. Mobile phone was used as a terminal

to capture images in order to perform the diagnosis.

MASK R-CNN was used to perform the detection

process by applying the training on 12,6000 collected

images. The trained model achieved accuracy of 90%

to detect and recognize 7 different types of dental

diseases. The researchers in (Al Kheraif et al.,2019)

collected 800 of X-ray images and then used adaptive

histogram equalization which helped to divide the

images into back-ground bones and foreground teeth.

After that, they used hybrid graph cut to perform the

segmentation process to separate the oral cavity and

the tissues. Finally, deep learning networks were

trained using the segmented images to perform the

predication with accuracy of 97%.

Orthopantomogram (OPD) images were collected in

(Laishram and Thongam, 2020), and then pre-

processing techniques were applied to prepare for the

training process of faster-RCNN to perform the

detection and the classification on the same time. The

trained convolutional neural network achieved 90%

in the detection process and 99% in the classification

process. Mobile app (OralCam) was proposed in

(Liang et al., 2020) to offer an end to end complete

system with self- examination of five different

diseases. 3,182 oral images were taken from 500

participants to train a conventional neural network

which was tested to give on average detection

sensitivity of 78.7%. 620 photographic images were

captured of extracted molars using smartphone in

(Duong et al., 2021). The collected images were

labelled manually into three classes by four dentists.

After that, a series of image pre-processing

techniques were applied to enhance the gathered

pictures and perform the segmentation process.

Finally, the classification process was implemented

using SVM classifier which was trained using colour

intensity features of the collected dataset. 640

photographic images of different patients’ oral

cavities were captured using a smartphone in (Ding et

al., 2021). Images enchantment and data

augmentation were applied on the collected dataset.

Data augmentation was used to increase the number

of images to 3,990 to prepare the collected data for

the training process. Then, transfer learning technique

was used by retraining YOLOv3 CNN model to

detect and recognize two types of caries.

The authors in (Zhu et al., 2022) presented a deep

learning network as U-shape architecture to perform

the segmentation process of 3127 panoramic

radiograph images. The pro-posed network was

called CariesNet to determine three different degrees

of caries based on panoramic X-ray images.

Additionally, they used full- scale axial attention

module to enhance the segmentation process and

improve the results. The proposed method achieved

93.61% of accuracy. The researchers in (Rashid et al.,

2022) proposed A hybrid system to localize regions

of caries by combining photographic and X-ray

images. They used the collected dataset to train mask

R-CNN deep learning model to perform the

segmentation process to detect regions of cavities and

oral diseases. The proposed system achieved about

92% of accuracy. 1902 photographic images were

taken using a smartphone of 695 participants in

(Thanh et al., 2022) to detect three different classes of

caries. Four different deep learning architecture were

re-trained to detect the oral lesions from the collected

images. The trained models were Faster R-CNNs,

YOLOv3, RetinaNet, and SSD. A retrospective study

was presented in (Keser et al., 2023) by collecting

photographic pictures of 65 healthy and 72 oral

lesions. Inception V3 deep learning network was

trained using the collected dataset to create a binary

classifier. The trained architecture achieved accuracy

of 100% for healthy and Oral lichen planus lesions

cases. The re-searchers in (Gomes et al., 2023)

collected 5069 images for six different types of oral

mucosal lesions. The images were labelled and

cropped manually by specialists. They trained four

different convolutional neural networks using 70% of

the collected dataset, the rest of the data was used to

test the trained models. ResNet-50, VGG16,

InceptionV3 and Xception were used as base

classifiers to perform the learning process of the

proposed models. A dataset of 470 Panoramic X-ray

images was labelled and segmented in (Haghanifar et

al., 2023). A genetic algorithm was proposed to

perform the segmentation process with image

processing operations to slice each tooth individually.

Finally, capsule network was trained using the

extracted features from different deep learning

networks to achieve accuracy of 86.05%.

Oral Diseases Recognition Based on Photographic Images

487

3 DATA MATERIAL AND

METHODOLOGY

The first step of our project is data collecting and

images gathering. The data collection process is

carried out at some local clinics with the help of

medical team of specialist dentists. We collect 1600

photographic images representing six common dental

diseases. The collected dataset is assembled from

both genders (male/female) with ages between 7 to 65

years. Since, the capturing process of the

photographic images is an easy process, simply using

a smartphone, we were able to collect images even

from children. On the other hand, capturing X-ray

images for such age is complicated task and prone to

errors and mistakes. The collected dataset was

obtained in an anonymous manner all recordings of

any private information regarding the patients' names,

ages, medical history, or even their status have been

omitted. Figure 1 shows some samples of the

collected images for the six cases of the oral diseases.

In the second stage of our project, we perform the

annotation process by applying image labelling to

separate our dataset into six categories of dental

diseases. This process requires some manual effort to

ensure accurate labelling. Unfortunately, the manual

labouring of labelling process cannot be carried out

by ordinary labellers. Hence, the manual annotation

of the collected images is required to be performed by

professionals with expertise in the field of dentistry.

So, three dentists examined the collected dataset and

categorized the images into six cases of oral diseases.

Then we divide our dataset into two segments with

the ratio of 80%:20%. The first part of our data is used

for the training process to learn predication models to

perform the detection and the recognition process.

The second part of our data is used to test and evaluate

the trained models and measure their performance.

Convolutional Neural Networks (CNNs) have shown

great success in dealing with image related learning

tasks. This is due to their natural compatibility with

the grid like structure of an image. Therefore, we

adopt CNNs to implement the detection and the

recognition operations. In our project, we propose a

deep learning network to detect the region of teeth to

localize the region of interest. Then, we use the

detection model to crop the images of our dataset to

train a CNN network to perform the classification of

oral diseases. We use data augmentation to create

more images and increase the dataset size. Data

augmentation can be considered as a regularization

technique by manipulating the original data to create

more copies and synthesize a different version of the

images through applying various types of

transformation such as rotation, translation, scaling,

and even light (brightness) changes. This technique is

used to improve the performance and reduce

overfitting by exposing the trained models to

augmented versions of the original dataset which

helps the models to generalize better and become

more robust. The collected images are categorized

into two kinds, as shown in figure 1. The first type

contains only the teeth (which is our region of interest

ROI). While the second type contains the teeth and

some other parts of the face such as cheeks, nose, lips,

and jaws. Therefore, we need a mechanism to

separate the two types of images in our dataset.

Manual splitting is not an option as it requires time

and effort. More importantly, manual split will

interfere with the automation process of the diagnosis

because we need to detect our region of interest and

then send the localized part for the classification

process.First, we determine whether the detection

process of ROI is required or not. After that, we

localize our ROI to send the cropped parts of the

images to training process. Finally, we use the

training data to train deep learning models and apply

the assessment and evaluation using the testing data

to measure the accuracy of the trained models. We

select 50 images for each case from our dataset,

where the first 25 images require detection to localize

the region of interest and the other 25 images do not

require any detection. The purpose of this collection

of images is to train a binary classifier to determine

whether the tested image needs detection to find the

region of interest or not, so we can use the important

parts of the images and remove any unnecessary

segments. To achieve the training process of this

binary classifier, we need an efficient and accurate

model. Therefore, we utilize the concept of the

transfer learning by selecting a pre-trained network

and modifying the chosen model to perform the

binary classification.

Hence, we use MobileNetV2 network as our base

model for the training process. Applying the transfer

learning approach leverages the prior knowledge of

MobileNetV2 network.

Firstly, we eliminate the last layer of the

MobileNetV2 network and flatten the resulted

features from the final layer. Subsequently, a fully

connected layer with 64 nodes is attached to the

model. Finally, we add the output layer with two

nodes of SoftMax layer to classify the images into

two classes (requires detection and does not require).

After that, we perform the training process using the

selected images.

ICAART 2024 - 16th International Conference on Agents and Artiﬁcial Intelligence

488

Figure 1: Photographic samples of our dataset.

During this stage, we freeze all the layers of

MobileNetV2 network (base layers) except for the

additional layers we added explicitly for binary

classification. This strategy allows us to optimize the

parameters for newly added layers while retaining the

original parameters of the MobileNetV2 network to

keep the previous knowledge of the model. In the

next phase, we need to train a fast and accurate

detector to determine the region of interest in the

contents of our images. For this particular task, we

use a light deep learning network to achieve the

training process of our detector model.

Adopting the transfer learning technique with

light pre-trained model can give us an efficient model

to implement the detection process. Hence, YOLO

V4-tiny model is used for this particular task. YOLO

V4-tiny is a single step object detector model which

means it can accomplish both the detection and

classification process in the same time in one step

instead of performing the two operations in separate

stages by applying initial detection in one step and

subsequent classification in the next step.

YOLO V4-tiny is a scaled down version of the

YOLO model has a smaller number of convolution

layers (smaller number of parameters) than the

ordinary YOLO. Therefore, adopting this model to

Case 1: Restorative

Dentistry

Case 2:

Pediatric

Case 3:

Orthodontic

Case 4:

Periodontics

Case 5:

Prosthodontics

Case 6: Oral medicine

Oral Diseases Recognition Based on Photographic Images

489

apply the training process reduces the cost of the

training time and the need of huge resources.

Additionally, by using transfer learning the pre-

trained YOLO V4-tiny returns the optimal values of

small number of parameters in the selected network.

Adopting this technique makes the learning process

possible despite the small size of the dataset.

Before the training process, we need to select

images from our data set to perform the detection

process to localize the region of interest inside the

images. Therefore, we selected 50 images, from each

case, which need detection. Then, we need to label

these images by providing a bounding box to

determine the coordinates of the region of interest for

each individual image. Since, we intend to train

YOLO V4-tiny model to perform the detection

process, the coordinates of bounding box for the

labelling operation should match the format of YOLO

network. Thus, we use Bbox-Label-Tool-Multi-Class

of the Darknet-library for the labelling process. It is

labelling tool that is completely compatible with the

YOLO format. This tool offers a programmable

configuration to initialize different modes of setting.

BBox-Label-Tool can create a simple GUI window to

input images and give the facility to label the region

of interest by applying a bounding box manually by a

user.

After the user localizes the region of interest,

BBOX tool will create the necessary files with the

required information for the training process. Figure

2 shows samples of the image labelling. Once the

labelling process of the selected images is done, we

can utilize these images to train our customized

detection model. As we mention before, we use for

this particular task YOLO V4-tiny network. We

retrain the pre-trained YOLO V4-tiny model with our

dataset to implement the learning process of the

detector. When the training process is completed, we

use the trained detector to localize the region of

interest to crop these regions. The cropped images are

used to prepare the training data to perform the

training of the disease classification. Figure 4 shows

the pipeline of the proposed method.

As shown in figure 3, after applying the detection

process, to determine the regions of interest, we

create database of the training images to learn

recognition models to classify 6 different types of

dental caries. Transfer learning approach is adopted

to achieve the training process of the classification

models. The analysis of the transfer learning achieves

the learning process by relying on the prior

knowledge from a pre-trained model. So, instead of

starting the training from scratch the learning process

starts with trained parameters of a base model which

has been trained using extensive amount of data.

Currently, numerous numbers of pre-trained models

are available to be utilized as base classifiers to

perform the training.

To implement the training of the disease’s

recognizer, we use VGG16 network as a foundational

classifier to exploit the prior knowledge of the

selected model. We modify the architecture of the

chosen network to achieve the training process with

our own images. We replace the last layer of VGG16

(which is responsible for classification of 1000

classes) with new classification layers for our 6

classes.

As shown in Figure 4, we include flatten layer to

make the size of vector features compatible with the

new attached layers. Then, we add a fully connected

layer of 512 nodes with a drop out layer of 0.5

dropping factor. The primary advantages of the drop

out layer is to decrease the effect of the overfitting

problem by skipping the update of the parameter’s

values during the training. Then, we add a second

fully connected layer of 256 nodes and followed by

another drop out layer. Finally, we wrap up the

designed network with a soft-max layer with 6 output

nodes to represent each individual disease.

After we complete the architecture of the

proposed network, we need to retrain the designed

model by applying fine-tune process to update the

weights and the parameters of our network to adjust

their values in a suitable configuration which allows

the trained model to capture and learn the most

relevant features from our dataset to achieve the

diseases classification task.

In order to improve the results, we expand our set

of experiments by testing different types of network

architectures as a base classifier. Hence, we use

additional pre-trained models to boost the accuracy of

our classifier. We exploit the prior knowledge

obtained from the following models: VGG16,

VGG19, GoogleNet, Xception, InceptionV3,

InceptionResNetV2, DenseNet201, MobileNetV2,

and NASNetLarge. Practically, we repeat the same

procedure of the fine-tune process to retrain these

models.

First, we remove the classification layer from the

selected network and we attach a fully connected

layers of 512 nodes with drop out layer, and followed

by another fully connected of 265 nodes with another

drop out layer. At the end of the network, we add a

soft max layer of 6 outputs nodes to classify 6

different diseases. All the results of these experiments

will be shown in the next section.

ICAART 2024 - 16th International Conference on Agents and Artiﬁcial Intelligence

490

Figure 2: Data annotation to train the detector model.

4 THE RESULTS

In the current section, we provide the finding results

of our experiments. We use the training data set to

learn the optimal features of our images to create

diseases classification model. Transfer learning

method is used to perform the recognition process.

First, we modify VGG16 network to implement the

re-training process to reshape the weights and the

parameters of the selected network to capture the

optimal representation of the learnt features of our

dataset. We utilize the testing segment of our dataset

to assess the quality of the trained models. Therefore,

we compute the accuracy, precision, and recall as

metric measurements to conduct comprehensive

Figure 3: Flowchart of the proposed algorithm.

Figure 4: The architecture of the designed CNN network.

Binary classifier ( to

deter mine whether

the i mage needs

detection or not )

YOLO Tiny(to

localize the region of

interest )

Crop the the region

of i nter est

Input imag

Pr epar e the dat a f or

tr aining

Train deep learning

models

Evaluate the tr ained

models

Soft max

layer (5 nodes)

Base Network(

GG 16)

Flatten

layer

Fully connected

layer (512

nodes)

Fully connected

layer (256 nodes)

Dropout

layer

Dropout

layer

Oral Diseases Recognition Based on Photographic Images

491

assessments, validations, and comparative analyses

of the trained classifiers. By adopting VGG16, we

obtain almost 92% of accuracy. Clearly, the prior

knowledge obtained by training a model using

extensive amount of dataset achieve reasonable

performance by applying a fine-tuning process to

learn the optimal features representation of

recognizing 6 dental caries in our dataset.

Table 1: Accuracy of deep learning networks.

Base network

model

Accuracy Precision Recall

VGG16

92.11% 92.42% 92.99%

VGG19

93.55%

93.27

93.13%

AlexNet

89.19% 88.99% 89.41%

Resnet50

92.74% 92.02% 92.83%

GoogleNet

91.25% 91.14% 90.79%

NasNet-Mobile

85.86% 86.32% 85.95%

DenseNet201

92.76% 92.58% 92.39%

MobileNetV2

87.82 % 88.01% 88.91%

InceptionResNet

90.68% 90.49% 90.18%

Xception

90.45% 90.61% 90.71%

InceptionV3

90.41% 90.97% 90.39%

It is important to highlight that all the experiment

and the obtained results are conducted by a personal

laptop type Lenovo where the processor is Core I7

with memory of 16 G RAM. The re-trained models

which are used in our experiments exhibit diversity in

their characteristics presenting variations in the

architecture design, connection mappings, layer

depth, and parameters quantities. Therefore, they

offer different performance and efficiency based on

their variations and properties. Basically, these

models may respond differently to the new given task

with unseen dataset to their previous knowledge.

Upon examining the results of Table 1, the

comparisons clearly demonstrate that VGG19

outshines as the most prominent base model which

offered an almost 93% of accuracy. The conducted

experiments show that the modified version of

VGG19 stands out by learning the best features of our

dataset to encapsulate the optimal representation of

different patterns for the oral diseases. Furthermore,

measured the confusion matrix of the modified

version of VGG19, these results are shown in figure

Figure 5: The confusion matrix of the re-trained model

VGG19.

5 CONCLUSIONS

In this paper, we propose a new dataset of

photographic images to train a predication model to

diagnose 6 different kinds of oral diseases. The

gathered images are annotated by expert dentists. The

collected images are used to train a binary recognizer

to determine whether detection is necessary inside the

dental images to find the region of interest (ROI).

After that, we deploy a modified version of YOLO

V4-tiny network to perform the detection process of

ROI. The detected parts of ROI within our data are

cropped to prepare our dataset for the classification

process. Finally, we adopt the transfer learning

strategy to train multiple pre-trained models to

implement the recognition process. The modification

of these models allows us to exploit their previous

knowledge and achieve 93% accuracy to classify six

different types of oral diseases.

REFERENCES

Araújo, A. L. D., da Silva, V. M., Kudo, M. S., de Souza,

E. S. C., Saldivia‐Siracusa, C., Giraldo‐Roldán, D., ...

& Moraes, M. C. (2023). Machine learning concepts

applied to oral pathology and oral medicine: A

convolutional neural networks' approach. Journal of

Oral Pathology & Medicine, 52(2), 109-118.

Amari, R., Noubigh, Z., Zrigui, S., Berchech, D., Nicolas,

H., & Zrigui, M. (2022, September). Deep

Convolutional Neural Network for Arabic Speech

Recognition. In International Conference on

Computational Collective Intelligence (pp. 120-134).

Cham: Springer International Publishing.

ICAART 2024 - 16th International Conference on Agents and Artiﬁcial Intelligence

492

Alhafidh, B. M. H., Daood, A. I., Alawad, M. M., & Allen,

W. (2018). FPGA hardware implementation of smart

home autonomous system based on deep learning. In

Internet of Things–ICIOT 2018: Third International

Conference, Held as Part of the Services Conference

Federation, SCF 2018, Seattle, WA, USA, June 25-30,

2018, Proceedings 3 (pp. 121-133). Springer

International Publishing.

Al Kheraif, A. A., Wahba, A. A., & Fouad, H. (2019).

Detection of dental diseases from radiographic 2d

dental image using hybrid graph-cut technique and

convolutional neural network. Measurement, 146, 333-

342.

Bellagha, M. L., & Zrigui, M. (2020, November). Speaker

naming in tv programs based on speaker role

recognition. In 2020 IEEE/ACS 17th International

Conference on Computer Systems and Applications

(AICCSA) (pp. 1-8). IEEE.

DAOOD, A., AL-SAEGH, A. L. I., & MAHMOOD, A. F.

(2023). HANDWRITING DETECTION AND

RECOGNITION OF ARABIC NUMBERS AND

CHARACTERS USING DEEP LEARNING

METHODS. Journal of Engineering Science and

Technology, 18(3), 1581-1598.

Ding, B., Zhang, Z., Liang, Y., Wang, W., Hao, S., Meng,

Z., ... & Lv, Y. (2021). Detection of dental caries in oral

photographs taken by mobile phones based on the

YOLOv3 algorithm. Annals of Translational Medicine,

9(21).

Farhani, N., Terbeh, N., & Zrigui, M. (2019). Object

recognition approach based on generalized hough

transform and color distribution serving in generating

arabic sentences. International Journal of Computer

and Information Engineering, 13(6), 339-344.

Duong, D. L., Kabir, M. H., & Kuo, R. F. (2021).

Automated caries detection with smartphone color

photography using machine learning. Health

Informatics Journal, 27(2), 14604582211007530.

Gomes, R. F. T., Schmith, J., Figueiredo, R. M. D., Freitas,

S. A., Machado, G. N., Romanini, J., & Carrard, V. C.

(2023). Use of artificial intelligence in the classification

of elementary oral lesions from clinical images.

International Journal of Environmental Research and

Public Health, 20(5), 3894.

Haghanifar, A., Majdabadi, M. M., Haghanifar, S., Choi,

Y., & Ko, S. B. (2023). PaXNet: Tooth segmentation

and dental caries detection in panoramic X-ray using

ensemble transfer learning and capsule classifier.

Multimedia Tools and Applications, 1-21.

Karaoglu, A., Ozcan, C., Pekince, A., & Yasa, Y. (2023).

Numbering teeth in panoramic images: A novel method

based on deep learning and heuristic algorithm.

Engineering Science and Technology, an International

Journal, 37, 101316.

Keser, G., Bayrakdar, İ. Ş., Pekiner, F. N., Çelik, Ö., &

Orhan, K. (2023). A deep learning algorithm for

classification of oral lichen planus lesions from

photographic images: A retrospective study. Journal of

Stomatology, Oral and Maxillofacial Surgery, 124(1),

101264.

Kim, H. N., Kim, K., & Lee, Y. (2023). Intra-oral

photograph analysis for gingivitis screening in

orthodontic patients. International Journal of

Environmental Research and Public Health, 20(4),

3705.

Laishram, A., & Thongam, K. (2020, February). Detection

and classification of dental pathologies using faster-

RCNN in orthopantomogram radiography image.

2020 7th international conference on signal processing

and integrated networks (SPIN) (pp. 423-428). IEEE.

Liang, Y., Fan, H. W., Fang, Z., Miao, L., Li, W., Zhang,

X., ... & Chen, X. A. (2020, April). OralCam: enabling

self-examination and awareness of oral health using a

smartphone camera. In Proceedings of the 2020 CHI

conference on human factors in computing systems (pp.

1-13).

Liu, L., Xu, J., Huan, Y., Zou, Z., Yeh, S. C., & Zheng, L.

R. (2019). A smart dental health-IoT platform based on

intelligent hardware, deep learning, and mobile

terminal. IEEE journal of biomedical and health

informatics, 24(3), 898-906.

Mansouri, S., Charhad, M., & Zrigui, M. (2017). Arabic

Text Detection in News Video Based on Line Segment

Detector. Res. Comput. Sci., 132, 97-106.

Mahmoud, A., & Zrigui, M. (2019). Deep neural network

models for paraphrased text classification in the Arabic

language. In Natural Language Processing and

Information Systems: 24th International Conference on

Applications of Natural Language to Information

Systems, NLDB 2019, Salford, UK, June 26–28, 2019,

Proceedings 24 (pp. 3-16). Springer International

Publishing.

Merhbene, L., Zouaghi, A., & Zrigui, M. (2010, June).

Ambiguous Arabic words disambiguation. In 2010 11th

ACIS International Conference on Software

Engineering, Artificial Intelligence, Networking and

Parallel/Distributed Computing (pp. 157-164). IEEE.

Rashid, U., Javid, A., Khan, A. R., Liu, L., Ahmed, A.,

Khalid, O., ... & Nawaz, R. (2022). A hybrid mask

RCNN-based tool to localize dental cavities from real-

time mixed photographic images. PeerJ Computer

Science, 8, e888.

Slimi, A., Hamroun, M., Zrigui, M., & Nicolas, H. (2020,

November). Emotion recognition from speech using

spectrograms and shallow neural networks. In

Proceedings of the 18th International Conference on

Advances in Mobile Computing & Multimedia (pp. 35-

39).

Thanh, M. T. G., Van Toan, N., Ngoc, V. T. N., Tra, N. T.,

Giap, C. N., & Nguyen, D. M. (2022). Deep learning

application in dental caries detection using intraoral

photos taken by smartphones. Applied Sciences, 12(11),

5504.

Zhu, H., Cao, Z., Lian, L., Ye, G., Gao, H., & Wu, J. (2022).

CariesNet: a deep learning approach for segmentation

of multi-stage caries lesion from oral panoramic X-ray

image. Neural Computing and Applications, 1-9.

Oral Diseases Recognition Based on Photographic Images

493