Article Type : Review Article
Authors : Addi RA, Benksim A and Cherkaoui M
Keywords : Artificial Intelligence (AI); Machine learning (ML); Deep learning (DP); Dentistry; Diagnostic
Although dated back to 1950, artificial
Intelligence (AI) has not become a practical tool until two decades ago. In
fact, AI is the capacity of machines to do tasks that normally require human
intelligence. AI applications have been started to provide convenience to
people’s lives due to the rapid development of big data computational power, as
well as AI algorithm. Furthermore, AI has been used in every dental
specialties. Most of the applications of AI in dentistry are in diagnosis based
on X-ray or visual images, whereas other functions are not as operative as
image-based functions mainly due to data availability issues, data uniformity
and computing power for processing 3D data. AI machine learning (ML) patterns
assimilate from human expertise whereas Evidence-based dentistry (EBD) is the
high standard for the decision-making of dentists. Thus, ML can be used as a
new precious implement to aid dental executives in manifold phases of work. It
is a necessity that institutions integrate AI into their theoretical and
practical training programs without forgetting the continuous training of
former dentists.
Artificial Intelligence (AI) is developing fast in all
sectors. It may assimilate human expertise and do tasks that required human
intelligence. It can be defined by the theory and development of computer
systems capable of executing tasks that need human understanding, such as
seeing perception, talk identification, resolution, and translation [1]. Also,
it a machine's ability to express its own intelligence by solving problems
based on data. Machine learning (ML) uses algorithms to anticipate outcomes
from a set of data. The aim is to facilitate it for machines without human
contribution to study from data and fix problems (Figure 1) [2]. Artificial
intelligence has been employed in every domain such as industry, medicine, dentistry,
research, portable display, hospital monitoring, automatic and non-Human
assistants. AI
may be often used as a practical implement helping dentists to minimize their
work time. In addition of diagnosing utilizing data feed directly, AI is able
to acquire a knowledge from several information origins to make a diagnostic
further on human capabilities.
AI may be sorted as weak AI and strong AI. Weak AI, utilizes an application skilled to fix unique or precise functions. Now, the most utilized AI is weak AI. For example, of AI in strengthening studying we can cite AlphaGo, and talk operating we have Google translation, and Amazon chat robot [3]. Strong AI calls attention to the competence and cleverness of AI equalling that of humans. It possesses its proper understanding and behaviour whose supplessness is comparable to humans [4]. Therefore, till then no strong AI applications are available. In Addition, ML is classified as supervised, semi-supervised and unsupervised learning. Supervised learning utilizes labelled inputs for learning to supervise the algorithm. The algorithm studies from the labelled input, releases and recognizes the shared characteristics of the labelled input to take auguries about unlabelled input [5]. At variance, unsupervised learning, performs automatically to discover the different characteristics of unlabelled data [6]. Semi-supervised learning reposes in mid of supervised and no-supervised learning, which employs a little size of labelled input jointly with a big size of unlabelled data during training [7]. Lately, a novel process named weakly supervised learning has been progressively common in the AI domain to reduce labelling expenses. Especially, the item division function solely utilizes picture-level marks as an alternative of item limit or position details for studying [8].
Figure 1: Key elements of artificial intelligence systems. (2).
Figure 2: Schematic diagram of deep learning (9).
Deep learning (DP) is now a significant
experimentation zone and constitutes a part of ML. It may use the two
supervised and unsupervised learning. DP represents an artificial “neural
network” composed of a base of three nodal layers—input, manifold “hidden”, and
output layers. Every layer is made of several interconnected nodes (synthetic
neurons) while every node is characterized by weight and biased threshold from
m crucial factors, provided by its proper linear regression model. The weight
is assigned when there is an input of the node. Similar to a decision tree
model, the neural network as a feedforward network is defined from the process
of passing data from one layer to the next (Figure 2) [9]. A deep neural
connection may bring characteristics from the input, without human
intervention. Neural networks (NN) are the mainstays of deep learning
algorithms. In fact, there are different variants of NN, the most important
sorts of neural networks are artificial neural networks (ANN), convolutional
neural networks (CNN) and generative adversarial networks (GAN).
ANN is composed of a set of neurons and layers, a
group of three layers corresponds to an elemental pattern for deep learning.
Only forward direction is allowed to inputs. Input neurons bring out
characteristics of input data from the input layer and dispatch data to hidden
layers, and the data traverse all the hidden layers consecutively. At the end,
the output layers expose and summarize the results. From previous layers, all
the hidden layers in ANN may weigh the data and perform adjustments to send
data to the next layer. Each hidden layer acts as an input and output layer,
allowing the ANN to understand more complex features [10].
CNN is a sort of deep learning pattern mostly utilized
for picture identification and production. The presence in CNN of convolution
layers, the pooling layer and the fully connected layer in the hidden layers is
the principal difference between CNN and ANN. Utilizing convolution kernels,
characteristic maps of input data were produced by convolution layers. The
input picture is bended by the kernels. Because of the weight sharing
convolution, the intricacy of pictures is decreased. The pooling layer is mostly
continued by every set of convolution layers, which decreases the size of
characteristic maps for more characteristic taking out. The fully connected
layer is utilized succeeding the convolution layer and pooling layer. The fully
connected layer links to every activated neurons in the previous layer and
converts the 2D characteristic maps into 1D. 1D characteristic maps are then
coupled with nodes of groups for categorization [11,12]. Finally, image
recognition is showing greater leverage and preciseness in CNN compared to ANN
due to the use of the functional hidden layers.
GAN is a sort of deep learning algorithm conceived by
Good fellow from the input data, this unsupervised learning method
automatically discover and generates new data with alike characteristics or
models in comparison with the input data [13]. 2 neural networks: a generator
and a discriminator. The principal aim for the generator is to produce input
which doesn’t allow to the discriminator to identify if the input is produced
by the generator or from the initial input data. The essential goal for the
discriminator is to differentiate between the output produced by the generator
and the initial input data as much as possible. The two GAN networks ameliorate
themselves and supplement each other. Furthermore, GAN has been spread fast
after its creation. They are mostly used picture-to-picture movement and
creating credible pictures of items, environments, and individuals [14,15]. A
new 3D-GAN structure was created founded on a conventional GAN connection [16].
It generates 3D items from a specified 3D spot by joining new discoveries in
GAN and dimensional convolutional networks. Different from a traditional GAN
network, it is competent to create items in 3D automatically or from 2D images.
It provides a larger spectrum of feasible utilizations in 3D input operating
juxtaposed with its 2D shape.
AI in operative
dentistry
Dentists identify dental decays by ocular and manual
investigation or by X-ray assessment but detecting early-stage lesions is
difficult when profound fissures, close interproximal joining. In fact, several
damages are detected uniquely in the late phases of dental decay, which conduct
to supplementary sophisticated treatment. Furthermore, most of diagnosis belong
to dentists’ experience despite of the wide use of dental radiography and
explorer in dental caries diagnosis. Each pixel has a degree of grey in
two-dimensional X-Ray which represents the object density. An AI algorithm may
assimilate the model and provide auguries to several dental lesions from this
concept. In fact, several studies performed a CNN algorithm dental caries
detection on periapical x-rays and intraoral images [17,18]. Others found that
AI in proximal caries detection was further productive and cheaper than
dentists [19]. Actually, AI showed encouraging results in precocious detection
of dental lesions, which accuracy was better than dentists or at less the same
(Figure 3,4).
AI in periodontics
Periodontitis is one of the most prevalent troubles.
It is a charge for billions of people and, if not well fixed, may conduct to
tooth mobility or loss [20]. It is well known that Prompt discovery and care
are required to avert acute periodontitis. In clinical practice, periodontal
illness determination is based on assessing pocket probing profundity and
gingival regress. Researchers used AI in diagnostic and periodontal disease
classification [21,22].
Others researchers utilized CNN in the discovery of
periodontal bone damage on panoramic radiographs [23]. In addition, studies
started that periodontal status may be inspected by a CNN algorithm utilizing
organizational health-related input [23].
AI in orthodontics
Orthodontic treatment organization is generally found on the experience and priority of the orthodontists. In fact, orthodontists spend a great effort to identify malocclusion, due to the multitude of changeable that must be examined in the cephalometric investigation, which makes difficult to establish the treatment program and anticipate the result [24].
Figure 3: CNN model to forecast the patient's dental condition from a panoramic radiograph (2).
Figure 4: Applications of AI in different subfields of dentistry (2).
Moreover, treatment planning and prediction of
treatment results, such as simulating the changes in the appearance of pre- and
post-treatment facial photographs are the most applications of AI in
orthodontics. Actually, thanks of AI, the orthodontic treatment outcome, the
skeletal class, and the anatomic landmarks in lateral x-rays may be examined
[25]. A
study performed an algorithm to diagnose if there is a requirement for
treatment by orthodontics on the base of orthodontics-related data [26].
On other study, an ANN model was proposed
to estimate if there is need of extractions based on lateral cephalometric
radiographs [27,28]. Also, several studies have demonstrated how AI may
automatically locate cephalometric landmarks with high accuracy as well as the
need of orthognathic surgery [29-32].
AI in oral and
maxillofacial pathology
Oral and Maxillofacial Pathology (OMFP) is a specialty
that examines pathological status and diagnoses sickness s in the buccal and
maxillofacial area. The most serious kind of OMFP is buccal cancer. World
Health Organization (WHO) reports over 657,000 patients with buccal cancer
which cause more than 330,000 deaths per year [33]. By utilizing x-rays,
pictures from microscope and ultrasonography AI may be utilized for tumour and
cancer identification by CNN algorithms [34,35]. AI is used to handle cleft lip
and palate in risk augury [36]. Further, with intrabuccal visual pictures and
using a CNN model, it was possible to spot buccal latent malignant troubles and
oral squamous cell carcinoma (OSCC). Also, optical Coherence Tomography (OCT)
has been utilized in the recognition of benign and malignant lesions in the
buccal mucosa in addition to intrabuccal visual pictures. In addition, a study has used ANN and Support
Vector Machine (SVM) patterns to identify neoplastic buccal lesions [37]. In
other study, researchers were able to mechanically identify oral squamous cell
carcinoma using a CNN algorithm from confocal laser endomicroscopy pictures
[34]. Finally, a study has used a CNN algorithm to recognize and determine
ameloblastoma and keratocystic odontogenic tumour (KCOT) [38].
AI in prosthodontics
AI is mostly used in prosthodontics to perform the
restoration design. CAD/CAM has digitalized the design work in profit-oriented
yields, like CEREC, 3Shape, etc. Some studies demonstrated novel methods
founded on 2D-GAN patterns creating a crown by studying shape technicians’
designs. Transformed from 3D mouth models, the forming input was 2D depth maps.
Other study utilized 3D data directly generating crown using a 3DDCGAN network
[39,40]. In addition, associating AI and CAD/CAM or 3D/4D printing could bring
a high effectiveness (88). Also, in debonding prediction and shade matching of
restorations, AI may be an unavoidable support [41, 42]. However, in removable
prosthodontics the design is more demanding as more elements and changeable
must be reviewed. Assisting the conception process of partial dentures is the
most used feature in recent ML algorithms [43,44].
AI in Endodontics
Using properties of periapical radiolucency, AI
algorithms may identify periapical disease [45]. Also, radiolucencies can be
recognized on periapical on panoramic radiographs with deep learning algorithm
model [46,47]. A study utilizing AI system identified 142 out of 153 periapical
lesions with a detection accuracy rate of 92.8%. In addition, utilizing
artificial neural connections the detection of cystic lesions has been done
[48]. Furthermore, a separation of granuloma from periapical cysts using CBCT
images was performed and three-dimensional teeth segmentation using the CNN
method was demonstrated [49,50]. AI can assimilate further on that human
competence. Also, the growth of computer tech is vital to promote the AI
development. Evidence-Based Dentistry (EBD) is “an approach to oral health care
that requires the judicious integration of systematic assessments of clinically
relevant scientific evidence, relating to the patient’s oral and medical
condition and history with the dentist’s clinical expertise and the patient’s
treatment needs and preferences”. ML models may be considered like another
helpful instrument for health professionals. Indeed, EBD and ML are matching to
better help dental professionals, in fact they may use it both to enhance their
benefits and place them to medical exercise.
A multiple of AI systems are being developed for
diverse dental disciplines and have produced encouraging results which predicts
a bright future for AI in dentistry. Thereafter, it is now a necessity that
institutions integrate AI into their theoretical and practical training
programs without forgetting the continuous training of former dentists.