Imaging Science in Dentistry 2023; 53: 193-8

https://doi.org/10.5624/isd.20230092

Efficiency and accuracy of artificial intelligence in the radiographic detection of periodontal

bone loss: A systematic review

Asmhan Tariq 1, Fatmah Bin Nakhi 1, Fatema Salah 1, Gabass Eltayeb 1, Ghada Jassem Abdulla 1,
Noor Najim 1, Salma Ahmed Khedr 1, Sara Elkerdasy 1, Natheer Al-Rawi 1, Sausan Alkawas 1,
Marwan Mohammed 1, Shishir Ram Shetty 1,*
1
Department of Oral and Craniofacial Health Sciences, College of Dental Medicine, University of Sharjah, Sharjah, United Arab Emirates

ABSTRACT

Purpose: Artificial intelligence (AI) is poised to play a major role in medical diagnostics. Periodontal disease is one
of the most common oral diseases. The early diagnosis of periodontal disease is essential for effective treatment and
a favorable prognosis. This study aimed to assess the effectiveness of AI in diagnosing periodontal bone loss through
radiographic analysis.
Materials and Methods: A literature search involving 5 databases (PubMed, ScienceDirect, Scopus, Health and
Medical Collection, Dentistry and Oral Sciences) was carried out. A specific combination of keywords was used to
obtain the articles. The PRISMA guidelines were used to filter eligible articles. The study design, sample size, type of
AI software, and the results of each eligible study were analyzed. The CASP diagnostic study checklist was used to
evaluate the evidence strength score.
Results: Seven articles were eligible for review according to the PRISMA guidelines. Out of the 7 eligible studies, 4
had strong CASP evidence strength scores (7-8/9). The remaining studies had intermediate CASP evidence strength
scores (3.5-6.5/9). The highest area under the curve among the reported studies was 94%, the highest F1 score was
91%, and the highest specificity and sensitivity were 98.1% and 94%, respectively.
Conclusion: AI-based detection of periodontal bone loss using radiographs is an efficient method. However, more
clinical studies need to be conducted before this method is introduced into routine dental practice. (Imaging Sci Dent
2023; 53: 193-8)

KEY WORDS: Artificial Intelligence; Radiography, Dental; Periodontitis; Alveolar Bone Loss; Diagnosis

addition to clinical examinations, which include periodontal

Introduction pocket probing, bleeding on probing, and determining the
Periodontal disease is characterized by destruction of the clinical attachment level, radiological examinations serve
periodontal attachment apparatus, loss of the alveolar bone, as a supplementary method for diagnosing periodontal dis-
and, in untreated cases, tooth loss.1 This oral disease is quite ease.4
prevalent, affecting between 20-50% of the population. It With ever-changing technology, increasingly sophistica-
is associated with various factors, including smoking, poor ted methods are being developed to enhance and refine med-
oral hygiene, and certain systemic diseases.2,3 An early diag- ical diagnostics. One such advancement is artificial intelli-
nosis of periodontal disease is crucial for devising an appro- gence (AI), a technology that emulates human intelligence.5
priate treatment plan and ensuring a positive prognosis. In One form of AI is the convolutional neural network (CNNs).
CNNs are a category of artificial neural networks that anal-
Received April 26, 2023; Revised June 18, 2023; Accepted June 23, 2023 yze visual imagery, making them particularly useful in the
Published online August 2, 2023
*Correspondence to : Dr. Shishir Ram Shetty radiographic assessment of bone loss in periodontal dis-
Department of Oral and Craniofacial Health Sciences, College of Dental Medicine, ease.6 This technology employs deep learning and is not lim-
University of Sharjah, 27272 Sharjah, United Arab Emirates
Tel) 971-6-505-7333, E-mail) [email protected] ited to executing pre-programmed instructions. Remarkably,

Copyright ⓒ 2023 by Korean Academy of Oral and Maxillofacial Radiology

This is an Open Access article distributed under the terms of the Creative Commons Attribution Non-Commercial License (http://creativecommons.org/licenses/by-nc/3.0)
which permits unrestricted non-commercial use, distribution, and reproduction in any medium, provided the original work is properly cited.
Imaging Science in Dentistry·pISSN 2233-7822 eISSN 2233-7830

- 193 -
Efficiency and accuracy of artificial intelligence in the radiographic detection of periodontal bone loss: A systematic review

CNNs can also “gain experience” or learn from practice, for Systematic Reviews and Meta-Analyses) guidelines
thereby assisting healthcare providers in achieving accurate (Fig. 1).8
diagnoses and effective future treatment outcomes.7 The screening procedure was independently conducted
To the best of our knowledge, very few systematic rev- by 10 reviewers, divided into pairs for each database. If a
iews have analyzed the role of AI in the detection of peri- disagreement arose, a third reviewer, who was a supervising
odontal bone loss. The purpose of this study was to system- faculty member, was brought in to resolve the issue.
atically review the published articles regarding the efficiency The keyword combinations specified earlier were utilized
of AI in detecting periodontal bone loss. to retrieve 2,290 articles from various databases: 1,125 from
the Health and Medical Collection (ProQuest), 358 from
ScienceDirect, 311 from Dentistry and Oral Science, 424
Materials and Methods from PubMed, and 72 from Scopus. These five groups then
Research question reviewed the 2,290 articles to eliminate any duplicates, re-
The application of AI in radiographic detection of peri- sulting in a total of 375 articles. After further screening for
odontal diseases will depend on the evidence reported in full-text availability, 37 articles remained. Twenty of the 37
clinical studies. This systematic review was centered around articles were removed because they were review articles,
the following research question: How efficient and accurate communications, commentaries, conference papers, and
are the current clinical applications of artificial intelligence opinion papers. This left 17 articles qualifying for the eligi-
in the radiographic diagnosis of periodontal diseases? bility stage, 10 of which were non-radiographic studies per-
taining to AI and periodontitis. Finally, 7 full-text articles
Search strategy and protocol were eligible for the qualitative analysis.
A comprehensive literature search was conducted across The raw data for this systematic review, recorded in an
several databases, including PubMed, ScienceDirect, Sco- Excel spreadsheet, are available at Figshare.com (doi:
pus, and Dentistry and Oral Sciences (EBSCO). This was 10.6084/m9.figshare.19586008).
supplemented with a manual search. The search process The studies were assessed using the Critical Appraisal
spanned from November 2, 2021, to February 16, 2022. All Skills Programme (CASP) diagnostic study checklist.9
the articles identified were published within the period of
2018 to 2022.
For the search strategy, a set of keywords was used in
Results
combinations in the databases to generate various results. Evidence table
The keyword combinations, structured according to the An evidence table was made to summarize the data of the
Boolean method, were “artificial intelligence (OR deep articles (Table 1). The table contains specific parameters
learning OR convolutional neural network)” AND “peri- such as author, year, country, sample type and size, the AI
odontal disease (OR bone loss OR periodontitis)” AND software used, main study findings, and statistical findings
“Radiography (OR panoramic radiograph OR periapical (area under the curve [AUC], sensitivity, specificity).10-16
radiography).” Other keywords, such as “detect” and “diag-
nosis,” were used as well. Study quality assessment
The studies were assessed using the CASP diagnostic
Eligibility criteria study checklist, which includes 12 questions.9 Nine of the
This study included full-text articles in English that de- 12 questions were assigned 1 point each (3 questions in the
scribed research utilizing AI to detect periodontal bone loss checklist required descriptive answers and were therefore
in dental radiographs, specifically panoramic and periapical not assigned a score). These 9 questions were used for the
radiographs. Review articles, communications, commentar- quality assessment of the 7 eligible articles. It was found
ies, case reports, and opinion papers were excluded. Non- that out of the 7 eligible studies, 4 had strong evidence
radiographic studies involving the use of AI to detect peri- (scores ranging between 7 and 8). The remaining studies
odontitis were also excluded. had intermediate-strength evidence (scores ranging between
3.5 and 6.5) (Table 2).
Data extraction
The detailed data extraction procedure was conducted in Study characteristics and outcomes
accordance with the PRISMA (Preferred Reporting Items All of the studies were equivalence trials, with sample siz-

- 194 -
 Asmhan Tariq et al

Fig. 1. PRISMA flowchart of the systematic review.

es ranging from 30 to 2276 patients. Three of the 7 studies the efficiency of the learning machines. This parameter
were conducted in South Korea.10,11,15 German, Turkish, Thai, represents a model’s ability to distinguish changes or ab-
and Swiss researchers conducted 1 study each.12-14,16 The normalities in images. The highest AUC was 94%.12 Lower
dates of the published studies ranged from 2018 to 2022. AUC scores were reported by Lee et al. and Kim et al.10,15
One study was published in 2018,15 2 in 2019,10,12 3 in Specificity was analyzed in 5 of the 7 studies. Zadrożny
2020,11,13,14 and 1 in 2022.16 All 7 studies evaluated the et al.16 reported the highest specificity (98.1%), followed by
use of AI models (specifically, the CNN subtype of deep Kim et al.10 (92%). In contrast, the lowest specificity score
learning) to detect periodontal bone loss.10-16 Five articles (63%) was reported by Krois et al.12 Two studies reported
demonstrated accuracy above 80%, while two had accuracy specificity ranging from 71% to 88%.13,14
results below 80%. Sensitivity metrics were reported by 6 of the 9 studies.
Three studies used the F1 score as the metric to measure Kurt Bayrakdar et al.13 (94%) and Krois et al.12 (92%) re-
accuracy. The F1 score combines the precision and recall ported the highest sensitivity.
score of an AI model.10,12,14 The study by Kurt Bayrakdar et
al.14 showed the highest F1 score (91%), while Kim et al.10
and Krois et al.12 found F1 scores between 75% and 78%
Discussion
in their studies. The application of deep learning-based technology for
The area under the curve (AUC) was also used to assess identifying periodontal bone loss in panoramic and periapi-

- 195 -
Efficiency and accuracy of artificial intelligence in the radiographic detection of periodontal bone loss: A systematic review

Table 1. Evidence table for the eligible studies

Author, Evidence
Sample type
year, AI software used Main study findings Statistical findings strength
and size
country score

Kim et al., 12,179 DeNTNet using DeNTNet was more effective F1 score: DeNTNet: 0.75 5.5/9
201910 panoramic deep learning in detecting PBL Clinicians: 0.69 Intermediate
South Korea radiographs (CNN) on panoramic radiographs AUROC: 0.95 evidence
than dental clinicians. Sensitivity: 0.87
Specificity:0.96

Chang et al., 518 Hybrid of The hybrid model showed AUC/Sensitivity/ 5.5/9
202011 panoramic deep learning for high accuracy and excellent specificity: not mentioned. Intermediate
South Korea radiographs detection and reliability in detecting The mean absolute difference evidence
CAD for radiographic bone loss. between AI and the radiologist
classification. diagnosis was 0.25

Krois et, al., 85 CNN The accuracy and reliability Mean accuracy: 7/9
201912 panoramic of the CNN were similar to Dentists: 0.06 Strong
Germany radiographs those of dentists, CNN: 0.067 evidence
but the diagnosis was faster Sensitivity:0.02
with CNN. Specificity: 0.14

Kurt Bayrakdar 2,276 CNN The utilization of AI for Sensitivity 0.9429 7/9
et al., 202013 panoramic interpreting dental radiographs Specificity 0.8857 Strong
Turkey radiographs facilitated the prompt diagnosis evidence
and treatment planning of
periodontal diseases.

Thanathornwong 100 Regional CNNs R-CNN tended to reduce Sensitivity: 0.84 8/9
and Suebnukarn, panoramic the effort and time required Specificity:0.88 Strong
202014 radiographs for the automated screening of evidence
Thailand periodontal diseases.

Lee et al., 1,740 Deep CNN Deep CNN proved to be Sensitivity specificity: 7.5/9
201815 periapical very effective and efficient not mentioned. Strong
South Korea radiographs in detecting periodontal bone CNN: AUC (95%) evidence
loss compared to experienced
periodontists.

Zadrożny et al., 30 Diagnocat The AI system can be very Specificity: >0.9 6.5/9
202216 panoramic software useful in the initial screening Sensitivity: >0.9 Intermediate
Switzerland. radiographs of panoramic radiographs. for assessment of missing teeth evidence
and prosthetic restorations,
and >0.8 for PBL.

PBL: periodontal bone loss, AI: artificial intelligence, CNN: convolutional neural network, R-CNN: region-based convolution neural network, AUROC: area under the
receiver operating characteristic curve, CAD: computer-aided design, AUC: area under the curve

cal radiographs can greatly assist dental clinicians in making sification in medical images. Current studies are exploring
accurate diagnoses. This technology significantly cuts down the use of CNN models for early detection of diseases such
on the time required for radiographic examinations, thereby as cancer, neurological and eye disorders, and dental con-
allowing more time for clinical examinations and treatment ditions.23,24 The majority of the articles included in this sys-
planning.17-20 tematic review utilized region-based CNNs (R-CNNs) for
Deep learning, a more complex subdomain of AI, utilizes the detection of periodontal bone loss.10-16
mathematical algorithms to mimic the way the human brain The process of using a CNN to detect radiographic changes
perceives, processes, and learns new information.21 CNNs, follows a specific sequence. In the initial stage the radio-
which constitute a specific subtype of deep learning, are used graphs are classified and annotated. They are then passed
to process and interpret radiographic and photographic through the layers of the CNN, which is trained by various
images with high accuracy.22 It is anticipated that CNN chain functions. Once the training is completed, validation
models will be extensively applied in healthcare settings, and testing are carried out, typically using new datasets.22-25
particularly for automatic feature learning and disease clas- Kurt Bayrakdar et al.13 used a CNN with 22 deep layers

- 196 -
 Asmhan Tariq et al

Table 2. The CASP diagnostic checklist scores for each article

Authors and year Title Evidence scores

Kim et al., 201910 DeNTNet: Deep Neural Transfer Network for the detection of 5.5/9
periodontal bone loss using panoramic dental radiographs (Diagnostic) Intermediate evidence
Chang et al., 202011 Deep Learning Hybrid Method to Automatically Diagnose Periodontal 5.5/9
Bone Loss and Stage Periodontitis Intermediate evidence
Krois et al., 201912 Deep Learning for the Radiographic Detection of Periodontal Bone Loss 7/9
Strong evidence
Kurt Bayrakdar et al., 202013 Success of artificial intelligence system in determining alveolar bone loss 7/9
from dental panoramic radiography images Strong evidence
Thanathornwong and Automatic detection of periodontal compromised teeth in digital panoramic 8/9
Suebnukarn, 202014 radiographs using faster regional convolutional neural networks. Strong evidence

Lee et al., 201815 Diagnosis and prediction of periodontally compromised teeth using 7.5/9
a deep learning-based convolutional neural network algorithm Strong evidence
Zadrożny et al., 202216 Artificial Intelligence Application in Assessment of Panoramic Radiographs 6.5/9
Intermediate evidence

to analyze panoramic radiographs of periodontally healthy ods using radiographs may not provide sufficient precision.
and diseased (with bone destruction) patients. In contrast, Dental specialists typically utilize a combination of methods,
Kim et al.10 used CNN with transfer networks (DeNTNet) including clinical findings and periodontal chartings, in con-
in their study. Thanathornwong and Suebnukarn14 used an junction with radiographs.
R-CNN, which primarily focuses on object detection at a
speed surpassing that of conventional CNNs. Krois et al.12 Conflicts of Interest: None.
used a deep learning CNN machine, supplemented with a
transfer network, to detect periodontal bone loss. Two stud-
References
ies (Chang et al.11 and Krois et al.12) reported that CNNs
could be used for overall assessments of the bone level, 1. H ighfield J. Diagnosis and classification of periodontal disease.
Aust Dent J 2009; 54 Suppl 1: S11-26.
cementoenamel junction, level, and tooth long-axis.
2. Bourgeois D, Inquimbert C, Ottolenghi L, Carrouel F. Periodon-
In 6 out of the 7 studies, panoramic radiographs were used tal pathogens as risk factors of cardiovascular diseases, diabe-
for the training, validation, and testing datasets.10-16 How- tes, rheumatoid arthritis, cancer, and chronic obstructive pulmo-
ever, Kim et al.15 used cropped panoramic images in their nary disease - is there cause for consideration? Microorganisms
study.10 Only 1 study used periapical radiographs. While 2019; 7: 424.
3. Nazir M, Al-Ansari A, Al-Khalifa K, Alhareky M, Gaffar B,
panoramic radiographs are commonly used for detecting
Almas K. Global prevalence of periodontal disease and lack of
alveolar bone loss, they may not offer the same level of im- its surveillance. ScientificWorldJournal 2020; 2020: 2146160.
age detail and resolution as periapical radiographs.26 4. Ramenzoni L, Lehner M, Kaufmann M, Wiedemeier D, Attin
In conclusion, the application of artificial intelligence in T, Schmidlin PR. Oral diagnostic methods for the detection of
the investigation of periodontal disease can be beneficial and periodontal disease. Diagnostics (Basel) 2021; 11: 571.
efficient in minimizing patient chairside time, though it is not 5. Al-Rawi N, Sultan A, Rajai B, Shuaeeb H, Alnajjar M, Alketbi
M, et al. The effectiveness of artificial intelligence in detection
without challenges. These challenges encompass ethical con-
of oral cancer. Int Dent J 2022; 72: 436-47.
cerns related to patient trust and reliance, as well as technical 6. Abiodun OI, Jantan A, Omolara AE, Dada KV, Mohamed NA,
issues such as the need for software updates, potential tech- Arshad H. State-of-the-art in artificial neural network applica-
nical difficulties, and the costs and setup of the machinery. tions: a survey. Heliyon 2018; 4: e00938.
The use of CNN-based deep learning models in clinical 7. Shetty S, Reddy S, Shetty R, Halkai, R, Shetty S, Halkai, K.
Machine learning in maxillofacial radiology: a review. J Datta
trials has demonstrated effectiveness in identifying peri-
Meghe Inst Med Sci Univ 2021; 16: 794-6.
odontally compromised teeth. This approach also reduces 8. Page MJ, McKenzie JE, Bossuyt PM, Boutron I, Hoffmann TC,
patient chairside time by providing a quicker, yet accurate, Mulrow CD, et al. The PRISMA 2020 statement: an updated
diagnosis. However, relying solely on deep learning meth- guideline for reporting systematic reviews. BMJ 2021; 372:

- 197 -
Efficiency and accuracy of artificial intelligence in the radiographic detection of periodontal bone loss: A systematic review

n71. periodontal disease: a panoramic radiograph study. Diagnostics

9. casp-uk.net [Internet]. Oxford: Critical appraisal skills pro- (Basel) 2023; 13: 1800.
gramme; 2023 [cited 2023 March 4]. Available from https:// 18. Ryu J, Lee DM, Jung YH, Kwon O, Park S, Hwang J, et al. Auto-
www.casp-uk.net. mated detection of periodontal bone loss using deep learning
10. Kim J, Lee HS, Song IS, Jung KH. DeNTNet: deep neural and panoramic radiographs: a convolutional neural network
transfer network for the detection of periodontal bone loss using approach. Appl Sci (Basel) 2023; 13: 5261.
panoramic dental radiographs. Sci Rep 2019; 9: 17615. 19. Chartrand G, Cheng PM, Vorontsov E, Drozdzal M, Turcotte
11. C
 hang HJ, Lee SJ, Yong TH, Shin NY, Jang BG, Kim JE, et al. S, Pal CJ, et al. Deep learning: a primer for radiologists. Radio-
Deep learning hybrid method to automatically diagnose peri- graphics 2017; 37: 2113-31.
odontal bone loss and stage periodontitis. Sci Rep 2020; 10: 20. N guyen TT, Larrivée N, Lee A, Bilaniuk O, Durand R. Use of
7531. artificial intelligence in dentistry: current clinical trends and
12. Krois J, Ekert T, Meinhold L, Golla T, Kharbot B, Wittemeier A, research advances. J Can Dent Assoc 2021; 87: I7.
et al. Deep learning for the radiographic detection of periodon- 21. Bordoloi D, Singh V, Sanober S, Buhari SM, Ujjan JA, Boddu
tal bone loss. Sci Rep 2019; 9: 8495. R. Deep learning in healthcare system for quality of service. J
13. Kurt Bayrakdar S, Çeliko, Bayrakdar IS, Orhan K, Bilgir E, Healthc Eng 2022; 2022: 8169203.
Odabas A, et al. Success of artificial intelligence system in deter- 22. Hwang JJ, Jung YH, Cho BH, Heo MS. An overview of deep
mining alveolar bone loss from dental panoramic radiography learning in the field of dentistry. Imaging Sci Dent 2019; 49:
images. Cumhuriyet Dent J 2020; 23: 318-24. 1-7.
14. Thanathornwong B, Suebnukarn S. Automatic detection of peri- 23. Miotto R, Wang F, Wang S, Jiang X, Dudley JT. Deep learning
odontal compromised teeth in digital panoramic radiographs for healthcare: review, opportunities and challenges. Brief Bio-
using faster regional convolutional neural networks. Imaging inform 2018; 19: 1236-46.
Sci Dent 2020; 50: 169-74. 24. Corbella S, Srinivas S, Cabitza F. Applications of deep learning
15. Lee JH, Kim DH, Jeong SN, Choi SH. Diagnosis and prediction in dentistry. Oral Surg Oral Med Oral Pathol Oral Radiol 2021;
of periodontally compromised teeth using a deep learning-based 132: 225-38.
convolutional neural network algorithm. J Periodontal Implant 25. Manerikar CR, Mani A, Mani S, Sachdeva S, Sodhi JK. Convo-
Sci 2018; 48: 114-23. lutional neural network in periodontology - innovative technol-
16. Zadrożny Ł, Regulski P, Brus-Sawczuk K, Czajkowska M, Par- ogy or new era? - a review. Ann Rom Soc Cell Biol 2021; 25:
kanyi L, Ganz S, et al. Artificial intelligence application in assess- 7453-63.
ment of panoramic radiographs. Diagnostics (Basel) 2022; 12: 26. H ou GL, Hung CC, Yang YS, Shieh TY, Tsai CC. Radiographic
224. alveolar bone loss in untreated Taiwan Chinese subjects with
17. Uzun Saylan BC, Baydar O, Yeşilova E, Kurt Bayrakdar S, adult periodontitis measured by the digital scanning radiogra-
Bilgir E, Bayrakdar İŞ, et al. Assessing the effectiveness of phic image analysis method. Dentomaxillofac Radiol 2003; 32:
artificial intelligence models for detecting alveolar bone loss in 104-8.

- 198 -