Applications of artificial intelligence in the analysis of dental panoramic radiographs: an overview of systematic reviews

Abstract

Objectives:

This overview of systematic reviews aimed to establish the current state of knowledge on the suitability of artificial intelligence (AI) in dental panoramic radiograph analysis and illustrate its changes over time.

Methods:

Medical databases covered by the Association for Computing Machinery, Bielefeld Academic Search Engine, Google Scholar, and PubMed engines were searched. The risk of bias was assessed using ROBIS tool. Ultimately, 12 articles were qualified for the qualitative synthesis. The results were visualized with timelines, tables, and charts.

Results:

In the years 1988–2023, a significant development of information technologies for the analysis of DPRs was observed. The latest analyzed AI models achieve high accuracy in detecting caries (91.5%), osteoporosis (89.29%), maxillary sinusitis (87.5%), periodontal bone loss (93.09%), and teeth identification and numbering (93.67%). The detection of periapical lesions is also characterized by high sensitivity (99.95%) and specificity (92%). However, due to the small number of heterogeneous source studies synthesized in systematic reviews, the results of this overview should be interpreted with caution.

Conclusion:

Currently, AI applications can significantly support dentists in dental panoramic radiograph analysis. As systematic reviews on AI become outdated quickly, their regular updating is recommended. PROSPERO registration number: CRD42023416048.

Introduction

Background

Artificial intelligence (AI), introduced in 1956, has become a hot topic in contemporary radiology and an area of interest for many researchers. ¹ Its integration into healthcare has expanded significantly in the past decade. ² AI is defined as the ability of a machine to perform complex tasks imitating specific human activities, such as solving problems, and making plans or decisions. ³ The main technology used in dentistry is known as image processing. These algorithms achieve high-accuracy classification and segmentation in radiographs, including panoramic ones. Modern analytical algorithms, including those used in medical analysis, can self-learn based on successive portions of data provided, which is referred to as machine learning (ML). ⁴ It is a part of AI that provides machines to learn from experience without explicit programming for a single task. ⁵ Neural networks are the heart of deep learning (DL) algorithms, which allow computers to learn by observing patterns in the data (Figure 1). ⁶ A type of neural network used mainly for image recognition and processing is called a convolutional neural network (CNN). It is crucial to provide a large data set of images to begin the training process and obtain a high-performance deep learning model.

Figure 1.

The major aspects of artificial intelligence.

According to Thurzo et al from 2011, publications on the use of AI in dentistry mainly concerned radiology (26.36%), orthodontics (18.31%), general scope (17.10%), restorative dentistry (12.09%), surgery (11.87%), and education (5.63%) (Figure 2). ² The applications of automated diagnostic and prognostic systems precisely identifying pathologies, cysts, tumors, periapical lesions, fractures, and other diseases are particularly promising in dental radiology. Most ML algorithms were developed using 2D diagnostic images such as cephalometric, periapical, and panoramic radiographs.

Figure 2.

Main AI focus in dentistry from 2011 to 2021. AI, artificial intelligence.

Dental panoramic radiographs (DPRs) are widely accepted as a primary source of information about oral health as they visualize all teeth and surrounding structures in one image, allowing for a basic assessment of dentition, periodontal bone loss, and lesions within the jaw bones. ⁷ Non-dental random findings on DPRs include pathological conditions such as maxillary sinusitis, sialoliths, tonsilloliths, elongated stylohyoid processes, and calcifications in the carotid arteries. ⁸ Besides the screening purpose, properly analyzed DPRs help to make therapeutic decisions. Their comprehensive analysis is time-consuming but desirable due to the identification of possible comorbidities.

Rationale

AI is considered useful in the detection of dental caries, vertical root fractures, apical lesions, periodontal bone loss, tumors and maxillary sinusitis on panoramic radiography, as well as in determining the proximity of the inferior alveolar nerve to the roots of lower third molars. ^9,10 The number of systematic reviews of various possibilities of using AI in analyzing panoramic images encourages a comprehensive summary of the current state of knowledge.

Objectives

The purpose of this overview of reviews was to establish the current state of knowledge on the suitability of artificial intelligence in DPR analysis and illustrate its changes over time. This research question belongs to the second type of research question according to the guidelines of Pollock et al for overviews of reviews. ¹¹

Methods

Eligibility criteria

The overview of reviews followed Preferred Reporting Items for Overviews of Reviews guidelines. ¹² The eligibility criteria were established following the PICOS methodology (Table 1). ¹³ A systematic review was considered to be one that contained the “systematic review” term in the title and/or abstract, had specific inclusion criteria and sources, and contained a description or visualization of the selection process. No supplemental primary studies were included.

Table 1.

Eligibility criteria

	Inclusion	Exclusion
Population	Panoramic radiograms obtained from human subjects	Panoramic reconstructions from CT images
Intervention	Diagnostics conducted by artificial intelligence algorithms	Analyses of no medical relevance
Comparison	Human-based diagnoses	Mixed human-algorithmic diagnoses
Outcomes	Qualitative (exclusion or confirmation) or quantitative (measurements against reference points) AI efficiency	No quantified results
Study design	Systematic reviews without time frame limits	Publication language other than English

Appendix A

PRIOR Checklist

Information sources

Bielefeld Academic Search Engine (BASE) and Google Scholar search engines were used as the ones with the greatest medical coverage according to Gusenbauer. ^14–16 Additionally, Association for Computing Machinery: Guide to Computing Literature (ACM) and National Library of Medicine: PubMed databases were searched. ^17,18

Search strategy

The search of medical databases was carried out on June 1, 2023, according to the following strategy: (“panoramic” OR “dpr” OR “dpt” OR “pantomogram” OR “orthopantomogram” OR “opg”) AND (“ai” OR “intelligence” OR “intelligent” OR “neural” OR “cnn” OR “machine” OR “deep” OR “automated” OR “learning”) AND “systematic” AND “review”. Due to Google Scholar search engine retrieving the most accurate results related to the given query initially, with their relevance gradually decreasing, only the first 200 out of over 100,000 items were retrieved.

Selection process

Deduplication and screening according to the Population and Intervention criteria were performed using the Rayyan tool (Qatar Computing Research Institute, Doha, Qatar and Rayyan Systems, Cambridge, MA). ¹⁹ Titles and abstracts were blindly assessed by three authors (NT, MC, and AB). The inclusion decision by at least one of the judges promoted the report for full-text evaluation (NT and MC). The entire selection process was visualized using a flow diagram. The overlapping issue was omitted at this stage and was addressed in the following steps.

Data collection process

Data from reports were independently extracted by two authors (NT and MC) without the use of automation tools. In cases of inconsistency, the third investigator (KC) had the deciding vote. The data were divided according to the problem assessed by the AI: (1) teeth identification and numbering; (2) detection of periapical lesions; (3) periodontal bone loss; (4) osteoporosis; (5) maxillary sinusitis; (6) dental caries; and (7) other tasks. The overlapping of primary studies was visualized on a timeline created in Microsoft PowerPoint (Microsoft Corporation, Redmond, WA).

Data items

From included systematic reviews, we extracted manually the following variables: (a) sensitivity; (b) specificity; (c) precision; and (d) accuracy of AI models used in the analysis of DPRs, omitting data not available. Data presented in decimal form were converted to percentages. The ranges of the above variables were presented consecutively from smallest to largest in the relevant tables. The arithmetic mean of each variable was shown in parentheses. The average value of individual variables in subsequent years was also calculated and included in the relevant figures.

Mean values of individual variables evaluating the performance of AI algorithms were extracted from each systematic review, if available. In the absence of data, an attempt was made to calculate individual means based on the data contained in the systematic review. Where necessary, the content of source studies was also used. Calculations of average performances were made on the basis of data from studies that met the quality criteria of a given systematic review (satisfactory sample size, low risk of bias).

Risk of bias assessment

In the course of this overview of reviews, the qualification of source studies for individual systematic reviews was not questioned, but only the risk of bias in systematic reviews was assessed. “ROBIS: Tool to assess risk of bias in systematic reviews” was used to evaluate eligible reports. ²⁰ Systematic reviews with a high risk of bias were rejected.

Synthesis methods

The efficiencies of AI in individual tasks were presented in tables and visualized in graphs using Microsoft Office (Microsoft Corporation, Redmond, WA).

Results

Systematic review

A systematic literature review yielded a total of 1021 entries, of which 983 remained after auto-deduplication. Manual deduplication resulted in the deletion of further items, which limited the selection to 933 unique entries. Blind screening by three authors resulted in the removal of 864 items unanimously indicated as non-compliant with the adopted criteria. Thus, 69 articles identified by at least one of the researchers as potentially eligible were evaluated in full text. Of these, 57 were rejected, with reasons given (Table A1). The concordance of decisions at this stage, expressed by Cohen’s κ coefficient, was 0.85, which means an almost perfect agreement. Ultimately, 12 articles were qualified for the synthesis (Figure 3). ^21–32

Figure 3.

PRIOR flow diagram. PRIOR, Preferred Reporting Items for Overviews of Reviews.

Characteristics of systematic reviews

12 eligible articles were published between 2019 and 2023. Data characterizing qualified systematic reviews are presented in Table 2. 'Coverage dates' include the years in which articles were searched for inclusion in a given systematic review. Some authors did not specify a time frame or did not introduce restrictions on the publication period (Figure 4). 'Included dates' cover the full years in which research papers qualified for a given systematic review were published. The oldest analyzed article was from 1988, while the latest was from 2022. AI has often been used to identify teeth and detect caries, osteoporosis, periapical lesions, and periodontal disease.

Table 2.

Characteristics of systematic reviews

First author	Publication date	Coverage dates	Dental field
		(Inclusion dates)
Almășan	2023	To 2022	Diagnosis of temporomandiubular joint osteoarthritis
		(2020–2021)
Chaurasia	2023	2011–2022	Identifying and classifying dental implant systems
		(2020–2022)
Mohammad-Rahimi	2022	2010–2021	Dental caries detection
		(2017–2021)
Revilla-León	2022	N/S, presumably to 2022	Diagnosis of alveolar bone loss
		(1988–2020)
Sadr	2022	2010–2022	Detecting periapical lesions
		(2019–2022)
Singh	2022	2016–2020	Tooth detection and numbering
		(2016–2020)	Dental caries detection
			Periodontal disease detection
			Osteoporosis detection
			Oral lesion detection
			Forensic dentistry
Umer	2022	To 2021	Teeth identification
		(2018–2021)
Revilla-León	2021	To 2021	Implant type recognition
		(2005–2020)	Osteointegration success or implant success prediction
			Implant design optimization
Khanagar	2020	2000–2020	Staging the development of lower third molar
		(2008–2020)	Detecting osteoporosis
			Detection of VRF
			Teeth detection and numbering
			Detecting apical lesions
			Detection of maxillary sinusitis
			Detecting and segmenting the approximate of inferior
			Alveolar nerve to the roots of lower third molars
			Detecting periodontal bone loss
			Gender determination
Prados-Privado	2020	To 2020	Dental caries detection and diagnosis
		(2008–2020)
Hung	2019	To 2019	Detecting osteoporosis
		(1998–2019)	Classification of cysts and tumors
			Tooth detection and numbering
			Detection of maxillary sinusitis
Hwang	2019	To 2018	Automatic teeth segmentation
		(2016–2018)	Tooth detection
			Detecting osteoporosis
			Staging lower third molar development for age estimation

N/S, not specified;VRF, vertical root fracture.

Figure 4.

Overlap of periods searched for systematic reviews (* - not specified, presumed time range visualized).

Primary study overlap

Eligible systematic reviews included articles published between 1988 and 2022 (Figure 5 illustrates the full years covered by each systematic review). Regardless of the wide search date ranges, most authors included only reports published after 2016. The oldest research paper dealt with the use of AI in the detection of periodontal disease. ³³ Over time, this technology began to be used also in the detection of periapical lesions, dental caries, osteoporosis, and tooth numbering.

Figure 5.

Overlap of primary studies included in systematic reviews.

Risk of bias in systematic reviews

The results of the assessment of the risk of bias are presented in Table 3. The distribution of the assessment results in individual domains is illustrated in Figure 6. One systematic review was disqualified from the quantitative syntheses due to the high risk of bias.

Table 3.

Risk of bias in systematic reviews

First author	Domain 1: study eligibility criteria	Domain 2: identification and selection of studies	Domain 3: data collection and study appraisal	Domain 4: synthesis and findings	Risk of bias in the review
Almășan
Chaurasia
Mohammad-Rahimi
Revilla-León (2022)
Sadr
Singh
Umer
Revilla-León (2021)
Khanagar
Prados-Privado
Hung
Hwang

“+”, low; “?”, unclear; “X”, high.

Figure 6.

Risk of bias distribution (Domain 1—study eligibility criteria; Domain 2—identification and selection of studies; Domain 3—data collection and study appraisal; Domain 4—synthesis and findings).

Synthesis of results

After assessing the risk of bias, 11 systematic reviews were qualified for the quantitative syntheses. The syntheses were carried out in seven domains: (1) detection of dental caries; (2) osteoporosis; (3) periapical lesions; (4) periodontal bone loss; (5) maxillary sinusitis; (6) teeth identification and numbering; and (7) other tasks.

Teeth identification and numbering

Four systematic reviews analyzed the use of AI in teeth identification and numbering on DPRs. 10 included articles published between 2017 and 2021 were qualified for synthesis (Table 4, Figure 7).

Table 4.

Outcome metrics of AI algorithms in teeth identification and numbering

First author	Years of publications	Sensitivity (mean)	Specificity (mean)	Precision (mean)	Accuracy (mean)
Hung	2019	98.7% (98.7%)	99.94% (99.94%)	N/S	99.87% (99.87%)
Hwang	2017	NS	N/S	N/S	91.74% (91.74%)
Khanagar	2019	98.7% (98.7%)	N/S	99.45% (99.45%)	N/S
Umer	2020–2021	96–98% (97%)	97% (97%)	85–99% (95.17%)	89–96% (94.83%)

AI, artificial intelligence; N/S, not specified.

Figure 7.

The number of reports on teeth identification and numbering and outcome metrics of AI algorithms by years. AI, artificial intelligence.

Detection of periapical lesions

Two systematic reviews analyzed the use of AI in the detection of periapical lesions on DPRs. Five included articles, published between 2019 and 2022, were qualified for the synthesis (Table 5, Figure 8).

Table 5.

Outcome metrics of AI algorithms in detecting periapical lesions

First author	Years of publications	Sensitivity (mean)	Specificity (mean)	Precision (mean)	Accuracy (mean)
Khanagar	2019	65% (65%)	87% (87%)	N/S	N/S
Sadr	2019–2022	48–92% (64%)	87–99.95% (93.48%)	49–84% (66%)	N/S

AI, artificial intelligence; N/S, not specified.

Figure 8.

The number of reports on detection of periapical lesions and outcome metrics of AI algorithms by years. AI, artificial intelligence.

Detection of periodontal bone loss

Six articles from two systematic reviews published between 2019 and 2020 analyzed the use of AI in the detection of periodontal bone loss on DPRs (Table 6, Figure 9).

Table 6.

Outcome metrics of AI algorithms in detecting periodontal bone loss

First author	Years of publications	Sensitivity (mean)	Specificity (mean)	Precision (mean)	Accuracy (mean)
Khanagar	2019	N/S	N/S	N/S	81% (81%)
Revilla-León (2022)	2019–2020	76–84% (80.33%)	81–93.75% (87.58%)	N/S	81–94.18% (89.06%)

AI, artificial intelligence; N/S, not specified.

Figure 9.

The number of reports on detection of periodontal bone loss and outcome metrics of AI algorithms by years. AI, artificial intelligence.

Detection of osteoporosis

Osteoporosis detection on DPRs with AI models was analyzed in 13 reports from three systematic reviews. The articles were published between 2007 and 2020 (Table 7, Figure 10).

Table 7.

Outcome metrics of AI algorithms in detecting osteoporosis

First author	Years of publications	Sensitivity (mean)	Specificity (mean)	Precision (mean)	Accuracy (mean)
Hung	2007–2017	76.8–99.1% (93.06%)	43.8–98.4% (82.86%)	N/S	62–98.9% (87.31%)
Hwang	2018	N/S	N/S	N/S	89.29% (89.29%)
Khanagar	2018–2020	N/S	N/S	N/S	N/S

AI, artificial intelligence.

Figure 10.

The number of reports on detection of osteoporosis and outcome metrics of AI algorithms by years. AI, artificial intelligence.

Detection of maxillary sinusitis

Two systematic reviews analyzed the use of AI in the detection of maxillary sinusitis on DPRs. Two included articles published between 2016 and 2019 were qualified for the synthesis (Table 8, Figure 11).

Table 8.

Outcome metrics of AI algorithms in detecting maxillary sinusitis

First author	Years of publications	Sensitivity (mean)	Specificity (mean)	Precision (mean)	Accuracy (mean)
Hung	2016	77.6% (77.6%)	69.4% (69.4%)	N/S	73.5% (73.5%)
Khanagar	2019	86.7% (86.7%)	88.3% (88.3%)	N/S	87.5% (87.5%)

AI, artificial intelligence; N/S, not specified.

Figure 11.

The number of reports on detection of maxillary sinusitis and outcome metrics of AI algorithms by years. AI, artificial intelligence.

Detection of dental caries

The use of AI in dental caries detection on DPRs was analyzed in two systematic reviews. Four included articles were published between 2017 and 2021 (Table 9, Figure 12).

Table 9.

Outcome metrics of AI algorithms in detecting dental caries

First author	Years of publications	Sensitivity (mean)	Specificity (mean)	Precision (mean)	Accuracy (mean)
Mohammad-Rahimi	2020–2021	N/S	86% (86%)	78.5–89.4% (84.97%)	86–96% (89.67%)
Prados-Privado	2017	N/S	N/S	N/S	86% (86%)

AI, artificial intelligence; N/S, not specified.

Figure 12.

The number of reports on detection of dental caries detection and outcome metrics of AI algorithms by years. AI, artificial intelligence.

Other tasks

In addition to the above-mentioned problems solved by AI algorithms, the analyzed systematic reviews also included others, such as classifying cysts and tumors, detecting vertical root fractures, temporomandibular joint (TMJ) osteoarthritis and the use of AI in implant dentistry and forensic dentistry (Table 10). In case of the systematic review of Almășan et al, results excluding indeterminate TMJ osteoarthritis diagnosis and without fine-tuning were analyzed. ³¹

Table 10.

Outcome metrics of AI algorithms in other analyzed tasks

First author	Dental field	Years of publications	Sensitivity (mean)	Specificity(mean)	Precision (mean)	Accuracy (mean)
Almășan	Diagnosis of temporomandiubular joint osteoarthrosis	2020–2021	39–94% (62%)	77–91% (83%)	78% (78%)	77–88% (81%)
Chaurasia	Identifying and classifying dental implant systems	2020–2022	N/S	N/S	N/S	70.8–98.2% (86.11%)
Hung	Classification of cysts and tumors	2013	N/S	N/S	N/S	81.8–88.9% (85.72%)
Hwang	Automatic teeth segmentation	2018	N/S	N/S	79% (79%)	N/S
	Staging the development of lower third molar for age estimation	2017	N/S	N/S	N/S	51% (51%)
Khanagar	Staging the development of lower third molar for age estimation	2017	N/S	N/S	N/S	51% (51%)
	Detection of vertical root fracture	2019	N/S	N/S	93% (93%)	N/S
	Detecting and segmenting the approximate of inferior alveolar nerve to the roots of lower third molars	2019	N/S	N/S	N/S	N/S
Revilla- León (2021)	Implant type recognition	2015–2020	N/S	N/S	N/S	93.8–98%
	Osteointegration success or implant success prediction	2005–2020	N/S	N/S	N/S	62.4–80.5%
	Implant design optimization	2009–2019	N/S	N/S	N/S	N/S

AI, artificial intelligence; N/S, not specified.

Discussion

General interpretation of the results

In recent years, the number of publications on AI rose notably. Most articles on the use of this technology in the analysis of DPRs were published in 2019 and beyond. Various outcome metrics were used to assess the effectiveness of AI systems. The most common ones were accuracy, sensitivity, specificity, and precision, which definitions are in Table 11. ^23,34

Table 11.

Definitions of outcome metric of evaluated AI models

Metric	Formula	Definition
Sensitivity = TP rate	$\frac{TP}{TP+FN}$	Probability of correct positive predictions in actual positives
Specificity = TN rate	$\frac{TN}{TN+FP}$	Probability of correct negative predictions in actual negatives
Precision	$\frac{TP}{TP+FP}$	Probability of correct positive predictions in positive results
Accuracy	$\frac{TP+TN}{AS}$	Probability of correct predictions in total number of predictions

AS, all samples; FN, false-negatives; FP, false-positives; TN, true-negatives; TP, true-positives.

The lowest reported sensitivity (48%) of the AI algorithm was noted in the review of Sadr et al, where segmentation with the data set size of 1300 DPRs was used to detect periapical lesions. However, its specificity was the highest (99.95%) among all studies in analyzed systematic reviews. ²³ The highest reported sensitivity (99.1%) occurs in the review of Hung et al, where a neural network developed on 141 DPRs from normal and osteoporotic female subjects was used to detect osteoporosis. ²⁹ The lowest specificity (43.8%) was also noted in the review of Hung et al. This outcome comes from the article from 2008, where the discriminant technique was used to detect osteoporosis. The data set used to develop this AI model consisted of 100 images from normal, low body mass density, and osteoporotic female subjects. ²⁹

The lowest reported precision (49%) appeared in the review of Sadr et al, where a 7-layer CNN based on a synthesized data set of more than 2000 tooth segments from panoramic radiographs was used to detect periapical lesions. ²³ The highest reported precision (99.45%) and accuracy (99.87%) occur in the review of Hung et al, where a data set of 1352 DPRs of adults was used to train the deep CNN. ²⁹

The lowest accuracy (51%) was noted in an article included in two systematic reviews (Khanagar et al and Hwang et al.). In the paper, the use of CNN to stage lower third molar development on DPRs for age estimation was assessed. 200 images were used for testing; however, the results were similar to those of the trained examiners. ^27,30

In general, outcome metrics of AI algorithms improved over time.

The results concerning teeth identification and numbering seem surprising. In 2020, there was a decrease in sensitivity by 2.84%, specificity by 2.94%, precision by 5.82%, and accuracy by 6.2%. Nevertheless, all these values were still very high, above 93%. The following year, there was an increase in sensitivity to 98% and specificity to 99%, but this result is based on only one study from 2021.

The specificity of detection of periapical lesions increased by 12.95% within 2 years. The sensitivity initially decreased from 65 to 48% for two consecutive years but then increased to 92% in 1 year. Nevertheless, only five studies were qualified for this domain, two from 2019 and one each from 2020, 2021, 2022. In both studies, from 2021 and 2022, U-Net type deep CNNs were trained to detect periapical lesions, but there was a difference in dataset size— 1300 radiographs (2021) and 470 radiographs (2022).

Six studies covered the detection of periodontal bone loss, three each from 2019 and 2020. In 1 year, there was an increase in accuracy by 12.09%, sensitivity by 5.5%, and specificity by only 0.62%.

The accuracy of detecting osteoporosis continued to grow until 2016, reaching 97.45%. In the next 2 years, there was a decrease of 8.16%, but generally, in 10 years, the accuracy rose by 27.29%.

In case of maxillary sinusitis detection, between 2016 and 2019, there was an increase in accuracy by 14%, sensitivity by 9.1%, and specificity by 18.9%. However, these differences were calculated based on only two studies concerning this domain.

For dental caries detection, the average accuracy did not change between 2017 and 2020 reaching 86%, but then increased by 5.5% in 1 year.

Teeth identification and numbering

Filling digital data charts may be time-consuming for many clinicians. Teeth identification refers to the detection of teeth in DPR. It facilitates dentists' jobs as they do not have to manually enter the details, which can result in higher efficiency. However, teeth identification may be challenging for AI models due to overlapping anatomical boundaries. Hwang et al included in the review the study from 2017, where the conventional CNN approach was used for tooth detection. The method was tested on a dataset containing 100 DPRs, achieving 91.74% average accuracy. ³⁰ Khanagar et al and Hung et al included the same study from 2019, where CNN-based AI system was used for teeth detection and numbering. The performance of this system was comparable to the level of experts. ^27,29 The results presented by Umer show that using CNN in teeth identification is promising. Nevertheless, there are some limitations, such as a high risk of bias and heterogeneity. ²⁵

Detecting periapical lesions

The most common radiographic findings associated with teeth are periapical lesions, the usual symptom of bacterial infection of the root canal system. ³⁵ As many of them are asymptomatic, frequently, they are incidentally diagnosed on radiographs taken routinely in the dental office. Their early detection is crucial, especially in patients with comorbidities. Khanagar et al included in the review the study where a system based on deep CNN turned out to be a successful tool for detecting apical lesions. ²⁷ Sadr et al included four studies using classification and segmentation methods, which showed relatively high specificity and sensitivity in detecting these radiolucent pathologies. However, the level of accuracy decreases due to superimposition and lack of homogeneity of two-dimensional radiographs. Ill-defined borders and the small size of lesions can influence the process of the detection of pathologies as well. ²³

Detecting periodontal bone loss

Periodontal diseases are highly prevalent and can affect up to 90% of the population all over the world. ³⁶ They are mainly a result of inflammation of the gums and surrounding bone, which can lead to early loss of teeth. Currently, many AI software can measure bone loss which helps clinicians reduce their diagnostic efforts. Khanagar et al included in systematic review a report from 2019, where a system based on CNN demonstrated similar results to six experienced dentists in detecting periodontal bone loss. ²⁷ Revilla-León et al included 11 studies that evaluated AI models for detecting periodontal bone loss from radiographic images, but only five analyzed panoramic radiographs. The average accuracy, sensitivity, and specificity suggest that AI models may be a powerful tool for diagnosing periodontal diseases. ²² Observing the effectiveness of currently used algorithms, differentiating extraction socket from medication-related osteonecrosis of the jaw also seems possible with the use of AI models, allowing their faster treatment. ³⁷

Detecting osteoporosis

Osteoporosis is a bone disease that occurs when mineral bone density decreases or the structure of bone changes. ³⁸ Orthopantomographic indices are considered tools for early screening of this condition. Hung et al included in his review nine studies that were published between 2007 and 2017. The lowest sensitivity, specificity, and accuracy occurred in the study from 2008, whereas the highest values of these metrics were noted in the study from 2016. ²⁹ Hwang et al included two studies from 2018 where deep CNN and Octuplet Siamese Network were used to detect osteoporosis with high accuracy. ³⁰ Khanagar et al also included two studies from 2018 and 2020, where deep CNN turned out to be a reliable tool for automated osteoporosis screening. ²⁷

Diagnosis of maxillary sinusitis

Maxillary sinusitis is the inflammation of the paranasal sinuses, which can be a result of bacterial, viral, or fungal infection. ³⁹ Undiagnosed and untreated sinusitis can lead to potentially life-threatening conditions like meningitis, vision changes, or olfactory dysfunction. ⁴⁰ Khanagar et al included in the review the study where the deep learning system showed higher diagnostic performance than two experienced radiologists. ²⁷ Hung et al described a study in which an AI technique called asymmetry analysis was used to support inexperienced dentists in diagnosing maxillary sinusitis. The diagnostic performance increased with the support of this computer-aided detection system. ²⁹ Above results lead to the statement that deep learning systems may be valuable diagnostic support, especially for inexperienced clinicians.

Dental caries detection

Dental caries involves the majority of the population. Dentists usually diagnose caries lesions on panoramic radiographs relying only on their visual inspection. Detection of early lesions is vital in maintaining oral health but sometimes may be challenging. ⁴¹ It is frequently difficult to identify caries lesions due to low image quality. Deep learning models may improve accuracy and support dentists in detecting them. All included studies in the systematic review of Mohammad-Rahimi et al were based on CNN, but different model structures were applied—PaXNet. MobileNet V2, and AlexNet. They showed relatively high mean accuracy, specificity, and sensitivity above 80%. ²¹ Prados-Privado et al included only one study using DPRs from 2017, in which a system based on Radon Transformation and Discrete Cosine Transformation was used to detect caries lesions. ²⁸

Other tasks

AI models can also be used in detecting TMJ osteoarthrosis, a chronic degeneration of hard and soft tissues around the joint, in which early diagnosis is vital to effective treatment planning. ⁴² Almășan et al included in the review three studies where authors used Residual Neural Network, a CNN VGG16, Inception V3, and Efficient Net-B to detect this disease on panoramic radiographs. The accuracies of methods varied between moderate to good and were higher when primary indeterminate diagnoses of TMJ osteoarthrosis were excluded. ³¹

Khanagar et al included in the review studies that use CNN in forensic odontology. DPR can be used in this field to estimate a person’s age based on the development of the lower third molar. The results of the studies are promising, which indicates that AI can be helpful in forensic dentistry. CNN was also successfully used to detect vertical root fractures on DPR, showing the high precision of the AI model and gaining a comparable level of performance as the experts. Deep learning can help with the segmentation of anatomical structures as well, for instance, in detecting the proximity of the inferior alveolar nerve to the roots of lower third molars, which is a risk factor for the occurrence of nerve damage. The results of the study were promising. However, further improvement of the system is advised. ²⁷ Hung et al included two studies where a support vector machine was used to detect cysts and tumors. The average accuracy of 85.72% indicates that cysts and tumors can be effectively diagnosed using AI techniques. ²⁹

Revilla-León et al included 17 studies where AI was applied in implant dentistry. This technology supports clinicians in implant type recognition, predicting osteointegration success, as well as helps to optimize implant designs, minimizing the stress at the implant–bone interface by 36.6% compared with the finite element analysis calculations. ²⁶ Chaurasia et al also analyzed the performance of DL models in identifying and classifying dental implant systems (DISs), which may be helpful to avoid unintended iatrogenic complications in case of, e.g. screw loosening or PI. In this systematic review, five studies used only panoramic images in recognizing up to 12 different DISs. The accuracy of these architectures was between 70.8 and 98.2%. Therefore, these DL models can be potentially used to facilitate the decision-making process for dentists. ³²

Limitations

Limitations of the evidence

Studies included in the analyzed systematic reviews used different neural networks and performance measurements. Because of this heterogeneity and limited quality of reporting in the studies, comparison of the outcomes was difficult and did not allow the meta-analysis to be performed. It would be recommended for future studies to apply standards like CLAIM—a checklist for AI in medical imaging, or STARD-AI for diagnostic studies using AI models. ^43,44

There was also a significant variety in the size of training data sets—from dozens to more than a thousand. Theoretically, the larger the data set, the more precise the AI algorithm will be. ⁴⁵ Comparing algorithms that differ in such a large amount of training data as well as applying different model structures are another limitations when comparing their effectiveness.

Prados-Privado et al drew attention to the importance of the definition of caries in analyzed studies. Depending on whether caries was diagnosed according to ICDAS II or defined as a loss of mineralization, the accuracy differed by up to 17%. ^28,46 Unfortunately, most studies did not specify how they described caries lesions.

There was no information in the analyzed systematic reviews about the dental X-ray machines’ vendors used in included studies. Various models have different technical specifications, which affect image quality. Optimal selection of image resolution has the potential for increasing NN performance. ⁴⁷

The review of Singh et al was disqualified from the quantitative syntheses due to the high risk of bias. ²⁴ It turned out to be not systematic, which was manifested by imperfections primarily in the domains identification and selection of studies and data collection and study appraisal. The search was performed using only one database when it is advised to explore multiple databases for a systematic review to identify available literature concerning the analyzed problem adequately. ⁴⁸ There is no information if the process of screening titles, abstracts, and full texts of manuscripts was performed independently by at least two reviewers, which minimizes errors in the selection of the studies. The authors did not assess the risk of bias of included articles that could establish transparency of findings and is an essential component of any review. ⁴⁹

Limitations of the review process

In this overview of reviews, non-English articles were rejected. Search queries were only in English as well.

Conclusions

According to the results of this overview of reviews, in the years 1988–2023, there was a significant development of information technologies for the analysis of DPRs. The effectiveness of AI algorithms has an upward trend. The latest analyzed AI models achieve high accuracy in detecting caries—91.5%, osteoporosis—89.29%, maxillary sinusitis—87.5%, periodontal bone loss—93.09%, and teeth identification and numbering—93.67%. The detection of periapical lesions is also characterized by high sensitivity (99.95%) and specificity (92%). The above results indicate that AI applications can significantly support dentists. However, due to the small number of heterogeneous source studies synthesized in systematic reviews, the results of this overview should be interpreted with caution. As systematic reviews in AI become outdated quickly, their regular updating is recommended.