AI Detects Dental Caries
Overview
Dental caries, a largely preventable condition, remains a significant global health concern. Numerous systematic reviews have explored the performance of artificial intelligence (AI) models in diagnosing and detecting dental caries. This umbrella review aimed to consolidate findings from these systematic reviews to evaluate the application and effectiveness of AI in dental caries detection and diagnosis.
A comprehensive search of databases, including MEDLINE/PubMed, IEEE Explore, Embase, and the Cochrane Database of Systematic Reviews, was conducted to identify relevant studies. After screening 1249 entries against eligibility criteria, seven studies were selected for analysis. Two independent reviewers assessed the studies, and findings were summarized in tabular format and analyzed narratively.
The most commonly utilized AI algorithms were multilayer perceptrons, support vector machines (SVMs), and neural networks. These algorithms were designed for various tasks, including segmentation, classification, caries detection, diagnosis, and prediction, using diverse data sources such as periapical and panoramic radiographs, smartphone images, bitewing radiographs, and near-infrared light transillumination images. Convolutional neural networks (CNNs) exhibited superior performance in caries detection, segmentation, and classification, showing high sensitivity, specificity, and area under the curve (AUC).
AI models demonstrated enhanced accuracy in detecting and diagnosing dental caries when paired with periapical and panoramic radiographic images, suggesting their potential to improve diagnostic precision in clinical and research settings.
Introduction
Dental caries, commonly referred to as tooth decay, is a condition caused by the demineralization of tooth structures due to sugar-driven cariogenic plaque bacteria. Despite being largely preventable, dental caries remains a significant global health concern, affecting approximately two billion adults and 514 million children worldwide, with Southeast Asia and the Western Pacific witnessing disproportionate impacts among impoverished and socially disadvantaged populations. Economically, the global cost of treating dental caries in both permanent and deciduous teeth is estimated to reach $22 billion annually.
Traditional diagnostic methods, such as intraoral periapical radiography (IOPAs) and radiovisiography (RVGs), are widely used in clinical practice to visually assess the severity of carious lesions. While these radiographic techniques can effectively detect early-stage lesions, their diagnostic accuracy is often compromised by subjectivity, relying heavily on the examiner’s expertise and experience. Consequently, they are prone to false-positive and false-negative results.
Digital health technologies, including artificial intelligence (AI), are rapidly emerging as promising tools to enhance the accuracy and efficiency of dental diagnoses. AI, a domain of computer science mimicking human cognitive functions, has garnered substantial attention in dental research. Various studies have demonstrated the potential of AI models, such as convolutional neural networks (CNNs) and artificial neural networks, in diagnosing, classifying, segmenting, and predicting dental caries. These findings are further synthesized in systematic reviews highlighting AI’s growing role in dentistry.
Despite encouraging evidence supporting the utility of AI, its integration into routine dental care remains limited, as the field is still in its early stages. An umbrella review is necessary to consolidate insights from systematic reviews, offering a comprehensive understanding of AI’s application in dental caries detection and diagnosis. Such a review aims to guide dental practitioners on the practical use of AI, assist researchers in identifying gaps in knowledge, and educate future dental professionals on innovative practices. Furthermore, it provides policymakers with evidence-based recommendations for responsible AI adoption, ultimately enhancing patient outcomes and streamlining the diagnostic process in dentistry.
Methods
An umbrella review was conducted to synthesize findings from systematic reviews on the application and effectiveness of artificial intelligence (AI) models for diagnosing and detecting dental caries. The study employed the PICO framework, focusing on patient dental image datasets, AI-based models, comparators such as conventional methods or alternative algorithms, and outcome metrics including accuracy, sensitivity, specificity, and area under the curve (AUC). This review adhered to guidelines from the Joanna Briggs Institute (JBI) and PRISMA, with the protocol registered under PROSPERO.
Comprehensive searches were performed on August 18, 2023, using databases such as MEDLINE/PubMed, IEEE Explore, Embase, and the Cochrane Database of Systematic Reviews. Google Scholar was also utilized to retrieve gray literature, with a focus on the top 100 relevant results. Additional studies were identified through backward and forward reference screening. Search terms were carefully curated to align with systematic reviews targeting dental caries and AI-based interventions, with detailed search strategies included in Supporting Information S1.
The inclusion criteria encompassed systematic reviews with or without meta-analysis that evaluated the effectiveness of AI-based methods in managing dental caries. Eligible studies needed to report at least one performance metric, such as accuracy, sensitivity, specificity, or AUC. There were no restrictions on data types, publication dates, settings, or languages.
Exclusion criteria included reviews without critical appraisal or robust methodologies (e.g., scoping reviews, editorials, preprints, commentaries, conference abstracts, or posters) and reviews reliant on secondary sources rather than original research.
Study selection was conducted in two stages: independent screening of titles and abstracts, followed by full-text review. Data extraction was performed using a standardized, pretested form to ensure accuracy and consistency. The quality of the included reviews was assessed using the JBI Critical Appraisal Checklist. Any disagreements during the process were resolved through consultation with senior experts.
The findings of this review provide a consolidated overview of AI’s potential in dental caries diagnosis, with a focus on improving clinical practice, identifying research gaps, and guiding policymakers on the responsible integration of AI in dental care.
Results
The review systematically evaluated the use of artificial intelligence (AI) models in diagnosing and detecting dental caries by analyzing seven high-quality systematic reviews published between 2020 and 2023. These reviews were selected from an initial pool of 1,249 entries sourced from multiple literature databases, including MEDLINE/PubMed, EMBASE, and Scopus. Using Rayyan software, 45 duplicate entries were removed, leaving 1,204 studies for title and abstract screening. This process resulted in the exclusion of 446 papers, and the subsequent full-text review of 20 studies led to the elimination of 13 that either lacked systematic review methodology or did not focus on dental caries. The remaining seven studies were selected based on strict inclusion criteria, ensuring relevance and quality.
The included studies represented diverse geographical contributions, involving researchers from Canada, Pakistan, Spain, and Brazil, among others. Five of the reviews were published in 2022, reflecting a surge of interest in AI applications in dentistry. Some of these studies, such as those by Mohammad-Rahimi et al. (2022) and Revilla-León et al. (2022), involved international collaboration, emphasizing the global interest in this field.
Methodological Approaches and Protocol Adherence
Only three studies had registered their protocols with the PROSPERO registry, a measure that ensures transparency and rigor in systematic reviews. Four reviews explicitly adhered to the PRISMA guidelines for systematic reviews, while two followed the PRISMA-DTA (Diagnostic Test Accuracy) guidelines, specifically tailored for diagnostic studies. One study did not specify adherence to reporting guidelines, potentially limiting its methodological transparency. All included reviews analyzed original research, and two also incorporated conference proceedings, highlighting a broad spectrum of data sources.
Language and Timeframe Variability
Language restrictions varied across studies. While three reviews limited their scope to English-language publications, three imposed no language restrictions, allowing for a more inclusive evidence base. The timeframe for included studies also differed, with some reviews considering research from inception and others focusing on the past 10–12 years. One study did not specify a timeframe, which could affect the comprehensiveness of its findings.
Also read Dental Extraction Prophylaxis Treatment With Antibiotics Is Controversial
PICO Framework Analysis
The PICO framework was applied across all studies, demonstrating some variability in focus. The population typically included radiographic image datasets, while interventions often comprised AI models, neural networks, and machine learning techniques. Comparator groups ranged from expert clinical judgments to other diagnostic models or no comparators. Outcome variables such as accuracy, sensitivity, specificity, and the area under the curve (AUC) were reported, with accuracy emerging as the most frequently assessed metric. Notably, one study uniquely compared different machine learning techniques for predicting caries, offering a nuanced perspective.
Search and Quality Appraisal Techniques
The included reviews utilized diverse database search strategies, with MEDLINE/PubMed, EMBASE, and Scopus being the most commonly employed. Reference list checking and citation analysis were also frequently used to ensure comprehensive study identification. Methodological quality assessment tools varied among studies, including the Cochrane Risk of Bias Tool, the JBI Critical Appraisal Checklist, and QUADAS-2. One study conducted quality assessments without disclosing the specific tool used, which could raise questions about its methodological rigor. Due to the heterogeneity of AI models and algorithms across the primary studies, all reviews adopted narrative synthesis methods.
Significant Observations
The systematic reviews highlighted the increasing use of AI in dentistry, focusing on its potential to enhance diagnostic accuracy and efficiency. Neural networks and machine learning models emerged as prominent interventions, showcasing their ability to analyze radiographic images and predict caries. The variability in outcomes and methodologies underscores the complexity of integrating AI into standard dental care.
Implications for Future Research and Practice
This umbrella review underscores the need for standardized methodologies and comprehensive reporting to advance AI applications in dental care. It also highlights gaps in the current literature, such as limited geographic representation in primary studies and inconsistent adherence to reporting guidelines. Future research should prioritize standardized protocols, broader population inclusion, and comparative studies to evaluate AI’s practical integration into clinical workflows. Such efforts could inform policymakers, clinicians, and educators, ultimately improving patient outcomes and streamlining caries detection practices.
The included reviews covered a broad range of retrieved studies, numbering between 133 and 3,410. Among these, the studies ultimately included in the analyses varied from 12 to 42. Data set sizes utilized for training and validating AI algorithms ranged widely, from as few as 32 data points to as many as 12,600, reflecting significant variation in sample sizes (Supporting Information S1: Table S3). The studies employed diverse types of dental images for algorithm development, including periapical radiographs (used in 4 studies), bitewing radiographs (6 studies), near-infrared light transillumination (NILT) images (5 studies), intraoral or oral photographs (4 studies), panoramic radiographs (3 studies), smartphone photos (2 studies), optical coherence tomography (OCT) images (2 studies), cone-beam computed tomography (CBCT) images (2 studies), and other medical records or data sets (2 studies). This diversity highlights the breadth of imaging modalities leveraged to enhance AI model performance.
The most commonly used AI algorithms across the included studies were multilayer perceptrons (3 studies), support vector machines (SVM) (3 studies), convolutional neural networks (CNN) (6 studies), and artificial neural networks (ANN) (3 studies). These algorithms were employed to execute tasks such as classification (6 studies), segmentation (4 studies), caries detection (4 studies), caries diagnosis (4 studies), and caries prediction (2 studies). Despite the wide array of approaches, only two studies explicitly identified the forms of caries detected, including precavitated lesions, initial caries, occlusal, proximal, root, enamel, and dentinal lesions, mainly affecting premolar and molar teeth (Prados-Privado et al. 2020; Talpur et al. 2022).
Performance Metrics of AI Algorithms
Six of the included studies evaluated the accuracy of caries detection models, reporting a wide range of results. Accuracy rates spanned from 65.7% for RetinaNet to 99.1% for VGGNet-16. Among the various imaging types, periapical and panoramic radiographs demonstrated the highest accuracy, with ranges of 82%–99.2% and 86.1%–96.1%, respectively. In contrast, near-infrared transillumination images showed lower accuracy, ranging between 68.0% and 78.0%.
Sensitivity, which measures the true positive rate, was reported in four studies. Deep learning models exhibited the highest sensitivity across caries detection, segmentation, and classification tasks, with values ranging from 25% to 99.7% (Mohammad-Rahimi et al. 2022). Similarly, neural networks achieved sensitivity rates between 59% and 99.6% (Khanagar et al. 2022). Specificity, indicating the true negative rate, was reported in five studies, with custom CNN models displaying the highest specificity at values ranging from 81% to 100% (Moharrami et al. 2024).
The area under the curve (AUC) metric, which evaluates the overall efficacy of a binary classification model, was documented in four studies. Neural networks achieved AUC values ranging from 0.69 to 0.99 (Khanagar et al. 2022; Prados-Privado et al. 2020). Machine learning models, in contrast, reported AUC values ranging from 0.740 to 0.987 for caries diagnosis, 0.74 to 0.917 for proximal caries lesion detection, 0.857 to 0.987 for caries classification, and 0.836 to 0.856 for segmentation tasks (Reyes et al. 2022).
Positive predictive values (PPV), which represent the likelihood that positive test results are true positives, were reported in only two studies (Khanagar et al. 2022; Prados-Privado et al. 2020). These ranged from 63% to 99%. Similarly, negative predictive values (NPV), or the likelihood that negative test results are true negatives, were disclosed in two studies (Khanagar et al. 2022; Moharrami et al. 2024), with values ranging from 73% to 98.15%.
Concluding Observations
The findings indicate that AI models, particularly CNNs and neural networks, exhibit strong potential for accurate caries detection and classification. The performance metrics across studies reveal robust sensitivity, specificity, and AUC values, underscoring their utility in dental diagnostics. However, the variability in imaging modalities, sample sizes, and algorithmic approaches suggests that standardization in methodology and evaluation criteria is crucial for future advancements in the field.
