Artificial Intelligence Tools for Dental Caries Detection: An Exploratory Systematic Review

1. Introduction

Dental caries continues to represent a major public health problem in Chile, with a prevalence of 73.9% among 15-year-olds and 99.2% in adults aged 35 to 44 years [1]. Traditionally, its detection and diagnosis have been based on conventional methods such as visual inspection [2], the use of dental probes [3], and radiographic imaging [3]. Although these approaches remain widely implemented in daily practice, they present inherent limitations in terms of sensitivity, reproducibility, and standardization. In this context, artificial intelligence (AI) has emerged as a promising technological innovation, relying on algorithms capable of simulating human learning and decision-making processes [4] for the analysis of a wide range of dental images [5].

The incorporation of AI into dentistry has demonstrated multiple advantages. These include reductions in diagnostic time [5], decreases in error rates [6], and improvements in diagnostic accuracy [7], all of which contribute to optimized treatment planning [8], enhanced clinical efficiency [3], and the reduction of risks associated with diagnostic uncertainty [7]. More specifically, techniques such as machine learning (ML), artificial neural networks (ANN) and convolutional neural networks (CNN) have exhibited the ability to process and analyze large volumes of data with high precision [9], accelerate evaluation processes [5,10], reduce inter-professional variability [4], facilitate the early identification of incipient lesions that allow for minimally invasive interventions [11], and ultimately improve the overall efficiency of dental practice [12].

Nevertheless, despite the considerable potential of AI, its implementation in dentistry continues to face significant challenges. The quality, diversity, and representativeness of training data are decisive factors, since biased or insufficient datasets may compromise the performance and generalizability of algorithms [4]. Moreover, appropriate technological infrastructure and professional training are indispensable for the effective adoption of these tools [13]. In parallel, the establishment of robust regulatory frameworks is required to promote responsible and ethical integration of AI, ensuring both patient safety and the preservation of professional standards. This scenario explains the heterogeneity observed in the existing literature, which complicates the systematic synthesis of knowledge regarding these emerging technologies.

The justification for addressing this problem stems from the pressing need to overcome such barriers while capitalizing on the proven benefits of AI to transform dental practice. A systematic analysis and organization of the available scientific evidence are therefore essential to maximize the clinical applicability of these tools and to contribute to the global advancement of oral health. Accordingly, the objective of this study is to provide a comprehensive overview of the current landscape of AI in the detection of dental caries, emphasizing its potential to reshape conventional diagnostic paradigms and strengthen evidence-based decision making in dentistry.

In alignment with this objective, the study specifically aims to: (i) compile the scientific evidence published in the last five years on AI tools for the detection of dental caries; (ii) compare the architectures employed in the development of the algorithms; (iii) analyze the functioning and reported performance of AI tools; and (iv) examine the preprocessing methods applied in the included studies.

2. Materials and Methods

This scoping review was conducted following the Preferred Reporting Items for Systematic Reviews and Meta-Analyses extension for Scoping Reviews (PRISMA-ScR) guidelines. The present review incorporated academic publications and scientific articles published within the last five years, indexed in PubMed, Scopus, and Web of Science (WoS), and available up to June 2024. Eligible studies were required to investigate the use of AI for the detection of dental caries, provide access to the full text, and be published in either English or Spanish. In contrast, opinion pieces, editorials, review articles, and studies that employed AI for pedagogical purposes were excluded, as were articles related to periodontics, endodontics, or oral cancer.

The search was conducted in June 2024 using the electronic databases PubMed, Scopus, and Web of Science. The strategy was based on standardized Medical Subject Headings (MeSH) terms, specifically “Artificial Intelligence” and “Dental Caries,” combined with the Boolean operator “AND,” resulting in the query “Artificial Intelligence AND Dental Caries.” This process generated a set of records that, after the removal of duplicates across databases, were subjected to an initial screening by title to retain those potentially relevant to the research question. Subsequently, the remaining publications underwent a more detailed evaluation, during which studies that did not meet the predefined inclusion criteria or for which full-text access was not available were excluded.

The selection process was independently conducted by two authors, and in cases where discrepancies arose, these were resolved through discussion with a third author. The remaining studies were then subjected to abstract screening, after which only those that directly addressed the research question were retained. The final pool of documents underwent critical appraisal in accordance with the TRIPOD-AI guidelines [14], with particular emphasis on methodological rigor, ethical considerations, and the clarity and reproducibility of the reported results, ensuring that each study adequately addressed the research objectives. The use of TRIPOD-AI served as a framework for bias control, ensuring that the included studies provided transparent, reliable, and verifiable evidence that adequately addressed the research objectives.

3. Results

The systematic search identified a total of 617 records, of which 159 were retrieved from PubMed, 290 from Scopus, and 166 from Web of Science, along with two additional studies identified outside the automated search strategy (Figure 1).

After the removal of duplicate entries across databases, the records were screened by title, resulting in the selection of publications that were potentially relevant to the research question. In the subsequent evaluation phase, studies that did not meet the inclusion criteria were excluded, as well as those for which full-text access was not available. Following abstract review, only studies that directly addressed the research objectives were retained, ultimately leading to the inclusion of 30 publications in this review (Table 1).

The temporal analysis of the included studies shows a steady increase in research output between 2019 and 2024, with a notable peak in 2022 (Figure 2). The geographic distribution of publications, highlighting the countries contributing the most to this research field, is presented in Figure 3.

The average compliance with the TRIPOD-AI checklist among studies developing and evaluating prediction models for dental caries was estimated between 80% and 85%. However, several recurrent deficiencies were identified. The most critical was the lack of a definitive gold standard for ground truth, as most studies relied on expert clinical annotation and consensus—often with high inter-rater agreement (Kappa > 0.75)—instead of histological or micro-CT verification. In addition, formal sample size calculations were generally absent, and the generalizability of results was limited by the predominant use of data from single institutions or ex vivo settings. Finally, a recurring shortcoming was the inadequate handling and reporting of class imbalance, particularly regarding the low prevalence of rare or severe lesions, which contributed to reduced model performance in minority classes despite high overall metrics (Table 1).

The studies included in this review assessed a wide spectrum of artificial intelligence applications for dental caries detection across different imaging modalities and diagnostic techniques (Figure 4). The performance of these approaches varied depending on the type of input data, the architecture employed, and the methodological rigor of the studies. Below, the main findings are summarized by technique, highlighting the strengths, limitations, and clinical implications of each approach.

The analysis of intraoral photography revealed heterogeneous performance across architectures, with models such as MobileNetV2 (sensitivity 0.925; specificity 0.896) and VGG-16 (AUC 0.8565; sensitivity 0.819) standing out for their accessibility and clinical potential, though their accuracy was strongly influenced by image quality [18,20,27,32]. More advanced frameworks, including Cascade R-CNN (specificity 0.96; sensitivity 0.732) and Mask R-CNN (accuracy 0.889), demonstrated higher robustness but required considerable computational resources, limiting their applicability [10]. In contrast, UV absorption spectroscopy, applied to salivary samples, achieved perfect sensitivity and specificity, highlighting its potential for noninvasive diagnosis [15].

CBCT yielded high diagnostic values (accuracy 0.953; sensitivity 0.921; specificity 0.963), though its integration into routine practice remains constrained by the high technical and resource requirements of this imaging modality [11]. Dental (periapical) radiography studies reported variable outcomes: while Faster R-CNN demonstrated moderate accuracy (0.7349) with rapid processing times (0.1923 s per image), Deeplabv3 achieved very high average accuracy (0.994) but poor segmentation performance (IoU mean 0.5073), reflecting challenges in lesion delineation [6,16].

Emerging methods, such as TFSNs, supported by U-Net and NASNet, reported moderate diagnostic performance (sensitivity 0.8026; PPV 0.7636), particularly for early-stage caries [8]. In panoramic radiographs, architectures such as ResNet18 (accuracy 82.72%; sensitivity 0.8538; specificity 0.877), MobileNetV2 (accuracy 0.87; specificity 0.88), and DCDNet (accuracy 0.72) showed variable but promising results for large-scale screening, with MobileNetV2 offering practical advantages for resource-limited environments [17,18,19].

NILT emerged as a radiation-free alternative, with ResNet18–ResNeXt50 combinations achieving moderate accuracy (0.69) and AUC (0.74), while U-Net with VGG-16 improved segmentation performance (mIoU 0.727). Nonetheless, the sensitivity remained relatively low (0.59), limiting its reliability despite favorable AUC values in vivo (0.78) compared to in vitro settings (0.65) [20,21,22]. Studies exploring clinical data and salivary biomarkers through neural networks also demonstrated encouraging performance, with accuracy values above 83%, reinforcing their potential role in complementary diagnosis [23].

Bitewing radiographs, the most frequently studied modality (13 publications), confirmed the versatility of deep learning architectures. U-Net models consistently reached high accuracy (0.9491) and F1-scores (0.8818), while AlexNet (accuracy 0.903) and VGG19 (accuracy 0.94) underscored the potential of classical CNNs (24,25,26,27). Advanced models such as Faster R-CNN and VGG-16 reported robust metrics (accuracy 0.861; specificity 0.985; AUC 0.948), and DarkNet-53 achieved strong discriminatory power (AUC 0.9564; specificity 98.18%) despite a lower sensitivity (72.26%) (7,28,29). RetinaNet, YOLOv5, and EfficientDet applied in the HUNT4 Oral Health Study reached a mean Average Precision (mAP) of 0.647, supporting the utility of large datasets for model validation [4]. Finally, Mask R-CNN obtained consistent AUC values (>0.80) for both primary and secondary caries, with data augmentation improving sensitivity and robustness across clinical conditions [12].

4. Discussion

This review highlights a research field in notable expansion, driven by technological advancements and a growing interest in its integration into dental practice. A multidisciplinary effort is evident to identify the most effective AI applications, exploring a variety of imaging techniques and analytical methods. While bitewing radiographs are the most used, there is a trend towards investigating new imaging technologies, seeking advantages in early and non-invasive detection.

The analyzed studies underscore the transformative potential of AI, especially CNNs, in caries diagnosis, enhancing clinical accuracy and efficiency. Deep learning (DL) algorithms such as U-Net, Faster R-CNN, and Mask R-CNN have demonstrated substantial improvements in diagnostic accuracy, with models like U-Net achieving superior sensitivity (0.75) compared to professionals (0.36) in certain contexts. These automated systems can distinguish between primary and secondary caries and identify incipient lesions with greater sensitivity than traditional methods, also optimizing workflow by directing the clinician’s attention to relevant areas of interest.

The premise of this review regarding the ability of AI models, such as the U-Net architecture, to surpass clinical performance is validated by the literature, which consistently reports that AI enhances the accuracy and sensitivity of dentists [30,31]. For example, AI sensitivity (0.75) has been shown to be significantly higher than the average dentist sensitivity (0.36) in the detection of proximal caries on bitewings [30,32]. This improved sensitivity is particularly relevant, as AI assistance has been observed to substantially increase clinicians’ ability to identify initial or moderate lesions (the categories most easily overlooked) with significant improvements in sensitivity for these subgroups [33]. Nonetheless, this high diagnostic precision must be weighed against reported limitations: AI often shows weaker performance for caries detection (lower sensitivity and lower AUC) compared to other dental pathologies, such as residual roots or crowns, in both panoramic radiographs and CBCT [34,35]. For instance, in a multi-diagnostic framework, AI demonstrated very high specificity (0.990) but the lowest sensitivity (0.554) for caries, with an AUC significantly lower than that achieved by highly experienced dentists [34]. In line with the methodological shortcomings identified in this review, the literature confirms that the lack of a strict ground truth (e.g., histology or micro-CT) is the most critical limitation [5,33,35], since most models are trained on expert consensus as the gold standard, which can introduce biases inherent to human judgment [5,35]. Generalizability is further limited by the predominant use of single-source data [34]. Unlike the earlier review by Mohammad-Rahimi et al. [36], which highlighted promising results but limited study quality, our work not only applies TRIPOD-AI criteria to mitigate bias but also provides clinically oriented insights into the applicability of AI tools for caries detection in real-world practice.

Regarding clinical workflow impact, beyond the optimization noted in this manuscript, eye-tracking studies confirm that dentists assisted by AI display more efficient visual behavior, focusing on relevant regions (caries and restorations) in significantly less time [31]. This efficiency translates into substantial time savings: one AI system for panoramic radiographs achieved an average diagnostic time of 1.5 seconds per radiograph, compared with 53.8 seconds for dentists [34]. Finally, although AI increases diagnostic precision, its cost-effectiveness is only realized when early detection leads to non-restorative, minimally invasive management rather than an increase in invasive interventions [30].

Nevertheless, clinical implementation faces important challenges, mainly the generalizability of models and variability in image quality. This underscores the need for extensive, high-quality datasets representative of diverse clinical scenarios. Other obstacles include the transparency and explainability of AI models (the “black box” problem), the requirement for large volumes of data for training, and the imperative need to establish robust ethical and regulatory frameworks to ensure safe and effective use. Interdisciplinary collaboration between AI developers and dental professionals is indispensable to ensure that the tools meet clinical requirements and are of practical utility.

This review confirms that AI (particularly CNN-based DL architectures) has considerably improved sensitivity, specificity, and reduced inter-observer variability, addressing a persistent challenge in conventional dentistry. However, despite the promising performance reported in controlled studies, validation in real-world clinical settings remains imperative. Future research should focus on the development of more robust, clinically applicable, transparent, and reproducible models, trained on large and diverse datasets. Moreover, improving the interpretability of AI to foster clinician trust, together with the establishment of ethical and regulatory frameworks that safeguard safety and privacy, will be essential for translating these advances into sustainable clinical practice.

5. Conclusions

The AI frameworks, particularly those leveraging DL and CNN architectures, exhibit robust diagnostic efficacy in the detection of dental caries. This efficacy is demonstrated across a spectrum of imaging modalities, with bitewing radiography being the most extensively investigated. The aggregated findings substantiate that these computational models consistently yield performance metrics (notably accuracy, sensitivity, and specificity) that are commensurate with, and in several instances surpass, the diagnostic benchmarks of conventional clinical assessment. Consequently, this work systematically fulfills its primary objectives by not only charting the predominant algorithmic architectures and data preprocessing pipelines but also by critically evaluating their reported efficacy. Finally, this review positions AI not merely as a technological novelty but as a validated and increasingly integral adjunct poised to redefine the paradigms of diagnostic accuracy in contemporary dentistry.