Accuracy of deep learning in the detection of carotid calcifications on cone-beam computed tomography: A systematic review

Abstract

Purpose

This systematic review aimed to describe the diagnostic performance of AI-based models in identifying carotid calcifications using cone-beam computed tomography (CBCT) images.

Materials and Methods

A comprehensive search was conducted in the PubMed/MEDLINE, Embase, IEEE Xplore, SciELO, and LILACS databases to identify relevant studies published up to 2025 that evaluated the diagnostic accuracy of artificial intelligence or deep learning systems in detecting carotid artery calcifications using CBCT. Grey literature sources were systematically searched. A qualitative synthesis was conducted for the included studies, followed by a diagnostic accuracy meta-analysis using sensitivity and specificity data. Analyses were performed using bivariate random-effects models (Diagnostic Random-Effects Model). Heterogeneity among studies was assessed using Cochran’s Q test, the I² index, Tau² statistic, and the corresponding P-value.

Results

A total of 529 records were identified. No additional studies were retrieved from the grey literature. Application of inclusion and exclusion criteria resulted in the selection of four studies for qualitative synthesis.

Conclusion

Despite variations in convolutional neural network (CNN) model architectures, all studies demonstrated that deep learning algorithms applied to CBCT achieve high performance levels in detecting carotid calcifications. The meta-analysis demonstrated that CNN-based models have high diagnostic potential for detecting carotid artery calcifications on CBCT. Although CBCT does not replace gold-standard diagnostic modalities, its use may represent a supportive tool for early screening and clinical referral. The expansion of datasets and the standardization of image acquisition protocols and quality are recommended.

Introduction

Mural carotid calcifications of major arteries have been consistently reported in the scientific literature as readily detectable through various imaging methods.1,2,3,4,5 These deposits, composed of lipids, inflammatory cells, connective tissue, and calcium-phosphate complexes, are commonly found at the bifurcation of the common carotid artery and are considered a hallmark of arterial aging and might be a strong indicator of atherosclerotic vascular disease.6,7,8

Although these plaques are known to be a significant risk factor for ischemic cerebrovascular events, they are usually asymptomatic in their early stages. As a result, they are frequently detected incidentally through imaging examinations performed for unrelated reasons, such as head and neck computed tomography (CT) or cone-beam computed tomography (CBCT) scans.3,5 Approximately 20% to 30% of ischemic strokes are thought to be caused by carotid artery stenosis.9,10 The risk of stroke can be considerably decreased by prompt intervention, such as medication, lifestyle changes, or surgical procedures like carotid endarterectomy, which is made possible by early detection of carotid calcifications.11

Even though duplex ultrasound and CT angiography are established screening tools, their routine use may not be feasible in asymptomatic individuals.3,12 Incidental findings of carotid artery calcifications on dental imaging modalities—such as panoramic radiographs and, more recently, CBCT have demonstrated potential as early indicators that warrant further vascular evaluation.13,14

However, the identification of such findings still relies on the radiologist’s expertise, and diagnostic accuracy may vary. Recent studies have shown that deep learning can enhance image interpretation across various imaging modalities. For instance, convolutional neural networks applied to panoramic radiographs have achieved high accuracy in detecting carotid calcifications.15 In CT angiography, deep learning models have demonstrated over 90% precision in identifying atherosclerotic plaques, compared to 78% for radiologists.16 In carotid ultrasound, artificial intelligence (AI)-based systems have shown high sensitivity (0.94), specificity (0.71), and AUC (0.91) for detecting vulnerable plaques and estimating stenosis severity, even when used by less experienced operators.17,18

Deep learning approaches, especially convolutional neural networks (CNNs), have shown exceptional performance in interpreting complex medical imaging, often being able to surpass human diagnostic accuracy. These systems are capable of extracting rich, objective data from images, leading to notable improvements in disease diagnosis even in early stages. Moreover, decision-support models contribute to more efficient and standardized image analysis, providing reliable information that can support the formulation of therapeutic strategies.19,20

Despite studies have explored the use of AI and deep learning for the detection of carotid calcifications on CBCT scans, this field remains in its early stages. Systematic evidence regarding the diagnostic accuracy of these approaches is still limited. Therefore, this review aims to describe the diagnostic performance of AI-based models for identifying carotid calcifications using CBCT imaging.

Materials and Methods

This systematic review was conducted in strict accordance with the PRISMA guidelines (Preferred Reporting Items for Systematic Reviews and Meta-Analyses).21

This systematic review was guided by the following clinical question: “What is the diagnostic accuracy of AI systems for detecting carotid calcifications on CBCT images?” The PIRD framework (Population, Index test, Reference test, and Diagnosis of interest) was used to select studies on diagnostic accuracy, as follows: Population (P): patients whose CBCT examinations revealed the presence of carotid artery calcification, Index Test (I): the use of AI algorithms and deep learning models, Reference Test (R) radiology report by a board-certified Oral and Maxillofacial Radiologist, and Diagnosis of interest (D): diagnostic accuracy outcomes of AI models, including sensitivity, specificity, log diagnostic odds ratio, and the area under the receiver operating characteristic (ROC) curve (AUC), and predictive values. This structured approach aimed to ensure clarity in eligibility criteria, guide the search strategy, and facilitate objective data synthesis.

The protocol for this systematic review was registered in the PROSPERO database (registration number: CRD420251036846) and is accessible at http://www.crd.york.ac.uk/PROSPERO/.

Search strategy and eligibility criteria

A comprehensive systematic review was conducted to identify relevant studies evaluating the diagnostic accuracy of AI or deep learning systems in detecting carotid calcifications using CBCT. Searches were performed in the following electronic databases; PubMed/MEDLINE, Embase, IEEE Xplore, Scientific Electronic Library Online (SciELO), and Latin American and Caribbean Health Sciences Literature (LILACS).

Additionally, grey literature sources were systematically searched to identify supplementary studies eligible for inclusion in the systematic review. These sources comprised Google Scholar, academic theses and dissertations, governmental and institutional reports, clinical practice guidelines, conference proceedings, patents, and technical documentation.

The search strategies were developed using Boolean operators (AND/OR/NOT) and were adapted for each database: MEDLINE/PubMed, Embase, LILACS, IEEE Xplore, and Google Scholar, to identify studies meeting the inclusion criteria, using the terms detailed in Table 1. The searches included studies published up to May 2025, with no restrictions on language or year of publication.

Table 1. Search strategies used in the databases.

To increase specificity, terms related to CT angiography and coronary vessels were excluded. In addition, the initial screening was based on available titles and abstracts, and potentially relevant records were selected for further evaluation in subsequent stages (full abstract and full-text review, when available).

To ensure consistency and accuracy, references were managed using EndNote Web software, which was also used to identify and remove duplicate records.

Studies were considered eligible for inclusion if they met the following criteria: original research articles, including prospective, retrospective, or cross-sectional studies, that evaluated the diagnostic accuracy of AI, machine learning, or deep learning algorithms for detecting carotid calcifications using CBCT in human subjects. Eligible studies were required to compare AI-based evaluations with manual interpretations by radiologists and/or a reference diagnostic standard. To be included, studies had to report at least one diagnostic accuracy measure-such as sensitivity, specificity, AUC, or predictive values. There were no restrictions regarding publication date or language, provided that the full text of the study was available.

Exclusion criteria comprised studies that utilized imaging modalities other than CBCT (CT, CT angiography, or magnetic resonance imaging), studies not involving AI or computational algorithms, animal or in vitro studies, as well as reviews, editorials, letters to the editor, commentaries, technical notes, or conference abstracts lacking full data. Duplicates or secondary analyses of previously published datasets were also excluded.

For inclusion in the meta-analysis, studies were required to report diagnostic accuracy outcomes at the examination or patient level, allowing the construction of 2×2 contingency tables (true positives, false positives, false negatives, and true negatives).

Studies that met the inclusion criteria but reported performance metrics derived from image segmentation frameworks or predictive modeling approaches were included exclusively in the qualitative synthesis.

Process of data collection and data extraction

Data extraction and review were independently conducted by 2 reviewers (GTTA and KMC). Any disagreements regarding the inclusion or exclusion of studies were resolved through re-evaluation by a third reviewer (VAS).

The study selection process was carried out in three distinct stages. In the first stage, after removing duplicates using EndNote Web, the reviewers examined the article titles to identify potentially eligible studies. In the second stage, the article abstracts were assessed. Finally, in the third stage, the full texts of the selected articles were thoroughly reviewed based on the inclusion and exclusion criteria. Discrepancies at both stages were resolved through consensus meetings, including the third reviewer (VAS).

The data extracted from each included study were systematically reviewed and organized according to predefined characteristics. These included: the author and year of publication, country of origin, study design, deep learning methodology employed, the comparative method used, sample size, and primary and secondary outcomes such as diagnostic accuracy metrics and additional technical or clinical findings. Data collection was performed using a standardized extraction form and was based solely on information explicitly reported in the articles. A detailed summary of all included studies is presented in Table 2.

Table 2. Methodological characteristics of studies evaluating artificial intelligence (AI) models for carotid calcification detection in cone-beam computed tomography (CBCT) imaging.

OMFR: oral and Maxillofacial Radiologist, FOV: field of view, CNN: convolutional neural network, ACC: accuracy, AUC: area under the curve, bbox_acc: bounding box accuracy, DSC: Dice similarity coefficient, ECAC: extra-cranial artery calcification, F1: harmonic mean of precision and recall, ICC: internal carotid calcification, IoU: intersection over union, HD95: 95th percentile Hausdorff Distance, MCC: Matthews correlation coefficient, NN: neural network, NR: not reported, OMFP: oral and maxillofacial pathologist, NPV: negative predictive value, PPV: positive predictive value, SE: sensitivity, SP: specificity

Synthesis methods

A qualitative synthesis was conducted for all four included studies to summarize their methodological characteristics and main findings. For the quantitative synthesis, a diagnostic accuracy meta-analysis was performed using sensitivity and specificity data extracted from two primary studies.22,23 Analyses were conducted using bivariate random-effects models (diagnostic random-effects model) to calculate pooled estimates with their respective 95% confidence intervals. Heterogeneity across studies was assessed using Cochran’s Q test, the I² index, Tau² statistic, and corresponding P-value. Results were graphically represented through forest plots. All analyses were performed using the OpenMeta [Analyst] software.

Analysis of the risk of bias of included studies

The risk of bias of the included studies were independently assessed by two reviewers (GTTA and KMC) using the QUADAS-2 tool (Quality Assessment of Diagnostic Accuracy Studies-2),24 which is validated for evaluating diagnostic accuracy studies. This tool assesses four key domains; 1) patient selection, 2) index test, 3) reference standard, and 4) flow and timing.

Each domain was evaluated for risk of bias and concerns regarding applicability to the review question. The index test referred to the AI/deep learning model applied for carotid calcification detection in CBCT images, while the reference standard varied across studies, including expert radiologist annotation, manual segmentation, or clinical confirmation. Each domain was classified as “low risk,” “high risk,” or “unclear,” resulting in an overall judgment for each study. Discrepancies were resolved by consensus or by involving a third reviewer (VAS). A summary of this assessment is presented in Table 3.

Table 3. QUADAS-2 evaluation summary.

: low risk, : high risk, ?: unclear risk.

Results

Study selection

A total of 530 records were identified across the selected databases: PubMed/MEDLINE (488), Embase (25), LILACS (1), and IEEE Xplore (16). No additional studies were retrieved from the grey literature. After removal of 8 duplicates, 522 records remained for screening. Titles and abstracts were reviewed according to the predefined inclusion and exclusion criteria. As a result, 28 full-text articles were sought for retrieval, of which 18 were not retrieved, and 10 reports were assessed for eligibility. Following full-text evaluation, 6 articles were excluded: five for using panoramic radiography instead of CBCT, and one for not evaluating AI. At the conclusion of the screening process, four studies met all eligibility criteria and were included in the qualitative synthesis. The study selection process is illustrated in the PRISMA flow diagram (Fig. 1).

Fig. 1. PRISMA flow diagram illustrating the selection process of studies included in the systematic review.

Study characteristics

Among the four studies included in this review, all employed experimental designs published recently, between 2022 and 2025 and from the United States. Ajami et al.,22 Nelson et al.,23 Alajaji et al.25 conducted experimental studies using retrospective CBCT exams evaluating the performance of CNNs in detecting carotid artery calcifications in CBCT scans. Lastly, Mahdian et al.26 performed a retrospective validation study using a U-Net segmentation model combined with XGBoost for classification, comparing model predictions with radiological features and clinical cardiovascular outcomes. None of the studies followed a prospective or randomized protocol.

The included studies collectively analyzed CBCT datasets from a total of 734 exams. The CBCT acquisition protocols implemented across the studies demonstrated heterogeneity in technical parameters. In the study by Ajami et al.22 an I-CAT digital imaging device (Imaging Sciences International, Hatfield, USA) was used. Alajaji et al.25 did not report the equipment used, as well as the acquisition parameters, solely the field of view (FOV) of 13.5 × 17.0 cm was indicated.

Results of individual studies

Nelson et al.23 acquired CBCT images using a 3D Accuitomo (J. Morita Corp, Kyoto, Japan) scanner. Mahdian et al.26 utilized 3 different devices such as I-CAT (Imaging Sciences International, Hatfield, USA), Instrumentarium OP300 (Kavo Dental, Charlotte, USA), and Carestream 9600 (Carestream Dental LLC; Atlanta, GA, USA). Due to equipment variability, acquisition parameters were adjusted for an average-sized patient, with kVp ranging from 70 to 120, mA from 3.2 to 10, voxel sizes from 0.25 to 0.4 mm, and a scan time of 19.0 seconds. The FOV ranged from the region of the external (cervical) carotid artery to the posterior vertebral artery. Detailed information about the CBCT protocols used in the included studies is summarized in Table 2.

Varieties of AI architectures were employed across the included studies. Ajami et al.22 used a Inception V3 + U-Net, while Alajaji et al.25 applied a U-Net based neural network architecture. Nelson et al.23 adopted a retrospective design, training a custom CNN architecture based on TensorFlow, to classify the presence or absence of calcifications based on CBCT images, with manual annotation by expert radiologist as the reference standard, and Mahdian et al.26 used nnU-Net.

The diagnostic performance metrics varied considerably across the included studies. Ajami et al.22 reported the highest AUC (0.97) and accuracy (94%) using a CNN architecture. In Alajaji et al.25 study, for extracranial calcified carotid artery atheroma (CCAA) segmentation, the model demonstrated robust performance, with a mean training accuracy of 92% and a validation accuracy of 82%. The testing accuracy was also 92%, with an intersection over union (IoU) of 0.90, a Dice Similarity Coefficient (DSC)/F1-score of 0.95, a positive predictive value (PPV) of 0.90, and a negative predictive value (NPV) of 1.0. Sensitivity was 1.0, specificity was 0.69, and the area under the ROC curve (AUC) reached 0.84. In contrast, the performance for intracranial CCAA segmentation was considerably lower, with a training accuracy of 61% and a validation accuracy of 70%. The testing accuracy was 38%, with an IoU of 0.35, DSC/F1-score of 0.52, PPV of 0.36, NPV of 0.67, sensitivity of 0.93, specificity of only 0.08, and an AUC of 0.50.

Nelson et al.23 reported a k-fold cross-validation accuracy of 76%, with a recall of 66% and a precision of 79%. The combined F1-score was 0.72, indicating a good balance between sensitivity and precision. Additionally, the Matthews Correlation Coefficient (MCC) was extrapolated to be 0.53, suggesting a strong balance across the confusion matrix categories and overall model robustness.

Mahdian et al.26 reported bounding box accuracy values of 0.71 for ECC, 0.78 for ICC, and 0.53 for VAC. The accuracy of cardiovascular disease classification varied depending on the type of calcifications analyzed. For stroke classification based on ECC, three models were used: clinical-only, radiomics-only, and a combined model (clinical + radiomics). Among them, the combined model showed superior performance, with an AUC-ROC of 0.69 ± 0.04, sensitivity of 0.87 ± 0.02, and specificity of 0.55 ± 0.12. For ICC, the combined model also outperformed the others, achieving an AUC-ROC of 0.94 ± 0.09, sensitivity of 0.92 ± 0.11, and specificity of 0.85 ± 0.21. Regarding VAC, the combined model achieved an AUC-ROC of 0.73 ± 0.22, sensitivity of 0.72 ± 0.12, and specificity of 0.68 ± 0.39.

For myocardial infarction classification, the combined model again demonstrated superior results. Based on ECC, the model achieved an AUC-ROC of 0.88 ± 0.04, sensitivity of 0.83 ± 0.05, and specificity of 0.52 ± 0.18. For ICC, the model reached an AUC-ROC of 0.84 ± 0.16, sensitivity of 0.88 ± 0.06, and specificity of 0.73 ± 0.22. The VAC-based model achieved an AUC-ROC of 0.77 ± 0.33, sensitivity of 0.83 ± 0.24, and specificity of 0.70 ± 0.42.

The reference standards among all the included studies was based on radiographic assessment by a board-certified oral and maxillofacial radiologist.

Additional information is presented in Table 1, including key methodological and performance features of each study, such as study design, anatomical region assessed, AI model used, task type, reference standard, CBCT settings, sample size, ground truth definition, use of augmentation, diagnostic metrics, main findings, and the specific validation method applied.

Results of quantitative synthesis

The diagnostic accuracy meta-analysis included 2 primary studies,22,23 which evaluated the performance of CNNs for detecting carotid artery calcifications in CBCT examinations.

The pooled sensitivity was estimated at 0.851 (95% CI: 0.424–0.978; P = 0.095), with substantial heterogeneity across studies (Tau² = 2.080; Q value = 20.627; I² = 95.152%, P < 0.001). Although the mean value indicates that the test correctly identified approximately 85% of true positive cases, the wide confidence interval reflects high statistical uncertainty and a possible influence of methodological variability among studies.

For specificity, the pooled estimate was 0.928 (95% CI: 0.733–0.984; P = 0.001), also with significant heterogeneity (Tau² = 1.173; Q value = 14.314; I² = 93.014%, P < 0.001), demonstrating high performance and statistical significance in identifying true negative cases (92.8%). The relatively narrow confidence interval and the strong P-value further support the robustness of the test’s specificity. These findings suggest that CNN-based models exhibit high diagnostic potential, with elevated specificity and moderate-to-high (Fig. 2)

Fig. 2. Forest plots representing the individual and pooled estimates of sensitivity and specificity of the studies included in the meta-analysis. A. Forest plot representing the pooled and individual estimates of sensitivity of the included studies. B. Forest plot representing the pooled and individual estimates of specificity of the included studies.

Analysis of the risk of bias

Based on the QUADAS-2 criteria, the study by Ajami et al.22 was assessed as having a low risk of bias in patient selection, given the use of well-defined and consistently applied inclusion and exclusion criteria. However, the reference standard raised some concerns. Although the evaluation was conducted by an expert radiologist, the study did not provide sufficient detail on how carotid calcifications were identified or verified, making the reliability of the diagnostic process unclear.

Also, the index test domain, referring to the application of AI to CBCT for the detection of carotid calcifications, was rated as having unclear risk of bias and applicability concerns. This was due to the lack of information regarding whether assessors were blinded to the reference standard, as well as the absence of detailed validation procedures for the deep learning model.

The flow and timing domain was rated as low risk, since all cases were assessed in a consistent manner, with no reported exclusions or delays between the index test and the reference standard.

The studies conducted by Alajaji et al.,25 Nelson et al.23 and Mahdian et al.26 presented similar results, all of them assessed as having a low risk of bias in patient selection, due to its retrospective collected exams used and clearly described criteria. However, in agreement with the study of Ajami et al.,22 the references standard domains, the study were judged to have an unclear risk of bias, also due to the lack of detailed information on how the validation for carotid calcification detection was established. The study did not clarify whether calcification findings were confirmed through gold-standard imaging modality, such as Doppler ultrasound or CT angiography. Additionally, there was no mention of blinding in the interpretation of the reference standard, raising concerns about potential incorporation bias. The results of these assessments are summarized in the Table 3.

Discussion

This systematic review evaluated the diagnostic accuracy of AI algorithms for the detection of carotid artery calcifications (CACs) in cone beam computed tomography scans. The restriction to three-dimensional examinations was based on the greater reliability of volumetric imaging for assessing cervical calcifications, as multiplanar visualization reduces anatomical superimposition and the potential confusion with non-vascular calcified structures. Thus, CBCT-based analysis enhances diagnostic precision and improves the clinical applicability of AI models. This approach differs from the study by Arzani et al.,27 which included both two-dimensional and three-dimensional examinations and also reported high diagnostic performance using AI.

Mahdian et al.26 highlighted that accurate segmentation of carotid calcifications remains challenging due to limitations inherent to small object detection, particularly the reduced size of lesions and subtle contrast differences. This issue is especially relevant in dental imaging, as extracranial carotid artery calcifications are commonly located near the carotid bifurcation at the C3–C4 level, which is an anatomical region frequently included in CBCT and panoramic radiographs — thereby enabling potential early identification of calcifications associated with cardiovascular disease despite segmentation complexity.

Image quality plays a critical role in the diagnostic performance of deep learning models, particularly in segmentation tasks. As highlighted by Albitar et al.28 and Fu et al.,29 suboptimal resolution and image quality compromise the efficiency of feature extraction techniques required for the identification of vascular calcifications. Among the included studies, only Alajaji et al.25 did not report voxel size, whereas the others used values ranging from 0.25 mm to 0.4 mm, which directly influence spatial resolution and the visualization of small calcifications. Therefore, methodological differences related to voxel size may contribute to the observed heterogeneity in diagnostic accuracy.

Alajaji et al.25 frequently misidentified structures such as the anterior clinoid process and the sphenoid bone as intracranial CCAAs, likely due to their anatomical proximity and similar radiodensity. Detection on axial slices proved particularly challenging, resulting in a high rate of false positives, especially within the ethmoid and sphenoid sinuses. Consequently, specificity on axial images was very low (0.08) compared with the extracranial dataset (0.69). The authors suggest that incorporating multiplanar reconstructions may improve diagnostic accuracy.

These findings emphasize the importance of precisely directing AI models toward the carotid region in three-dimensional imaging, such as CBCT, due to the frequent presence of anatomically adjacent structures with overlapping radiographic characteristics. For example, in the study by Ajami et al.,22 cases involving calcified triticeous cartilage and the superior cornu of the thyroid cartilage both of which may closely mimic vascular calcifications, were explicitly excluded from the training set to reduce the risk of misclassification. Such strategies are essential to improving the specificity and clinical reliability of AI-based diagnostic tools.

Ground truth definition and dataset composition varied notably among the included studies and may have contributed to the differences in model performance. In the study by Alajaji et al.,25 ground truth was generated through meticulous, color-coded manual annotation of carotid calcifications and multiple surrounding anatomical structures. Additionally, incidental findings such as sialoliths were annotated to reduce the risk of misclassification, although calcified triticeous cartilage, despite its acknowledged radiographic similarity to CCAA, was not present in their dataset.

In contrast, Mahdian et al.26 included scans with various types of calcifications, which may have introduced greater diagnostic complexity and partially explain the variation in performance metrics. Nelson et al.23 and Ajami et al.22 used binary labels indicating the presence or absence of calcifications to each CBCT scan based on the expert radiologist’s interpretation.

Moreover, calcifications are often located near the edges of the image volume, where quality tends to be lower due to noise and artifacts, making their detection more challenging, particularly for models trained on limited or heterogeneous datasets. Their small size requires the capture of fine details that are often imperceptible to the human eye, which is limited in differentiating subtle grayscale variations in CBCT data.30 This reinforces the need for AI models capable of identifying patterns beyond clinical visual perception.

An important aspect observed in the reviewed studies is the discrepancy between the establishment of the ground truth, often defined by specialists through comprehensive three-dimensional evaluation of CBCT scans, and the training of CNNs, which in most studies use only two-dimensional images.

The diagnosis of carotid calcifications is generally performed using gold-standard imaging modalities such as CT angiography,31,32 Doppler ultrasound, or MR angiography, particularly when interpreted alongside clinical findings. Nevertheless, the identification of these calcifications in dentomaxillofacial imaging remains clinically valuable. Although CBCT does not replace these methods for vascular assessment, its ability to reveal calcifications as incidental findings in dental examinations represents an important opportunity for early suspicion and referral. This is particularly relevant in the dental context, where oral and maxillofacial radiologists do not always systematically interpret images, and dentists often focus on dentoalveolar structures, potentially overlooking clinically significant vascular findings.

Most evidence comes from retrospective analyses or single-center cohorts, which may not reflect the complexity of real-world clinical settings. Although preliminary results are promising, the true test of these tools will depend on large, multicenter prospective studies with standardized protocols and external validation, ensuring consistent performance and practical clinical applicability.

In agreement with a limitation highlighted by Nelson et al.,23 the model used identified only the presence or absence of calcifications, without considering their severity or morphological patterns. All positive cases were treated equivalently, regardless of the number, size, or location of the calcifications. Although useful for initial detection, this approach limits diagnostic specificity and clinical utility. Therefore, we emphasize the importance of developing models capable of stratifying the severity of calcifications.

When suspected calcifications are identified and confirmed by the radiologist, standardized referral pathways to primary care physicians, cardiologists, or vascular specialists must be activated to ensure appropriate medical evaluation using confirmatory imaging and cardiovascular risk assessment. Such an approach reinforces the role of dental imaging as a complementary component of interdisciplinary healthcare, promoting early identification of individuals at potential cardiovascular risk while respecting diagnostic boundaries and maintaining patient safety.

Although four studies met the eligibility criteria and were included in the qualitative synthesis, only two were eligible for the meta-analysis due to limitations in the availability and compatibility of diagnostic accuracy data. Alajaji et al.25 and Mahdian et al.26 were excluded because their performance metrics were based on image segmentation approaches and predictive modeling rather than conventional diagnostic accuracy analyses. Specifically, they reported results at the pixel or slice level, without providing exam-level data in 2×2 contingency tables required for calculating true and false classification measures necessary for meta-analytic pooling.

A key limitation was the heterogeneity in the objectives of the included studies, which involved different approaches such as detection, classification, and segmentation of calcifications. This methodological variability, combined with the lack of standardization in outcome reporting, may have contributed to the observed heterogeneity and limited the comparability and interpretability of the pooled results.

Additionally, all included studies presented small datasets, and deep learning models depend on both the volume and diversity of data to learn meaningful patterns. Limited datasets compromise the ability to generalize to new cases. Furthermore, the development of robust models requires high-quality training data, as the performance of the algorithms depends directly on the reliability of expert annotations.

Finally, limitations related to the model architecture itself should also be considered. These limitations reduce the model’s ability to learn the full anatomical context, as illustrated in the study by Alajaji et al.,25 in which intracranial findings were frequently misclassified due to radiographic similarity with adjacent structures that are difficult to differentiate in a single plane. This example highlights that restricting the model to one or two dimensions may reduce diagnostic accuracy and reinforces the need to develop truly three-dimensional models capable of replicating clinical expertise and detecting subtle changes that remain imperceptible to the human eye.

Beyond methodological considerations, other limitations related to real-world clinical implementation should also be recognized. Regulatory approval processes may represent an additional layer of complexity for incorporating AI into daily clinical practice. This technology must undergo rigorous validation to ensure clinical safety and reliability before being integrated into routine use.

Although AI can be integrated into dental radiology workflows as an opportunistic screening tool and assist in identifying regions suggestive of calcifications, it does not replace professional judgment; the dentomaxillofacial radiologist remains responsible for the accurate description of findings. In addition to accuracy, cost-effectiveness and efficient integration into existing diagnostic workflows must also be considered, as the potential benefits—such as early detection of systemic vascular risks—still require systematic evaluation and economic analyses.

AI models have demonstrated potential as auxiliary tools for detecting carotid artery calcifications on CBCT scans. However, the findings of this systematic review should be interpreted as preliminary and exploratory, considering the limited number of studies providing compatible diagnostic accuracy data, the methodological heterogeneity observed, and the variability in reference standards across investigations.

Although different convolutional neural network architectures were evaluated, the available evidence collectively suggests that deep learning approaches applied to CBCT imaging may support the identification of carotid calcifications. Nevertheless, current results remain insufficient to establish consistent diagnostic performance or broad clinical applicability. Rather than replacing gold-standard vascular imaging, CBCT-based AI should be viewed as a supportive tool for opportunistic detection and referral. Definitive diagnosis and cardiovascular risk assessment remain the responsibility of medical evaluation following appropriate patient referral.

Future research based on standardized diagnostic accuracy methodologies, larger multicenter datasets, and fully validated three-dimensional analytical frameworks is required to strengthen evidence reliability and clarify the potential role of these models in cardiovascular risk assessment.