Reliability of artificial intelligence algorithms in automated age estimation using orthopantomograms: A scoping review

TL;DR

This study reviews AI algorithms for estimating age from dental X-rays, finding they are more accurate than traditional methods and could aid forensic and clinical work.

Contribution

The study systematically evaluates AI models for age estimation from orthopantomograms, highlighting their reliability and potential clinical impact.

Findings

AI models like CNNs and EfficientNet show higher accuracy than traditional age estimation methods.

AI-driven age estimation reduces human error and enhances decision-making in forensic and clinical settings.

Standardized protocols are needed to integrate AI models into forensic odontology practices.

Abstract

This study aims to evaluate the efficiency of AI (artificial intelligence) algorithms for automated age estimation using orthopantomograms (OPGs) and to determine whether these models can effectively replace conventional age estimation techniques.

Three independent literature searches were conducted in PubMed, Scopus, and Embase. Studies published in the English language were considered, focusing on age estimation using AI. A total of 1519 articles were screened, and 24 articles were included in the study. The data was extracted in a standardized, predefined manner. After finalizing the search, the data collected was tabulated, interpreted, and verified. The selected studies were analyzed for methodological rigor, algorithmic performance, and comparative effectiveness against traditional age estimation methods.

AI-based models, especially deep learning architectures like convolutional neural networks, EfficientNet, DenseNet, and hybrid models such as Age-Net, demonstrated superior accuracy, precision, and reliability compared to traditional age estimation methods. These AI-driven models show promising results in reducing human error, increasing efficiency, and enhancing forensic and clinical decision-making.

AI-driven age estimation using OPGs represents a transformative advancement with considerable forensic and clinical potential. Although these AI models may not yet fully replace conventional techniques, they offer a substantial value as complementary tools, improving both accuracy and operational efficiency. To foster wider adoption and improve reliability, ongoing research and the development of standardized protocols are essential for integrating these methods into forensic odontology and related fields.

Figures

Total number of articles screened in the databases.

Full text

Introduction

The advent of artificial intelligence (AI) has revolutionized various fields, including healthcare and dentistry. One of the promising applications of AI is in the realm of age estimation, particularly through the analysis of orthopantomograms (OPGs). Accurate age estimation is crucial in various contexts, such as forensic science, anthropology, and pediatric dentistry, where chronological age may influence treatment decisions and legal proceedings.^ 1 ^

In India, the growing population and increasing reliance on digital health technologies necessitate the development of reliable AI algorithms for age estimation. Traditional methods of age determination, which often rely on manual assessment by dental professionals, can be time-consuming and subjective. While AI has been extensively adopted in various medical fields, its integration into dentistry, particularly in forensic odontology, is still in its preliminary stages. The integration of AI into this area promises enhanced accuracy and efficiency, potentially reducing human error and improving diagnostic outcomes.^2,3^

AI is rapidly transforming the field of dentistry by automating diagnostic, analytical, and decision-making capability.^4,5^ Digital dentistry (digital imaging, computer-aided design and computer-aided manufacturing) is already reshaping clinical workflows such as auto-designing dental restorations, optimizing implant placement, and analyzing datasets to predict outcomes, personalize treatment, and reduce the risk.^ 6 ^ In forensic odontology, AI has significant potential to enhance human identification and age estimation, which are crucial aspects of forensic practice. AI technologies, particularly deep learning (DL) and convolutional neural networks (CNNs), can analyze panoramic radiographs, intraoral scans, and cone beam computed tomography images to predict dental maturity and chronological age. Automated systems have demonstrated accuracy in staging third molar development, which is an important tool for determining legal age thresholds in forensic cases. Additionally, AI can quantify pulp–tooth ratios (PTRs) and extract radiomic features, thus improving the accuracy of adult age estimation. Beyond age estimation, AI can also classify dental structures for comparative identification, perform bite-mark analysis, and recognize unique dental patterns.^4,5^ DL approaches can also be used to automate and enhance the detection of dental issues in OPGs, such as broken roots, periodontally compromised teeth, and the Kennedy classification of partially edentulous arches.^ 7 ^ Despite the potential benefits, the reliability of AI algorithms in automated age estimation remains underexplored. Specialized postgraduate programs are crucial to prepare the next generation of dentists who are equally skilled in traditional clinical dentistry and digital/AI-driven dentistry.^ 5 ^ A thorough review is warranted to evaluate existing literature, identify gaps in the current understanding, and assess the applicability of AI-driven techniques.

The main research question of this scoping review was to assess the latest developments in using AI for age estimation from OPGs, and to determine if these models demonstrate an advantage over conventional techniques by achieving lower mean absolute error (MAE) and improved diagnostic performance. Thus, the main objective of this paper was to determine the efficiency of AI-driven algorithms derived by the automated age estimation process using OPGs and to validate the potential of automated age estimation, utilizing OPGs, and to assess whether AI-based models can effectively replace conventional techniques of age estimation using OPGs. This scoping review offers a unique synthesis of 24 original studies focused on AI-based age estimation using OPGs, building on existing reviews of AI applications in forensic and dentistry. It critically compares various AI models with traditional techniques. The findings indicate that AI consistently achieves lower MAE, greater accuracy, and less subjectivity compared to manual assessments. The paper also highlights the forensic applications of AI in determining legal age thresholds (14, 16, and 18 years), disaster victim identification (DVI), and human identification tasks. The study evaluates the limitations of algorithms, focusing on issues such as data scarcity, imbalanced datasets, and the lack of generalizability across diverse populations. It also emphasizes the need for standardized protocols and validation frameworks.

Methods

Search strategy

Our study followed the preferred reporting items for systematic reviews and meta-analyses extension for scoping reviews framework, a thorough digital search for data using reliable databases, including PubMed, Scopus, and Embase was done.^ 8 ^ To find relevant articles in these electronic databases, we employed a variety of key terms: “Artificial Intelligence,” “Machine learning,” “Deep Learning,” “Deep neural networks,” “Convolutional neural networks,” “Age estimation,” “Forensic dentistry,” and “Forensic odontology.” We also utilized Boolean operators (AND, OR) and applied filters for articles written in English. Table 1 outlines the detailed search strategy used for this scoping review.

Table 1.: Summary of search terms, medical subject headings terms, and Boolean operators/filters applied.

Eligibility criteria

Inclusion criteria

1. Study design: Original research articles (observational or retrospective) evaluating age estimation.
2. Population: OPGs were used as the primary radiographic data source.
3. Intervention: Use of AI, machine learning (ML), or DL models for age estimation.
4. Comparisons: Studies comparing AI-based models with traditional techniques and solely AI models were included.
5. Outcomes: Measured in terms of at least one performance metric, including accuracy, sensitivity, specificity, precision, recall, MAE, root mean-squared error, coefficient of determination (R^2^), or standard error of estimate.
6. Publication type and language: Full-text, peer-reviewed articles published in English.

Exclusion criteria

Nonoriginal works (reviews, systematic reviews, meta-analyses, commentaries, technical notes, letters to the editor).Case reports, pilot studies, and conference abstracts.Articles not available as full text or published in languages other than English.Studies not directly involving AI/ML/DL algorithms for age estimation.Research without human OPG data (e.g. animal studies or other imaging modalities).Studies with insufficient methodological details or without clearly reported performance outcomes.

Data charting process

The initial screening of titles and abstracts, along with the removal of duplicates in an Excel sheet (version 1.19.4), was conducted based on the established inclusion and exclusion criteria. Data charting was performed independently by two reviewers (S.K. and S.P.), with subsequent verification by the senior reviewer (V.P.). Any discrepancies were resolved by a third reviewer (A.J.) to ensure the accuracy and reliability of the extracted information. The Excel sheet recorded details such as the authors, study population, year of publication, and country of study. The intervention focused on the application of AI models in age estimation, along with various conventional techniques used, as well as the outcomes and limitations of the studies were evaluated.

Results

Selection of the source of evidence

A total of 1519 records were retrieved across PubMed (n = 1433), Scopus (n = 61), and Embase (n = 25). After removing 12 duplicates, 1507 unique records were screened. Of these, 1466 articles were excluded at the title/abstract stage (1264 did not align with population intervention comparison outcome, 175 review papers, 5 pilot studies, 20 case reports, and 2 commentaries/technical notes). Forty-one full-text articles were assessed for eligibility, and 17 were excluded (1 due to unavailability of full text, 15 not meeting strict inclusion criteria, and 1 not in English). Ultimately, 24 studies met the inclusion criteria and were included in this scoping review. Figure 1 illustrates the search strategy employed in this scoping review.

Total number of articles screened in the databases.

Characteristics of sources of evidence

The included studies varied widely in population with sample size ranging from 554 to 50,000 OPGs and geographical settings (South Korea, India, Turkey, Belgium, Brazil, Thailand, China, Germany, Malaysia). Both conventional forensic methods (Demirjian's, Kvaal's, Gustafson's) and AI models (CNNs, EfficientNet, DenseNet, visual geometry group [VGG], Inception, MobileNet, Age-Net, transfer learning approaches, and XGBoost) were employed. The main outcome of our research question was measured in terms of MAE, accuracy, area under the curve (AUC), and F1 scores.

Results of individual sources of evidence

AI for age estimation using developmental stages (Demirjian's method): YOLOv5, U-Net, and EfficientNet achieved mean average precision (mAP) of 0.995 and segmentation accuracy of 0.978.^ 9 ^ DenseNet-201 produced an MAE of 0.73 years, considerably lower than traditional Demirjian staging.^ 10 ^ Studies highlighted that the mandibular molars, especially second molars, as the most reliable markers for CNN-based age estimation. Artificial neural network-multilayer perceptron (ANN-MLP) models in specific populations minimized the difference between chronological age and dental age.^11,12^ These findings demonstrate that AI markedly improves accuracy compared with Demirjian's conventional approach.

AI for age estimation based on third molar development: DenseNet-121 with transfer learning achieved AUC values of 0.94 (males) and 0.83 (females) at the 18-year threshold. CNNs applied to large datasets classified individuals under 18 versus over 18 with accuracies of 87–88%.^13,14^ Root length-based DL models achieved 87.2% accuracy, slightly outperforming support vector machines (SVMs).^ 15 ^ Quadratic regression with Demirjian staging confirmed a 100% probability of being over 18 once mandibular third molars reached stage H.^ 16 ^

AI for PTRs: InceptionV4-based CNN models analyzing 12,827 radiographs achieved an MAE of 3.1 years (R^2^ = 95.5%).^ 17 ^ ML (XGBoost) reduced MAE to 4.65 versus 5.68 years with Kvaal's method.^ 18 ^

AI using Gustafson's criteria: Partial least squares (PLS) regression achieved an MAE of 4.15 years in males, while support vector regressor (SVR) achieved 3.81 years in females.^ 19 ^

AI in human identification, gender determination: Beyond age estimation, the role of AI is extended to tasks such as tooth identification, sex determination, and DVI. CNNs trained on large datasets of dental radiographs have achieved accuracies between 87% and 99% for tooth type classification, approximately 76% for sex determination, and median age estimation errors around 4.9 years. Notably, the predictive accuracy was higher for healthy, untreated teeth compared with those affected by restorations or pathology, underscoring the influence of dental alterations on model reliability.^ 20 ^

Transfer learning approaches, such as AlexNet-based models, further enhanced human identification by localizing and classifying, reaching 95% accuracy for tooth classification and 100% accuracy in distinguishing upper versus lower jaws.^ 21 ^ DL model like DENT-net achieved strong performance in forensic identification, reporting Rank-1 identification accuracy of 85%, with high precision (0.90), recall (0.95), and F1 score (0.92), while also operating at rapid processing speeds suitable for large-scale forensic casework.^ 22 ^ EfficientDet-D3 CNN architectures showed high accuracy in detecting dental features in mass disaster victims, with average precision of 99.1% and recall of 99.6%, though recall was lower for prostheses (84.3%) and root canal-treated teeth (89.2%).^ 23 ^ Additionally, CNN-based models trained on broader datasets of panoramic radiographs demonstrated age group classification accuracies of 53.8% within ±5 years, 95.1% within ±15 years, and 99.6% within ±25 years, with performance again favoring healthy dentitions.^ 24 ^

Comparative performance of AI models for age estimation: With hyperparameter optimization, EfficientNet-B4 achieved the lowest MAE of 0.562 years, while DenseNet-201 and MobileNet V3 demonstrated MAEs of 0.58 and 1.78 years, respectively.^ 25 ^ Large-scale CNN applications also reported strong results; a standardized CNN trained on 50,000 OPGs attained MAEs of 2.76 years (postmortem) and 3.26 years (antemortem) while significantly reducing processing time. Other models showed variable performance depending on dataset size and age group.^26?–28^ DentAge, a DL system trained on over 21,000 radiographs, achieved an overall MAE of 3.12 years, with its best performance in adolescents (MAE 1.94 years) but limited accuracy in the elderly (MAE 13.40 years).^ 29 ^ Similarly, EfficientNet-B5 performed best in younger cohorts (12–21 years), achieving an MAE of 2.8 years. Heuristic grouping strategies, such as applying a ±3-year tolerance, improved performance metrics for WideResNet and DenseNet architectures.^ 30 ^

Hybrid models integrating DL with ML classifiers, such as Age-Net (EfficientNetB0 + SVM), achieved an AUC of 0.970 and an accuracy of 76.41%.^ 31 ^ The studies comparing VGG19, ResNet152, SqueezeNet, and EfficientNet showed that VGG19 consistently outperformed others in simpler binary tasks (92.95% accuracy) but struggled in more complex classifications.^ 32 ^ Table 2 provides an overview of the data extraction process and key characteristics from the studies reviewed.^9??????????????????????–32^

Table 2.: Summary of data extraction from various studies.

Discussion

The integration of AI into dental radiology has shown significant promise in advancing clinical applications, particularly in age estimation, identification, and classification tasks. This scoping review demonstrates that AI-based models, especially CNNs, DenseNet, and EfficientNet architectures, achieve significantly higher accuracy in dental age estimation compared to traditional techniques. Several studies reported significantly lower MAE values, highlighting the precision of DL approaches in forensic applications. However, limitations such as a lack of dataset diversity, the need for external validation, and insufficient methodological transparency hinder the immediate application of these results in practice. The studies presented in this paper emphasize the potential and challenges of using AI models in the field of forensics.

Effectiveness of AI in automated age estimation using developmental stages of teeth in children and adolescents (Demirjian's method)

AI-based models have shown notable advancements over Demirjian's traditional method in estimating dental age in children and adolescents. Ong et al.^ 9 ^ developed a fully automated staging system on 5133 panoramic radiographs that integrated detection (YOLOv5), segmentation (U-Net), and classification (EfficientNet). The system achieved an mAP of 0.995 for detection and 0.978 accuracy for segmentation, while classification performance varied by tooth type, with the molar model yielding the highest F1 scores. Thus, the results highlight the ability of DL approaches to effectively replicate and, in some cases, enhance the staging reliability of Demirjian's technique. However, limitations such as low-resolution images, patient positioning errors, and imbalanced datasets were identified as critical barriers to achieving uniformly high accuracy. Also limited number of samples for early developmental stages was one of the limitations, as panoramic radiographs are not routinely taken at a young age, indicating a need for studies with larger sample sizes. Adding to these findings, Sivri et al.^ 10 ^ conducted a comparative study on 5898 pediatric radiographs (ages 4–17 years) and reported an MAE of 0.73 years using DenseNet-201, which was considerably lower than values typically achieved with the traditional Demirjian approach. This reinforces the evidence that CNNs can minimize prediction errors and provide greater consistency compared with conventional staging methods.

Another study by Matthijs et al.^ 11 ^ explored the application of the DenseNet-201 model for automated age estimation using 1639 panoramic radiographs. Their approach focused on five mandibular teeth (central incisor, canine, first premolar, first molar, and third molar), wherein a 10-stage modified Demirjian technique for staging was applied. The model's performance varied across tooth types, with the left mandibular second molar (tooth 37) showing the highest accuracy (MAE = 0.71), whereas the left mandibular central incisor (tooth 31) presented greater challenges. These findings highlight that mandibular molars, particularly second molars, offer more reliable developmental markers for CNN-based age estimation. A study by Bunyarit et al.^ 12 ^ in the Malaysian population investigated the applicability of Chaillet and Demirjian's modified eight-tooth method for dental age estimation among Malaysian Indian children and adolescents aged 5.00–17.99 years. A total of 1015 dental panoramic radiographs was analyzed. The study employed an ANN-MLP approach to develop new dental maturity scores for Malaysian Indian children and adolescents. The conventional method gave a difference between the chronological age and dental age of approximately 2.09 years in boys and 2.79 years in girls across all age groups. The ANN-MLP model produced a new prediction formula, resulting in a mean difference of approximately 0.035 ± 0.84 years for boys and 0.048 ± 0.93 years for girls, indicating enhanced accuracy. Variability in dental development due to genetic and environmental factors may also contribute to discrepancies in age estimation.

Effectiveness of AI in automated age estimation based on third molar development

Several studies have emphasized the potential of AI in improving the accuracy of age estimation through third molar assessment. Franco et al.^ 13 ^ investigated third molar development around the legal thresholds of 14, 16, and 18 years using a large dataset of 11,640 panoramic radiographs. Employing DenseNet-121 with transfer learning, their model achieved strong performance, with AUC values of 0.94 for males and 0.83 for females at the 18-year threshold, indicating its potential forensic utility. Murray et al.^ 14 ^ also applied CNNs to third molar development for determining the legal age of majority. Using 4003 panoramic radiographs, the CNN effectively classified individuals as “under-18” or “over-18” with high predictive accuracy. This study emphasizes the legal and judicial implications of using AI tools, especially in enhancing courtroom decision-making. Patil et al.^ 15 ^ assessed age estimation in an Indian population by applying a fully connected DL model based on second and third molar root lengths. The DL approach yielded higher predictive accuracy (87.2% for 2-Class classification) compared with traditional ML models such as SVM (86.4%). However, classification accuracy diminished for more complex stratifications (3-Class: 66%, 5-Class: 42.8%), suggesting that root length alone may not capture the full variability of dental maturation. Similarly, Duangto et al.^ 16 ^ employed quadratic regression based on Demirjian staging of third molars in a Thai population (n = 1867). Their results demonstrated low error values and found that once mandibular third molars reached stage H, the probability of being over 18 years was 100% in both sexes. However, the limited number of younger participants reduced the strength of conclusions for early developmental stages.

Overall, these findings indicate that AI models, especially CNNs and transfer learning architectures, are very effective at using third molar development for legal age determinations. However, differences among populations, sex variations, and dependence on single indicators like root length remain challenges that future research needs to address to improve broad applicability and forensic use.

Effectiveness of AI in automated age estimation based on PTR

The application of AI in PTR analysis has shown promising improvements over conventional methods. Oliveira et al.^ 17 ^ evaluated a large dataset of 12,827 dental images from Brazilian patients using an InceptionV4-based CNN model that incorporated radiological features such as pulp chamber dimensions and permanent tooth calcification stages. Their model demonstrated strong predictive performance, with an MAE of 3.1 years and a coefficient of determination (R^2^) of 95.5%, highlighting its reliability across different age groups and its ability to capture a broader range of anatomical features compared with traditional approaches.

Similarly, Pereira de Sousa et al.^ 18 ^ assessed 554 panoramic radiographs using ML algorithms in combination with Kvaal's method. Among the tested models, the XGBoost classifier outperformed the conventional Kvaal analysis, achieving an MAE of 4.65 years compared to 5.68 years with the manual approach. One of the limitations of the study was a smaller sample size that limited the generalizability of ML models.

Effectiveness of AI in automated age estimation based on Gustafson's criteria

Dai et al.^ 19 ^ investigated 10 ML models, based on modified Gustafson's criteria for dental age estimation in Southwest China, in their retrospective study involving 851 samples. The PLS regressor model achieved the best performance in males with an MAE of 4.151 years, while the SVR model performed well in females with an MAE of 3.806 years.

Effectiveness of AI in automated identification of tooth, gender determination, and DVI

Milošević et al.^ 20 ^ analyzed a dataset of 86,495 tooth X-ray images using CNN models, with emphasis on sex and tooth type classification. Their findings revealed accuracies of 76.41% for sex assessment, 87.24% to 99.15% for tooth type determination, and a median absolute error of 4.94 years for age estimation. Notably, the study highlighted that models trained on healthy, unaltered teeth performed significantly better, emphasizing the impact of dental alterations (such as restorations and pathology) on predictive accuracy. Sathya and Neelaveni^ 21 ^ proposed a transfer learning approach using AlexNet for human identification based on dental traits. Their model operated in three stages: localizing the query tooth, classifying it into one of the four categories (molar, premolar, canine, incisor), and numbering it according to the universal system. This system achieved a classification accuracy of 95% across tooth types, 100% accuracy in distinguishing between upper and lower jaws, and high accuracy in numbering teeth. It surpassed traditional manual methods in both precision and speed. Fan et al.^ 22 ^ developed DENT-net, a DL system for automated human identification trained on 15,369 panoramic radiographs from 6300 individuals. This system achieved a Rank-1 identification accuracy of 85.16% and demonstrated strong performance metrics (precision 0.90, recall 0.95, F1 score 0.92, AUC 0.996). Additionally, its rapid processing speed (approximately 10 ms per feature extraction) makes it suitable for large-scale forensic casework. However, the system had difficulty with cases involving mixed dentition, which restricts its universal applicability. Choi et al.^ 23 ^ employed the use of AI in DVI by employing an EfficientDet-D3 CNN architecture to detect dental features on 1638 panoramic radiographs. Their model showed an average precision 99.1% and recall 99.6%. For teeth with dental prostheses the was recall 84.3% and for root canal treated teeth the recall was 89.2%, Thus the accuracy of AI models was more for natural teeth compared to those with treated teeth. Furthermore, Kim et al.^ 24 ^ explored deep neural networks for age estimation across 10,023 images, to check whether CNNs can classify dental panoramic radiographs into age groups even without precise age information, using approximate age groupings. Their study results gave an accuracy of 53.8% within ±5 years, 95.1% within ±15 years, and 99.6% within ±25 years. Also the performance was best with healthy teeth (accuracy = 96.45%), and slightly lower with treated teeth. In summary, these studies illustrate the expanding role of AI in forensic odontology beyond merely estimating chronological age. AI-driven models exhibit high accuracy in tooth classification, gender determination, and human identification, proving useful in mass disaster scenarios. However, challenges remain regarding variability due to dental alterations, population diversity, and mixed dentition, which future research should address to enhance forensic applicability.

Comparison of various AI models used for age estimation

Hyperparameter-optimized deep learning models

Büyükçakır et al.^ 25 ^ tested EfficientNet, DenseNet-201, and MobileNet V3, highlighting that EfficientNet-B4 achieved the lowest MAE of 0.562 years. Their study highlighted the use of hyperparameter optimization, which substantially enhanced accuracy. Nonetheless, challenges such as limited training data and computational complexity restricted generalizability, reflecting a common barrier across dental AI studies.

CNNs for identification

Kim et al.^ 26 ^ employed a modified VGG16 model for human identification, which gave accuracies above 80%, despite image variability. The application of Grad-CAM visualization provided interpretability by highlighting diagnostic features. While effective, dataset imbalance and image distortion reduced reliability, emphasizing the need for data augmentation and preprocessing techniques. Bussaban et al.^ 27 ^ compared VGG19, ResNet152, and SqueezeNet across multiclass classification tasks. VGG19 consistently outperformed others, achieving 92.95% accuracy in binary classification but lower performance in more complex multiclass settings. Though EfficientNet showed promising results, smaller sample sizes constrained model learning. Heinrich^ 28 ^ trained a standardized CNN on 50,000 OPGs, achieving MAEs of 2.76 years (postmortem) and 3.26 years (antemortem), while reducing database processing times by 96%. Bizjak and Robič^ 29 ^ used DentAge, which achieved an overall MAE of 3.12 years. It gave the best results in adolescents (MAE 1.94 years), while in the elderly individuals, it did not give significant results (MAE 13.40 years), illustrating the difficulty of estimating age in older individuals. Similarly, Kahm et al.^30^ found that heuristic grouping with a tolerance of ±3 years improved F1 scores for WideResNet and DenseNet models.

Hybrid approaches

Baydogan et al.^ 31 ^ introduced Age-Net, a hybrid model integrating CNNs with ML (e.g. EfficientNetB0 + SVM). The model achieved an AUC of 0.970 and accuracy of 76.41%, demonstrating that collaborative frameworks can balance precision and strength. However, their dataset's restricted age range limited external applicability. Expanding datasets to broader age groups would enhance translational value.

Transfer learning and anatomical feature utilization

Mu and Gang^ 32 ^ assessed the accuracy of transfer learning models (ResNet, EfficientNet, VGGNet, DenseNet) for age estimation using panoramic radiographs. The best model was EfficientNet-B5 with an MAE of 2.8 years. The younger age group (12–21 years) gave the best results with this model. Their findings also highlight that the different anatomical structures, such as maxillary sinus and angle of the mandible, could contribute to age prediction.

Thus, after comparing the various AI models, EfficientNet variants and VGG-based models often gave the best results, with hybrid and group strategies (e.g. Age-Net) providing enhanced robustness. Dataset size and age distribution were consistently identified as key factors influencing performance, with smaller or skewed datasets leading to less generalizability. Moreover, interpretability techniques (e.g. Grad-CAM) and heuristic groupings (±3 years) have become strategies to improve clinical and forensic suitability. Collectively, these findings show that while AI models can significantly outperform traditional methods, standardizing datasets, expanding age coverage, and validating across populations are still essential for forensic adoption.

AI applications in forensics extend to various other dental fields, including periodontics, endodontics, oral medicine/pathology, prosthodontics, pediatric dentistry, orthodontics, and orofacial pain. These AI models are utilized for diagnosis, treatment planning, clinical decision-making, and prognosis prediction across different dental specialties. This technology benefits both individual patients and the community as a whole. Recently, the implementation of generative AI has emerged as a valuable tool for dental practitioners.^4,6,7^ As dental education lacks AI training, incorporating this in the postgraduate program can help to bridge the gap.^ 5 ^

Challenges and limitations

Data and bias

The lack of standardized data exchange frameworks and high-quality datasets is the main obstacle to AI in dentistry. Data access and sharing are also hampered by privacy and security issues. Furthermore, the fairness and generalizability of AI models can be impacted by biases in training datasets, which can result in uneven performance across demographic groups.^ 4 ^

Ethics and accountability

Many AI systems lack transparency and explainability, which hinders accountability in clinical decision-making and erodes trust. Ethical concerns include risks to patient privacy and the potential for algorithmic bias.^ 4 ^

Implementation barriers

The transition of AI from research to practical application is hindered by various regulatory, technical, and infrastructural challenges. Creating clinically reliable AI models requires considerable time, access to correctly labeled datasets, and substantial computational resources. Additionally, limited understanding of AI among dental professionals, along with the high costs associated with implementation and maintenance, further limit the widespread acceptance.^ 4 ^

Conclusion

The findings of this review stress the significant potential of AI algorithms in automated age estimation processes using OPGs. The studies reviewed demonstrate that AI-based models, particularly DL architectures such as CNNs, EfficientNet, DenseNet, and hybrid models like Age-Net, exhibit superior accuracy, precision, and reliability compared to conventional age estimation techniques. These AI-driven models show promising results in reducing human error, enhancing efficiency, and improving forensic and clinical decision-making.

Despite these advancements, several challenges persist, including data scarcity, imbalanced datasets, variability in dental development across populations, and the need for standardized protocols. Although AI models have shown high performance across various age groups, further research is needed to improve generalizability across diverse populations and optimize models for forensic and clinical applications. Future studies should prioritize larger, broad-based datasets, strong validation techniques, and the integration of AI-based age estimation models into routine forensic field.

In conclusion, AI-driven age estimation using OPGs presents a transformative approach with significant forensic and clinical implications. While AI-based models may not yet fully replace traditional methods, they serve as powerful adjuncts that can enhance accuracy and efficiency. Continued research, along with the establishment of standardized AI protocols, will be essential for the widespread implementation of these models in forensic odontology and related fields.