Clinically interpretable deep learning for pediatric dental age estimation with explainable AI

Abstract

Background

Chronological age estimation is essential for forensic science, law enforcement, and treatment planning. Dental orthopantomography (OPG) is widely used for this purpose, yet most conventional approaches remain rule based, slow, examiner dependent, and sensitive to image quality and dental variations.

Methods

Pediatric OPGs from children aged 5–15 years were used to train and evaluate five deep learning networks, namely ConvNeXt-Tiny, ResNet-50, EfficientNet-B0, DenseNet-121 and Xception. With these models, ensemble strategies were developed and evaluated. Model performances were assessed using MAE, RMSE and R². Additionally, explainable AI (XAI) heatmaps were generated to visualize and examine age informative regions.

Results

The proposed median ensemble achieved the best performance with an MAE of 0.616 years (≈ 7.4 months), outperforming all single networks. XAI analyses indicated that the models based their predictions on clinically plausible dental regions, with Xception providing the most precise and concentrated attributions.

Conclusions

This study demonstrates that the median ensemble deep learning approach provides robust and examiner-independent pediatric age estimation from OPGs. This framework significantly reduces prediction errors compared with individual models. Furthermore, the integration of XAI-based visual explanations enhances interpretability, positioning this approach as a valuable decision support tool for forensic and dental practice.

Supplementary Information

The online version contains supplementary material available at 10.1186/s12903-026-07961-z.

Background

There is a continuous need for reliable age estimation in forensic, legal, criminal, law enforcement, and clinical contexts [1–3]. According to UNICEF’s December 2024 update, although global birth registration coverage for children under five has risen to 77%, approximately 150 million children remain unregistered. In addition, around 55 million children reported as registered do not possess a birth certificate [4]. In sub-Saharan Africa, the birth registration rate is about 51%, and roughly half of the world’s unregistered children (~ 90 million) are in this region [5]. Teeth, being among the most indestructible and best preserved parts of the body throughout life, are particularly suitable for age estimation [6, 7]. Due to the increasing importance of age estimation in forensic practice, the interdisciplinary Study Group on Forensic Age Diagnostics (AGFAD) was founded including medical doctors, dentists, radiologists, and anthropologists. AGFAD aims to define scientific and ethical standards while increasing the reliability of applications in this area [8]. According to their report, among the most suitable methods is the X-ray examination of dentition, and an expert report is required for the most accurate age estimation [9]. As X-ray imaging can display all teeth in a single frame, it is helpful not only for diagnosis and treatment but also for estimating age. Many methods have been developed utilizing X-ray images, such as Cameriere, Demirjian, Moorrees, Fanning & Hunt, and Nolla [10–13]. These approaches are based on the eruption stages, developmental mineralization, and root growth of teeth and require time consuming and examiner dependent manual processes which can lead to inconsistent outcomes [14]. Recent advancements in machine learning (ML) and pattern recognition have produced automated solutions to these problems [15–18].

Several studies have used supervised ML to enhance conventional dental age estimation methods. In these ML-assisted approaches, tooth development is first scored by an expert using schemes such as Demirjian, Cameriere or Willems, and regression models are then trained to predict chronological age from these indices. For example, Shen et al. [15] applied ML algorithms to Cameriere based measurements in Eastern Chinese children showed that ML regressors can reduce error compared with the original Cameriere formulas. Patil et al. [19] trained ML and artificial neural network models on manually measured root lengths of mandibular second and third molars to estimate age in South Indian subjects. Galibourg et al. [20] compared several supervised ML algorithms with Demirjian and Willems methods. Demirjian stages and sex as inputs are required to be assigned manually before ML prediction. Taken together with other ML assisted studies [21–24], these results show that supervised models can improve Cameriere and Demirjian based formulas across different populations. However, such approaches still depend on manual staging or morphometric measurements, which are time consuming and subject to inter and intra-observer variability.

To overcome these limitations, several studies have begun to apply deep learning models to dental radiographs [25, 26]. However, many of these methods still rely on predefined staging systems. Balel et al. [27] addressed the limitations of manual staging by developing a fully automated YOLO-based pipeline that detects mandibular teeth and assigns Demirjian stages via a graphical interface, achieving accuracy comparable to human assessors. Shi et al. [28] proposed a three-step framework combining YOLO-based tooth localization with a custom staging network and a final dental age assessment step based on single-group meta-analysis of predicted stages to derive dental age on an external test cohort. Other studies have adopted deep learning formulations that treat age estimation as a classification problem rather than continuous regression [29–31]. Kahaki et al. [32] proposed a multi-class classification approach for Malaysian children and trained a deep convolutional neural network using engineered features extracted from segmented mandibular regions. Murray et al. [33] employed a DenseNet-121 to classify individuals as above or below the legal age of majority using panoramic radiographs.

Looking across these studies, the current deep learning literature on dental age estimation still leaves several important gaps. Existing methods often remain tied to Demirjian-type staging or other handcrafted indices, focus on specific tooth groups instead of exploiting information from the entire OPG, and provide limited insight into which dental structures drive predictions. In addition, many reports formulate age assessment as a classification task instead of continuous age regression. Although such classifiers can be effective for legal age decisions, they typically output broad age categories rather than precise age estimates, limiting their suitability for individualized clinical and forensic assessment.

To address these issues, we propose lightweight ensemble methods that aggregate per-image predictions from diverse deep learning models. The advantages and contributions of the developed method are as follows:

• Providing a fast, objective, reproducible, and accurate alternative to conventional age estimation methods.

• Leveraging information from the entire OPG image, rather than restricting the analysis to predefined regions of interest.

• Instead of classification, performing continuous age regression to enable more precise age estimates for forensic, legal, law-enforcement, and clinical applications.

• Designing and systematically evaluating lightweight ensembles that combine predictions from heterogeneous deep learning models to improve accuracy and robustness.

• Providing post-hoc explainability via Integrated Gradients (IG) to verify that model attention aligns with clinically plausible, age-informative anatomical structures.

The paper is organized as follows. Section 2 presents the materials and methods, with emphasis on the dataset and the proposed ensemble model. Section 3 reports the experimental findings, including the explainable artificial intelligence (XAI) analysis. Section 4 provides a discussion of the results, clinical relevance, and future directions. Section 5 concludes the paper.

Materials and methods

Data collection and preprocessing

This retrospective study was conducted on 1,222 pediatric orthopantomographic (OPG) images from 608 girls and 614 boys aged 5–15 years, representing the Turkish pediatric population in Karabuk, Turkiye. The images were acquired between 2010 and 2020 at Karabuk Oral and Dental Health Education and Research Hospital using an Orthopantomograph OP100 D unit (Instrumentarium Dental, Tuusula, Finland) operating at 60 kV and 6.4 mA, with a 138 mm vertical field of view and an exposure time of 17.1 s. The cohort covered the full pediatric range seen in routine care. As shown in Fig. 1, the mean age was 9.48 ± 2.69 years with a median of 9.11 years. Chronological age was calculated to one decimal place using the dates of birth and imaging, while all other metadata were fully anonymized. Ethical approval was granted by the Non-interventional Clinical Research Ethics Committee of Karabuk University (Approval Date: 25/7/2025, No: 2025/2416).

Fig. 1.

Age distribution of dataset

All images were exported from the clinical archive as DICOM files, standardized in orientation, and then converted to PNG format. Quality control procedures were applied. Patients with generalized hypodontia or extensive tooth loss unrelated to normal exfoliation were excluded. Radiographs showing obvious pathological changes or developmental anomalies, as well as images that were completely black, severely blurred due to motion or positioning errors, or had incomplete views of the dentition, were also excluded. Cases with orthodontic appliances and minor tooth loss were retained in the dataset to reflect routine clinical imaging conditions. Example OPG images are shown in Fig. 2. Metadata fields were checked for consistency and cleaned where necessary. To prevent duplication, both file hashes and perceptual hashes were computed and duplicate images were removed. The final CSV file contained the image filename and the continuous age in years for each case.

Fig. 2.

Example OPG images from the dataset

Although OPGs are originally grayscale, each image was converted to three-channel RGB by replicating the single intensity channel across the R, G, and B channels to match the input requirements of the deep learning architectures employed in this study. After RGB conversion, pixel intensities were linearly scaled and normalized using the standard ImageNet mean and standard deviation. Images were resized to 224 × 224 pixels for the ConvNeXt-Tiny, ResNet-50, EfficientNet-B0, and DenseNet-121 architectures, and to 299 × 299 pixels for the Xception model.

Deep Learning Networks

ConvNeXt-Tiny [34], ResNet-50 [35], EfficientNet-B0 [36], DenseNet-121 [37], and Xception [38] are ImageNet pretrained deep convolutional neural networks. They are utilized for feature extraction from images. These networks have complementary architectural characteristics and represent widely used deep learning networks in medical image analysis [39–44]. Table 1 summarizes the key architectural differences between the deep learning networks used in this study. Architectural details and the modifications made to adapt each deep learning model for age regression are provided in Supplementary S1.

Table 1.

Comparison of implemented deep learning networks [45–49]

Model		Advantage	Disadvantage	Key Feature	Architecture
ConvNeXt-Tiny [34]	Modern blocks, good accuracy, stable optimization		Computationally intensive, memory cost	Depthwise convolutions, LayerNorm, GELU, LayerScale, DropPath	Modern residual deep learning network with optimized blocks
ResNet-50 [35]	Simple, reliable, strong transfer, large ecosystem		Less parameter efficient than newer architectures	Bottleneck blocks with skip connections	Deep learning network with residual connections
EfficientNet-B0 [36]	High accuracy per FLOP, compact, edge friendly		Sensitive to training settings, input resolution	Inverted residuals, SE, SiLU	Inverted residual deep learning network
DenseNet-121 [37]	Strong gradient flow, feature reuse, good generalization		Concatenation increases memory and latency	Each layer adds k channels, transition and compression layers	Densely connected deep learning network
Xception [38]	Depthwise separable convolutions, improved representation learning		Computational complexity	Residual shortcuts	Fully separable deep learning network

The proposed ensemble model

Three algebraic ensemble models were implemented to enhance the reliability of predictions and harness the diverse strengths of multiple deep learning networks. For each image, the scalar outputs produced by the individual models were combined as input to the proposed ensemble models. The first ensemble (Eq. 1) computes the unweighted mean of predictions from M trained models, where the 𝑚-th network generates a scalar prediction for image i (𝑚 =1, …, M). The second ensemble (Eq. 2) computes a weighted mean. The weight of the 𝑚-th model is inversely proportional to the squared out-of-fold (OOF) MAE, , computed across the validation folds. These weights are normalized and subsequently renormalized if specific model predictions are unavailable. A small constant 𝜀 = 10⁻¹² is included to ensure numerical stability. The third ensemble (Eq. 3) takes the median of the available predictions for each image. If fewer than 𝑀 predictions are available for an image, the median is computed over the available set. In the final decision stage, all models contribute to the ensemble, and predictions are aggregated using weighted averaging, simple averaging, or median-based fusion, as illustrated in Fig. 3.

1

Fig. 3.

Workflow of the proposed ensemble models

2

3

Training setup and hyperparameter tuning

Experiments were conducted on a single NVIDIA A100 (40 GB) GPU under Ubuntu 22.04 with CUDA 12.1 and cuDNN enabled. To ensure a fair comparison, all deep learning models were trained and evaluated on the same preprocessed OPG dataset using a common training pipeline. An age-stratified patient level split with two year intervals was implemented to prevent data leakage while ensuring balanced age distributions across subsets. A 20% held-out test set was isolated, and the remaining 80% was used for 5-fold CV (Fig. 4). Training augmentations (random horizontal flip, 𝑝=0.5; rotation, ± 10°; brightness/contrast jitter) were applied exclusively to the training folds.

Fig. 4.

CV procedure for each model

Hyperparameters were tuned separately for each backbone using a grid search over optimizer type, learning rate, and batch size. Detailed hyperparameter configurations, including search ranges and final selected values, are provided in Supplementary S2 for reproducibility. To verify the adequacy of the training duration, validation MAE was monitored across epochs for all deep learning models. As illustrated in Supplementary S3, rapid convergence was observed and validation MAE stabilized within the first 3–4 epochs. Therefore, training was capped at five epochs to reduce the risk of overfitting. Finally, each deep learning model was retrained on the combined training and validation data using the optimal hyperparameters to obtain the final model.

Explainable Artificial Intelligence

XAI methods aim to increase the transparency of model decisions by revealing which input regions contribute to individual predictions [50–52]. IG estimates per-pixel contributions by integrating gradients along a straight line path from a black reference to the input. The resulting attribution can be positive or negative according to the sign of its effect on the prediction. Following model and ensemble evaluation, post-hoc attribution maps was computed with IG to qualitatively assess where the networks focus on OPGs. IG was implemented using the Captum library, and all images underwent the same preprocessing pipeline as during model inference. Channel wise attributions were summed across RGB, z-score normalized, and only the top 10% of the positive attribution mass was visualized. Prior to rendering, Gaussian smoothing was applied (σ = 7).

Evaluation metrics

Performance was assessed using regression metrics, uncertainty estimates, statistical comparison, and qualitative XAI evaluation. Regarding predictive accuracy, MAE quantifies the average absolute deviation between the predictions, and the ground truth ages in the natural unit of the task (Eq. 4). It is directly interpretable in clinical terms. Root Mean Squared Error (RMSE) emphasizes larger mistakes by squaring residuals before averaging, making it sensitive to occasional substantial under or over-estimation (Eq. 5). The coefficient of determination (𝑅²) expresses how much of the variance in true ages is explained by the model relative to a baseline mean predictor (Eq. 6) [53]. Errors are reported in years unless stated otherwise.

4

5

6

In addition to point estimates, 95% confidence intervals (CIs) were computed for the test-set MAE of the best single model and the best ensemble. CIs were obtained by non-parametric bootstrap resampling of the test cases. To compare these two models, per-image absolute errors were assessed using the Wilcoxon signed-rank test. A p-value < 0.05 was considered statistically significant. Additionally, a three-level rating scheme was applied to the IG heatmaps to assess whether the models focused on age-relevant structures. A pediatric dental expert with over 10 years of experience performed a blind evaluation of the IG heatmaps. Without knowing the model identity, the expert assigned each image to one of three categories:

• Strong match: the heatmap concentrated on expected age-informative dental features, such as crown and root morphology of developing teeth, narrowing of the pulp chamber and root canals, root length and shape, and the presence of apex closure in premolars and molars.

• Partial match: the main focus was close to the teeth and included some age-informative structures, but the attention was more diffuse, extended into surrounding alveolar bone or only partially overlapped with clinically relevant regions.

• Mismatch: the model primarily focused on non-informative or irrelevant areas, such as radiographic labels, metal artifacts, soft tissues, or distant skeletal structures, rather than on the dentition.

Results

In this section, test-set performance of the individual networks and the proposed ensembles is reported using MAE, RMSE, and R² supported by error distribution and residual analyses. Uncertainty and model comparisons are summarized with 95% bootstrap CIs and the Wilcoxon signed-rank test. Finally, XAI findings are presented using IG heatmaps and expert ratings. Among the individual models, ConvNeXt-Tiny achieved the best performance, followed by Xception and DenseNet-121, whereas EfficientNet-B0 and ResNet-50 exhibited higher errors. The proposed median ensemble further reduced the test MAE from 0.640 to 0.616 years. Additionally, both RMSE and R² metrics were improved by the ensemble models compared to any single model (Table 2).

Table 2.

Results of trained models on test set

Model	MAE	RMSE	R²
ResNet-50	0.8951	1.1675	0.8130
EfficientNet-B0	0.8843	1.1570	0.8164
DenseNet-121	0.7852	0.9945	0.8643
Xception	0.6998	0.9142	0.8854
ConvNeXt-Tiny	0.6399	0.8319	0.9051
Ensemble (Mean)	0.6232	0.8177	0.9083
Ensemble (Weighted)	0.6203	0.8126	0.9094
Ensemble (Median)	0.6160	0.8211	0.9089

Figure 5 displays the distribution of prediction errors across models. Among the single models, ConvNeXt-Tiny shows relatively small interquartile ranges (IQR) with median values close to zero. Xception also has a moderate IQR but a slightly negative median and outliers on both sides. Both networks display centered spread and balanced error distributions. DenseNet-121 plot shows a relatively positive median which suggests that it tends to overestimate, and a moderate IQR compared to the other models. The EfficientNet-B0 plot shows a negative median and a wide IQR, which means that the model is consistently underestimating and higher variability in the prediction errors. On the other hand, ResNet-50 plot has the widest IQR and the longest whiskers. This pattern shows that there are more extreme values and outliers in the residuals. However, the mean ensemble significantly lowers the IQR when compared the single networks indicating lower error variance. Also, its median is the closest to zero among the ensembles. The weighted ensemble achieves small IQR compared to single deep learning networks, but its median moves slightly into the positive range. This shift is primarily due to contributions from certain models that tend to overestimate. The median ensemble displays similar IQR with weighted ensemble, showing a median nearly at zero. This indicates the highest resistance to outliers. Notably, this method exhibits fewer outliers than all the networks.

Fig. 5.

Prediction error boxplot

In Fig. 6a, true versus predicted ages for the median ensemble are plotted. Most data points are tightly aligned along the y = x identity line, and the model achieves R² = 0.9089. The majority of cases lie within ± 1 year and nearly all within ± 2 years, indicating good overall calibration and only moderate scatter compared with ConvNeXt-Tiny and Xception. The ECDF of the absolute error is shown in Fig. 6b. The median ensemble curve sits above others in the 0.5–1.5 year error range, indicating consistent performance in practice. At the 1-year threshold, approximately 80% of predictions from both the median ensemble and ConvNeXt fall below this error. At 2 years, 98% of predictions are captured for both the median ensemble and ConvNeXt, leaving only a small tail with errors ≥ 2 years. The median ensemble reduces these larger error outliers to some extent as well. Figure 6c displays the MAE values along with 95% CIs for ConvNeXt-Tiny (0.561–0.719 years) and the median ensemble (0.575–0.657 years). Per-image absolute errors were significantly lower for the median ensemble than for ConvNeXt-Tiny according to a Wilcoxon signed-rank test (p = 0.043). Figure 6d plots residuals against true age to highlight age-dependent patterns. Predictions for younger children (5–7 years) tend to be slightly overestimated, whereas those for older adolescents (13–15 years) show slight underestimation. Residuals remain approximately centered around zero across the 5–15 year range, and no strong systematic age-related bias is apparent.

Fig. 6.

Performance evaluation of the proposed method. a True vs. Predicted Age with ± 1 and ± 2 year error bounds. b Empirical Cumulative Distribution Function (ECDF) of absolute errors. c Comparison of MAE with 95% CIs between the best single model and the median ensemble. d Residual plot for the median ensemble

All IG heatmaps were generated post-hoc on test predictions. Warm colors such as red and yellow indicate regions with higher positive contribution to the predicted age, whereas non-highlighted areas contribute minimally. Heatmaps were produced with Gaussian smoothing, and representative examples are shown in Fig. 7.

In Fig. 7a, Xception gives a near-perfect estimate (true 7.0 y, pred 7.1 y). The IG heatmap concentrated over the mandibular posterior mixed dentition region. It particularly highlighted the first permanent molars and early premolar germs, which show short roots and wide open apices. In Fig. 7b, ConvNeXt-Tiny provides the closest single-network estimate (true 9.0 y, pred 8.8 y). IG attributions cluster bilaterally along the mandibular posterior mixed dentition arc. The highlighted areas include the first permanent molars and developing premolar and second molar germs, where the roots are elongated and the open apices are partially narrowed. In Fig. 7c, ResNet-50 and DenseNet-121 provided accurate estimates for a 14.7-year case (ResNet-50: 14.3 y; DenseNet-121: 15.1 y). IG attributions again concentrated on posterior dentoalveolar regions. DenseNet-121 mainly highlighted the mandibular posterior second molar areas, whereas ResNet-50 emphasized a band along the maxillary posterior arch, strongest around the upper right molars. Little relevance is assigned to anterior teeth or soft tissues. The highlighted teeth show fully developed premolar and first-molar roots, with second molars exhibiting nearly closed apices.

Fig. 7.

IG based sample test images predicted by single networks

For each model, IG heatmaps were generated for 50 randomly selected images from the test set. These visualizations were then evaluated by the expert using the three level rating scheme, and the results are summarized in Table 3. Strong matches were observed most frequently for Xception (60%) and ConvNeXt-Tiny (52%), whereas EfficientNet-B0 showed the lowest proportion of strong matches (22%). Partial matches were most common for EfficientNet-B0 and DenseNet-121 (62% and 44%), indicating that their attention was often more diffuse. Mismatches were frequent for ResNet-50 (32%) and least common for Xception (3%).

Table 3.

Three level rating results for IG heatmaps for each deep learning model

Model	Strong match (%)	Partial match (%)	Mismatch (%)
ConvNeXt-Tiny	52	36	12
DenseNet-121	38	44	18
EfficientNet-B0	22	62	16
ResNet-50	28	40	32
Xception	60	37	3

Discussion

In this study, several state-of-the-art deep learning models were evaluated for pediatric age estimation. After hyperparameter tuning using CV on the training and validation splits, all five deep learning models achieved sub-year errors on the held-out test set. ConvNeXt-Tiny was the best-performing single model for age estimation. Rather than relying on a single architecture, lightweight ensemble strategies were designed to aggregate per-image predictions from multiple deep learning models. Among these, the median ensemble is particularly suitable from a clinical perspective because it reflects a consensus across models and is less sensitive to extreme predictions from any individual model. This reduces the risk of rare but large errors that could arise if decisions were based on a single model, providing more stable and robust support for decision making.

A direct comparison with previous studies is difficult because datasets differ in sample size, age range, population characteristics, annotation protocols, and imaging quality. Nevertheless, when recent OPG based ML and deep learning systems are considered, our results compare favorably. The proposed median ensemble achieved a test MAE of 0.616 years (≈ 7.4 months) in 5–15 year-old children, placing its performance within the best reported range for pediatric dental age estimation while offering an end to end workflow. Galibourg et al. [20] analyzed a large French dataset of 3,570 panoramic radiographs from individuals aged 2–24 years, in which dental maturation was manually scored using Demirjian’s stages and then modelled with ten supervised regression algorithms. For the 4–16 year subgroup, which closely overlaps with our age range, support vector machines reached an MAE of 0.729 years and improved on the Demirjian and Willems reference tables. However, their workflow remains dependent on the initial manual staging of teeth, which subjects the process to inter and intra-observer variability. Our ensemble deep learning model further reduced this error to 0.616 years without manual staging. These findings suggest that deep learning networks capture age-related cues directly from the raw radiographs that are not fully represented by categorical staging labels. Similar trends are observed when comparing our results with recent automated systems. Shi et al. [28] employed a multistage framework for children aged 3–15 years, combining tooth detection, Demirjian type staging and meta-analysis to achieve an MAE of 0.72 years. In contrast, our end-to-end ensemble attains slightly lower error without relying on explicit staging steps, indicating that high precision can be achieved with simpler, image-driven architectures. Ataş et al. [54] proposed a modified InceptionV3 based model on panoramic radiographs covering a broad age range from late childhood to older adults and reported 2.39 years of MAE for the 10–19 year subgroup. Lower error observed in our pediatric cohort likely reflects both the narrower age range and the use of a specialized ensemble focused on developmental years. Sivri et al. [55] compared several pre-trained deep learning networks with Demirjian’s method in 4–17 year old Turkish children and reported that DenseNet-121 achieved an MAE of 0.73 years, outperforming Demirjian while providing a faster, examiner independent workflow. Despite the smaller dataset, aggregating multiple networks provided better results. Rather than interpreting this as a superiority, it is evidence that ensemble based, image level models are a promising complementary direction alongside single deep learning network and updated Demirjian type protocols.

Several authors have stressed that explicability of results is essential in forensic medicine and that high performing ML models risk functioning as “black boxes” with limited transparency [56–59]. In line with these concerns, we incorporated IG to generate visual explanations of model predictions. IG allows experts to verify whether the deep learning networks focus on clinically plausible, age-informative regions on OPGs and thus helps to bridge the gap between high-accuracy deep learning and the level of transparency expected in forensic reporting. In our analysis, IG maps showed that each deep learning network tended to emphasize slightly different age-related structures. The IG patterns in Fig. 7 can be related to well-known dental maturation indicators.

In the ~ 7-year case (Fig. 7a), the emphasis on mandibular first molars and early premolar germs with short roots and widely open apices is broadly compatible with Demirjian stages around D–E. In the ~ 9-year case (Fig. 7b), the highlighted posterior teeth show more advanced root elongation with partially narrowed apices, broadly consistent with stages around E–F. In the ~ 15-year case (Fig. 7c), the model attention remains concentrated in posterior regions where root maturation is advanced and apical closure is near complete, consistent with late stages around G–H. The qualitative agreement between the IG-highlighted regions and well-known Demirjian-style maturation cues provides a clinically plausible interpretability check. Specifically, attention being assigned to posterior tooth-root development and apical closure patterns suggests that predictions are driven by age-informative dental maturation features rather than by irrelevant background structures. Stages were not explicitly annotated, so these references are used only as interpretability anchors and should be interpreted as qualitative alignment rather than formal validation. Importantly, the case containing a fixed space maintainer shows minimal attribution on the metallic appliance, while attention remains focused on developing roots, suggesting that this artefact did not dominate the prediction in that example. This further reduces concern that obvious artefacts dominate the decision.

In Fig. 7b, ConvNeXt-Tiny based its correct prediction primarily on bilateral mandibular molars, whereas ResNet-50 focused more on the apices of the maxillary roots. This illustrates that the models do not rely on a single fixed region in the mouth but extract information from multiple areas, and that combining different architectures can increase the reliability of the overall decision. IG maps also make it possible to assess whether predictions arise from clinically meaningful regions rather than artefacts or background structures.

A three-level IG qualitative assessment was conducted by a pediatric dental expert to evaluate the anatomical relevance of each model’s attention. This evaluation showed that Xception and ConvNeXt-Tiny consistently aligned with clinically relevant, age-related structures, which correlated with their better performance. Despite the generally high accuracy, the XAI results also revealed cases in which some models did not concentrate on clinically meaningful areas (Fig. 8). Furthermore, expert evaluation of the heatmaps revealed specific behavioral patterns, characterized by a general tendency among the models to prioritize permanent teeth over deciduous ones. Specifically, Xception performed well in the permanent dentition stage, predominantly focusing on information from the mandible, whereas EfficientNet-B0 struggled to focus on relevant dental structures, often attending to air spaces instead of the teeth. ConvNeXt-Tiny displayed stage dependent behavior. In mixed dentition scans, its focus was moderate with a bias toward permanent teeth, whereas it demonstrated excellent attention alignment in the permanent dentition stage.

Fig. 8.

Examples with weaker focus on dentition: a ResNet-50, (b) EfficientNet-B0, (c) DenseNet-121, (d) ConvNeXt-Tiny

In practice, the achieved accuracy is particularly significant in contexts where reliable birth records are lacking, such as asylum procedures or determinations of criminal responsibility. However, given the error margin, predictions near critical legal thresholds require cautious interpretation. In such borderline forensic cases, this automated tool serves as valuable objective evidence to complement, rather than replace, comprehensive biological profiling. From a clinical perspective, predictive performance is highly effective for the mixed dentition phase. Since treatment timing relies more on developmental stages than precise chronological age, deviations of ± 1 year are generally considered as clinically acceptable. For example, decisions regarding the timing of preventive orthodontic treatment, space management, or growth-related interventions are usually based on the stage of tooth eruption, root development, and growth status, and a deviation of a few months rarely alters the clinical approach. Nevertheless, age estimation requires careful attention in specific scenarios, such as in pediatric endocrinology for distinguishing between early and late puberty or determining growth-related interventions. In these cases, significant overestimation could lead to premature intervention, while underestimation might delay necessary treatment. To mitigate these risks, dental age estimates must not be the sole criterion but should be interpreted in conjunction with skeletal maturation and other clinical findings, particularly when treatment timing is critical.

This study has several limitations, primarily concerning data generalizability, analytical scope, and technical constraints. In terms of generalizability, the data were obtained from a single center using a single device type. This uniformity restricts external validity and may limit the model’s robustness to variations in image quality found in other institutions. The cohort, drawn from a single institution in northern Turkiye, largely represents a relatively homogeneous local pediatric population. This limits applicability to populations with different ancestry, growth patterns, or nutritional status. Regarding analytical scope, although the sex distribution was balanced, sex-stratified training was not performed, meaning potential performance differences between sexes cannot be excluded. Similarly, while the dataset included cases with orthodontic appliances, the specific influence of such clinical covariates on prediction accuracy was not systematically analyzed. From a technical standpoint, although a three-level clinical assessment was performed, validating the IG maps with multiple examiners would further strengthen the results. Finally, regression to the mean remains an inherent challenge in age estimation models. Although only a mild pattern was observed in this study, predictions at the extremes of the age range should still be interpreted with caution.

Future work should address these limitations in several ways. Multisite cohorts with prospective external validation are needed to test the models on independent data from different regions, scanners, and clinical settings. In this regard, promoting open science initiatives and the curation of publicly available, large-scale benchmarks would greatly facilitate such cross-center evaluation. Future studies should focus on mitigating age dependent errors and regression to the mean effects, particularly at the extremes of the age range. To achieve this, strategies such as age aware loss reweighting, resampling for underrepresented ages, and heteroscedastic or quantile regression models combined with age stratified calibration could be explored. Approaches for domain adaptation may be investigated to handle scanner and site related distribution shifts, and the use of higher resolution or tiled inputs may be assessed to better capture fine-grained dental development cues. Furthermore, considering the biological variations in dental development across different ancestries, developing region specific models or employing transfer learning to adapt models to local populations would be particularly vital for unregistered children in specific global regions, offering a more accurate and rapidly deployable forensic tool than pursuing a single universal model. The incorporation of clinical covariates, together with fairness-oriented and stratified analyses, can also be considered to characterize and reduce potential biases. Finally, although the weighted ensemble method improved performance in this study, such schemes may be brittle under distribution shift and increase operational complexity. In this context, simpler aggregation schemes such as median ensembles should be systematically evaluated as potentially more robust alternatives.

Conclusion

The proposed median ensemble deep learning approach provides accurate, automatic, and examiner-independent pediatric age estimation from panoramic radiographs. Unlike classical methods that rely on handcrafted stages, these networks can autonomously identify age-informative regions directly from raw OPGs. The median ensemble further enhances accuracy and robustness by aggregating predictions from multiple deep learning models rather than relying on a single model. This aggregation mitigates the impact of individual model failures, providing the reliability required for safe clinical decision making. Furthermore, XAI analyses can provide visual, case level explanations allowing experts to identify cases where the model focused on less relevant areas. These properties suggest that the proposed system could serve as an interpretable decision support tool in forensic and dental workflows.

Authors’ contributions

A.D. conceived the study, developed the methodology, implemented the software, performed the visualization, validated the results, and contributed to writing, reviewing and editing. T.T.K. contributed to conceptualization, data curation, investigation, validation, and writing, reviewing and editing. All authors read and approved the final manuscript.

Funding

The authors did not receive support from any organization for the submitted work.

Data availability

The data generated during and/or analyzed during the current study are available from the corresponding author on reasonable request.

Declarations

Ethics approval and consent to participate

This study was performed in line with the principles of the Declaration of Helsinki. Approval was granted by the Karabuk University Non-interventional Clinical Research Ethics Committee (Approval No: 2025/2416, Date: 25 July 2025). Owing to the retrospective design and the use of anonymized radiographic data, informed consent was not required, as confirmed by the ethics committee.

Consent for publication

Not applicable.

Competing interests

The authors declare no competing interests.