The study workflow can be seen in Figure 1. This study used a deep learning pipeline organized into 3 sequential stages.

First, image preprocessing was performed. In this stage, cephalometric radiographs were cropped into mandibular regions—the mandibular angle and mandibular length. Four preprocessing strategies were then applied (original, SMOTE, StandardScaler, and SMOTE+StandardScaler), combined with image augmentation.

Second, model development was conducted. Six CNN architectures (MobileNetV2, ResNet50V2, InceptionV3, InceptionResNetV2, VGG16, and VGG19) were adapted through transfer learning. Each mandibular region was processed separately, features were flattened, and the outputs merged into a single representation.

Third, multitask prediction was performed. The integrated features were used to generate 2 outputs (age estimation and sex prediction).

The data used in this study were obtained from the Department of Radiology at the Universitas Airlangga Dental and Mouth Hospital between 2019 and February 2023. A total of 340 anonymized cephalometric radiographs were collected from Indonesian individuals aged 8 to 40 years (Figure 1A). All images were standardized to a resolution of 224×224 pixels before analysis. Each radiograph was manually cropped into 2 regions of interest—the mandibular length (Figure 1B) and the mandibular angle (Figure 1C)—by the research team and subsequently validated by licensed dentists with a minimum of 5 years of clinical experience. This procedure produced 680 image samples.

To examine the impact of balancing and normalization, 4 distinct scenarios were tested (Table 1):

Image augmentation was applied to the training set using the Keras ImageDataGenerator to mitigate the small number of datasets. These augmentation techniques synthetically increased dataset diversity and helped the deep learning models learn more invariant features from the mandibular anatomy. The augmentation configuration included random rotation range up to 90 degrees, zooming in or out up to 25%, random horizontal flips and vertical flips up to 25%, and nearest neighbor interpolation for missing pixels. Image augmentation was applied only to the training set and implemented uniformly across all experiments at the image level, using identical configurations for both age estimation and sex prediction tasks. The purpose of augmentation was to increase data variability rather than to achieve exact numerical class balancing.

Six CNNs were selected for their proven utility in medical imaging and differing levels of complexity:

All models were initialized with ImageNet weights and fine-tuned. Features from both mandibular regions were extracted, flattened, concatenated, and processed through a multitask output structure.

Images were resized to 224×224 pixels and the pixel values normalized to a 0 to 1 [0,1] range and divided into training (70%), validation (15%), and testing (15%) subsets, ensuring no participant overlap between sets. We used TensorFlow and the Adam optimizer with a learning rate of 1×10^–⁴. Huber loss was applied to the age estimation output, while binary cross-entropy with label smoothing (0.05) was used for the sex prediction output. We did not use class weighting in the loss function, but we relied on SMOTE to mitigate imbalance together with label smoothing. To prevent overfitting, we applied regularization techniques such as dropout (0.5), early stopping with a patience of 10, and learning rate reduction on a plateau. Training was conducted for up to 100 epochs with a batch size of 32, with early stopping based on validation performance. Given the relatively small dataset, model training was closely monitored using validation performance to further mitigate overfitting. A unified hyperparameter configuration was applied across all architectures to ensure a controlled and fair comparative evaluation.

We examined the CNN’s performance on each task separately. For sex prediction, the metrics included accuracy, precision, and F₁-score, while age estimation was assessed using mean absolute error (MAE) in years and mean absolute percentage error (MAPE) in percent. All evaluations were conducted on the held-out test set across every preprocessing scenario and CNN architecture to ensure consistency and comparability of results.

This research used an archived dataset of cephalometric radiographs sourced from the Department of Radiology at Universitas Airlangga Dental and Mouth Hospital from March 2019 to February 2023, and the requirement for informed consent was waived by the institutional review board. No intervention or direct contact with participants occurred. The dataset remains inaccessible to the public owing to institutional data-sharing policies and considerations regarding patient privacy. All methods adhered to applicable guidelines and regulations, including the Declaration of Helsinki and institutional ethical standards. The Dental Faculty of Universitas Airlangga approved the experimental protocols (316/HERCC.FODM/III/2023). We anonymized patient records before conducting the analysis to protect confidentiality and uphold ethical guidelines.

The dataset comprised 340 cephalometric radiographs collected from Indonesian individuals aged 8 to 40 years. Each image was manually cropped to isolate 2 mandibular regions—the mandibular length (Figure 1B) and the mandibular angle (Figure 1C), producing a total of 680 image inputs. Table 2 summarizes the distribution of image samples by sex and age group, showing a 3:1 female-to-male ratio and a strong concentration of samples in the age range of 16 to 25 years. Age-group frequencies are reported at the image level (2 cropped mandibular images per individual).

Six pretrained CNN architectures (MobileNetV2, ResNet50V2, InceptionV3, InceptionResNetV2, VGG16, and VGG19) were assessed under 4 preprocessing scenarios: original, SMOTE, StandardScaler, and SMOTE+StandardScaler. The dual-input framework combined mandibular length and angle features for joint prediction of sex and age.

Model performance in age estimation was evaluated using MAE and MAPE. Table 3 presents the results across all architectures and preprocessing strategies.

VGG16 achieved the lowest error rates, particularly in the original scenario (MAE=3.19 years; MAPE=13.19%), establishing a strong baseline. VGG19 also demonstrated robust performance across most scenarios. In contrast, InceptionV3 and InceptionResNetV2 consistently achieved higher errors, particularly under the SMOTE scenario, suggesting that synthetic oversampling might introduce noise detrimental to these architectures for regression. The StandardScaler and SMOTE+StandardScaler scenarios generally improved stability, reducing the performance gap between models and helping VGG19 achieve its best MAPE (14.35%). These results for MAE variation across models and preprocessing strategies are visualized in Figure 2.

Model performance on sex prediction was evaluated using accuracy, macro F₁-score, weighted F₁-score, and class-wise F₁-scores for female and male categories. Table 4 summarizes results across all architectures and preprocessing scenarios.

Table 4 shows VGG16 and VGG19 delivered the most accurate and balanced performance. VGG16 achieved the highest stand-alone accuracy of 86% under the StandardScaler scenario, with a male F₁-score of 63%. The application of SMOTE consistently improved the male F₁-score across nearly all models, for instance, raising ResNet50V2’s male F₁-score from 14% to 64%, confirming its efficacy in mitigating class imbalance. However, models such as InceptionV3 and ResNet50V2 exhibited high sensitivity to preprocessing, with performance fluctuating significantly across scenarios. Across the evaluated preprocessing scenarios, all CNN architectures demonstrated a male F₁-score above 33% in at least 1 scenario, except for InceptionV3 under the original (unbalanced) condition. The comparative accuracy trends are shown in Figure 3.

This study demonstrates that deep learning models, particularly VGG16 and VGG19, can effectively perform joint age estimation and sex prediction from cropped mandibular cephalometric images. The findings emphasize that performance is not solely determined by architectural design but critically depends on preprocessing strategies to address dataset limitations. Synthetic oversampling (ie, SMOTE) was essential in mitigating severe class imbalance and improving fairness in male sex prediction, while the use of accuracy and F₁-score provided complementary insights into classifier behavior under imbalance [19]. Accuracy offers a straightforward measure of correctness, yet can be misleading when 1 class dominates, whereas the F₁-score, by accounting for both precision and recall, provides a more reliable evaluation in such contexts [20,21]. These results emphasize the importance of integrating robust architectures with targeted preprocessing to achieve equitable and reproducible outcomes in forensic odontology.

The preprocessing step of cropping cephalometric images to the mandibular region provides substantial advantages. For example, by removing irrelevant anatomical structures such as the cranial vault and maxilla, background noise is reduced, and the models are directed to focus on the most discriminative features. This cropping approach improves computational efficiency by lowering input dimensionality and enhances interpretability, as the mandible is widely recognized as one of the most sexually dimorphic bones in the craniofacial complex.

Several studies support the relevance of mandibular-focused analysis. A study by Prabha et al [22] demonstrated that mandibular indices derived from lateral cephalograms are highly effective for sex determination, highlighting the forensic importance of isolating mandibular features. Similarly, Gao and Tang [23] showed that deep learning frameworks integrating cephalometric landmarks achieve greater accuracy when attention mechanisms prioritize mandibular regions, as these structures carry distinctive morphological cues critical for demographic prediction. In forensic odontology, the mandible is also considered more resistant to postmortem changes compared with other cranial structures, making it a reliable target for identification.

Cropping images enables more precise landmark detection and reduces interobserver variability from a computational perspective [24,25]. Preprocessing strategies such as region-specific cropping have been shown to improve model precision and generalization, particularly when combined with data-balancing techniques such as SMOTE. Clinically, mandibular length and angle are key determinants in orthodontic diagnosis and maxillofacial treatment planning, reinforcing the dual relevance of cropping for both forensic and medical applications.

The performance analysis of age estimation highlights the importance of selecting appropriate preprocessing strategies to optimize deep learning models in demographic prediction. While the study was not designed to forecast age within defined intervals, the use of MAE and MAPE as evaluation metrics provides a robust framework for assessing predictive accuracy. MAE serves as a metric used by various recommendation systems to measure the difference between user ratings and predicted scores [26,27]. However, a widely recognized accuracy metric across various fields, often referenced in scholarly articles, is the MAPE [28,29].

The comparative outcomes across different preprocessing scenarios suggest that data balancing and normalization exert distinct influences on model behavior. Oversampling techniques such as SMOTE may introduce synthetic variability that benefits certain architectures but disrupts others, reflecting findings in prior work where oversampling occasionally degraded model generalization in medical imaging tasks [30,31]. Conversely, normalization through StandardScaler consistently improved model generalization, aligning with evidence that standardized input distributions enhance convergence and stability in CNNs [32].

This study is consistent with prior work using VGG16 for age estimation from cervical vertebrae images, which reported an MAE of 3.53 years and an average MAPE of 16.36% in the original (unbalanced) scenario [33]. These results indicate that, on average, the predicted age deviated by approximately 3.5 years from the true chronological age, supporting the reliability of deep learning–based age estimation in craniofacial imaging [33].

In addition to preprocessing effects, the age distribution of the dataset was uneven, with a strong concentration of samples in the range of 16 to 25 years. This imbalance may have influenced age estimation performance, as models tend to achieve lower prediction errors in age groups that are more frequently represented during training, while performance for underrepresented age ranges may be less stable. Similar effects of age imbalance on regression-based age prediction tasks have been reported in prior studies, highlighting the importance of age-stratified sampling or regression-aware balancing strategies in future work [34,35].

The performance analysis of sex prediction highlights the persistent challenge of class imbalance, particularly in male prediction, despite strong overall accuracies achieved by deep learning architectures. This imbalance is consistent with prior literature, where female features are often more consistently represented in datasets, leading to biased learning outcomes. Franco et al [36] demonstrated that transfer learning approaches outperform models trained from scratch in dental radiograph classification, underscoring the importance of pretrained architectures, such as VGG16 and VGG19, in capturing subtle morphological differences. These findings emphasize that while high accuracy is achievable, equitable performance across sexes remains a methodological priority in forensic odontology.

Oversampling techniques such as SMOTE proved effective in mitigating imbalance by generating synthetic samples for minority classes. Elreedy et al [37] provided a comprehensive analysis of SMOTE, confirming its utility in addressing class imbalance across diverse domains [38]. More recent refinements, including abnormal minority handling and Outlier-SMOTE, demonstrate that oversampling can be adapted to improve generalization in sensitive datasets [39,40]. In forensic sex prediction, these approaches are particularly relevant, as they provide more representative training views and reduce bias in male prediction. Furthermore, advanced variants such as MeanRadius-SMOTE have shown superior reliability compared with conventional SMOTE and LR-SMOTE, achieving better predictive accuracy across both majority and minority classes [41]. Collectively, these studies reinforce that oversampling is a critical intervention, though its effectiveness remains architecture-dependent.

Normalization techniques also contributed to improved accuracy, particularly in complex architectures, by ensuring equal feature contributions and reducing the risk of dominant variables overshadowing relevant patterns. Practical guidelines such as those outlined by Brownlee [41] highlight the role of StandardScaler and Normalizer in stabilizing training, while empirical studies confirm their impact on supervised classification accuracy [42,43]. The broader generalizability of normalization has demonstrated significant performance gains in electricity consumption forecasting, underscoring its universal relevance across domains [44]. Nonetheless, normalization alone does not fully resolve sex prediction disparities, highlighting the need for targeted interventions that combine preprocessing with architectural optimization. Large-scale surveys of deep learning in medical imaging further emphasize that preprocessing and model design must be jointly considered to achieve equitable performance in forensic applications [34,35].

AI-assisted forensic odontology underscores the mandible as a resilient anatomical marker for sex prediction and age estimation when other skeletal elements are unavailable [14,45]. Consistent with Abdelhamid and Desai [19], our findings confirm that synthetic oversampling strategies, such as SMOTE, can effectively mitigate data imbalance and improve prediction robustness in limited radiographic datasets. This study also complements the work of Matsuda et al [45], who demonstrated that multitask deep learning frameworks improve learning efficiency and generalization across medical imaging tasks.

Within forensic practice, these results emphasize the mandibular-focused, multitask CNN framework as a practical tool for postmortem identification and disaster victim assessment. By integrating data-balancing and normalization techniques, the proposed approach enhances interpretability, reproducibility, and scalability, paving the way for broader AI applications in forensic odontology and demographic estimation.

This study demonstrates that cropped mandibular regions, particularly the mandibular length and angle, are reliable anatomical indicators for demographic prediction in forensic contexts. Among the CNN architectures evaluated, VGG16 and VGG19 consistently achieved superior accuracy and balanced sex prediction, confirming their suitability for forensic applications. MobileNetV2 benefited from oversampling strategies, while ResNet50V2 and InceptionV3 showed limited performance in male prediction, indicating the need for further refinement. The integration of robust CNN models with mandibular image analysis provides a scalable pathway for automated forensic identification, especially in disaster scenarios and resource-limited settings.

This study has several limitations that should be considered when interpreting the findings. First, the dataset was relatively small (680 images from 340 participants) and drawn from a single Indonesian population. This may limit the generalizability of the results to other ethnicities, age ranges, or geographic settings. Moreover, this constraint may increase the risk of overfitting in deep learning models. Future studies should consider using larger, multicenter datasets to enhance robustness and applicability. Second, the pronounced class imbalance, with a 3:1 female-to-male ratio, also influenced the model performance. In addition to sex imbalance, the age distribution was also uneven, with a strong concentration of samples in the age range of 16 to 25 years. This imbalance may have influenced age estimation performance across models and should be addressed in future studies using age-stratified sampling or regression-aware balancing strategies. Third, there are limitations related to methodological and practical considerations. Manual cropping of mandibular regions, despite clinical validation, introduces a degree of subjectivity that may affect reproducibility; automated landmark detection or segmentation methods could address this in future studies.

Fourth, the focus on only 2 mandibular parameters (length and angle) excludes other potentially informative craniofacial and dental features. Fifth, although VGG16 and VGG19 produced the strongest results, their higher computational demands may limit applicability in time-sensitive forensic workflows. Conversely, lightweight models such as MobileNetV2 offer greater efficiency but at reduced precision. Finally, the models were not evaluated under noisy or degraded imaging conditions common in postmortem or disaster settings, warranting future work on model robustness and real-life applications.