A retrospective radiological investigation of 1,857 cone-beam computed tomography scans of permanent first molars from Han Chinese adults has demonstrated that combining three-dimensional pulp chamber volume with machine learning algorithms can estimate chronological age with a mean absolute error of 3.14 years. Published in the Egyptian Journal of Forensic Sciences by researchers from North Sichuan Medical College, the study evaluated adults aged 18 to 65 years to establish whether non-linear computational models could overcome the precision limits of classical forensic dentistry. By analyzing three-dimensional pulp chamber volumes, the research team demonstrated that sex- and position-specific extreme gradient boosting models achieved a coefficient of determination of 0.87, reflecting a 36.5 percent error reduction compared with the best traditional cubic regression model. The findings establish that dental pulp regression, driven by continuous secondary dentin apposition, provides a reliable volumetric index capable of yielding high-precision age predictions for judicial, humanitarian, and forensic identification contexts.
Forensic age estimation in deceased individuals and living subjects has historically depended on skeletal development markers, such as epiphyseal fusion, medial clavicular ossification, and cranial suture closure. However, skeletal assessments are frequently compromised by taphonomic degradation, mechanical fragmentation, and individual biological variation, which together generate estimation errors averaging five years or more in compromised remains. In contrast, teeth offer superior taphonomic stability, displaying high resistance to environmental decomposition, physical trauma, and post-mortem alteration. Furthermore, dental development remains relatively independent from external factors. Within the adult dentition, the first molar represents an ideal target because it is the earliest permanent tooth to erupt, emerging into the oral cavity around six years of age. Having the longest functional duration in the human mouth, the first molar exhibits continuous secondary dentin deposition along the inner pulp cavity walls, producing progressive volumetric reduction that correlates strongly with chronological age while avoiding interference from occlusal wear.
To ensure morphological consistency and eliminate external pathological confounders, the researchers enforced strict inclusion criteria across an extensive single-center archival database. Approved by the Institutional Review Board of North Sichuan Medical College (Approval No. 2023006) with an informed consent waiver for pseudonymized archival data, the study reviewed cone-beam computed tomography scans acquired between December 2020 and June 2023 at the Affiliated Hospital of North Sichuan Medical College. The final cohort comprised 1,857 Han Chinese adults aged 18 to 65 years, including 883 males and 974 females. Scans were required to capture the full maxilla and mandible with complete visualization of the right first molars, designated as maxillary tooth 16 and mandibular tooth 46 under Fédération Dentaire Internationale notation. Every evaluated tooth was confirmed free of caries, periapical lesions, restorative treatments, endodontic therapy, and occlusal wear extending into the dentin layer, ensuring that all measured volumes represented physiologically normal pulp morphology.
Imaging data were acquired using an OP 3D Vision cone-beam computed tomography system from KaVo Dental GmbH operating at 120 kilovolts, 5 milliamperes, an exposure time of approximately 4 seconds, and a 0.2-millimeter slice thickness. Chronological age was calculated from documented birth and scan dates. Digital Imaging and Communications in Medicine files were processed in Mimics software version 21.0 following a standardized protocol combining semi-automatic segmentation with mandatory manual editing. Segmentation isolated the pulp chamber through Hounsfield Unit thresholding set between 200 and 1200 HU to separate low-density pulp tissue from higher-density dentin. A region-growing algorithm was initialized by placing seed points at the geometric center of the pulp cavity simultaneously across axial, coronal, and sagittal planes. When initial segmentation incompletely captured narrow isthmuses or pulp horns, additional seed points were placed in adjacent slices until the entire pulp cavity was completely extracted.
Because partial volume effects and similar radiodensities along the pulp-dentin junction create segmentation challenges, initial masks underwent systematic manual refinement in all three anatomical planes. Investigators utilized Edit Masks for voxel-level adjustments and Multiple Slice Edit to correct discontinuities across consecutive tomographic slices. To maintain quality control and measurement consistency, two independent operators evaluated all segmentations. Any case exhibiting a volumetric discrepancy exceeding 5 percent between operators was re-evaluated until consensus was reached, and a random 10 percent sample underwent repeated assessment to determine intra-observer reliability. Intraclass correlation coefficients confirmed excellent measurement stability, yielding an inter-observer coefficient of 0.84 and an intra-observer coefficient of 0.89, both significantly greater than zero (p < 0.05). Following manual verification, Mimics executed its Calculate Part function to generate watertight three-dimensional surface models, automatically recording pulp chamber volume in cubic millimeters for subsequent statistical modeling.
Statistical analysis revealed significant differences in pulp chamber volume across anatomical positions and between sexes (p < 0.05). Males exhibited significantly larger mean pulp chamber volumes than females, and maxillary right first molars displayed greater internal volume than mandibular first molars. These dimensional disparities reflect known developmental differences, masticatory load patterns, and sexual dimorphism during odontogenesis, where female dental development begins earlier and sex chromosomes influence mitotic activity in the dental papilla. Furthermore, estrogen-mediated inhibition of odontoclasts and X-chromosome stabilization of dentin mineralization may explain the enhanced estimation precision observed in female samples. Pearson correlation analyses demonstrated strong, statistically significant negative correlations between pulp chamber volume and chronological age across all groups, with correlation coefficients ranging from r = -0.81 to r = -0.88 (all p < 0.05). The strongest negative correlation occurred in female maxillary first molars (r = -0.88), confirming that volumetric pulp reduction accurately tracks advancing chronological age.
To evaluate traditional mathematical approaches, the researchers fitted eleven curvilinear regression models across training subgroups. The full dataset of 1,857 individuals was partitioned into an 80 percent training set (1,485 subjects) and a 20 percent independent test set (372 subjects) using stratified sampling based on age and sex. To avoid sample-size bias during regression development, equal random subsamples of 706 males and 706 females were drawn from the training data. Evaluated models included linear, logarithmic, inverse, quadratic, cubic, compound, power, S-curve, growth, exponential, and logistic curves. The cubic regression model achieved the highest coefficient of determination across all subgroups, with R² values ranging from 0.77 to 0.79. Among these cubic formulations, the female-specific model utilizing maxillary tooth 16 achieved the lowest mean absolute error at 4.95 years, whereas the sex-independent model utilizing combined molar volumes yielded an error of 7.43 years, highlighting the limitations of rigid polynomial curves.
To address non-linear morphological relationships and overcome the constraints of classical equations, the research team developed six machine learning algorithms in Python 3.11.6 using scikit-learn 1.3.0 and XGBoost 1.7.0 libraries. The investigated algorithms included Linear Regression, Lasso Regression, Neural Networks, Random Forest, Gradient Boosting, and extreme gradient boosting (XGBoost). Hyperparameter tuning was conducted through grid search with 5-fold cross-validation exclusively on the training set. Key strategies to prevent overfitting included L2 regularization, dropout layers in neural networks, bagging and boosting ensembles with a 0.1 learning rate, and 50-round early stopping. When evaluated on the held-out test set, position-specific models consistently outperformed position-agnostic models. In the sex-independent sample, position-specific XGBoost achieved a mean absolute error of 3.46 years (R² = 0.84) compared to 4.70 years (R² = 0.74) for position-agnostic Gradient Boosting. In males, position-specific XGBoost attained an error of 3.27 years (R² = 0.83).
The optimal predictive performance was attained by the female-specific, position-specific XGBoost model on maxillary first molars, which achieved a mean absolute error of 3.14 years (95% CI: 2.92–3.37) and an R² of 0.87 (95% CI: 0.84–0.89). Bonferroni-corrected paired t-tests confirmed that this result was statistically superior to the best cubic regression model (MAE difference = 1.81 years, p < 0.001), representing a 36.5 percent error reduction. This level of accuracy is critical in judicial and humanitarian contexts, including victim identification in mass disasters and age verification for criminal responsibility or unaccompanied minors. Because legal thresholds often hinge on specific chronological ages, such as 14 or 18 years, reducing estimation error prevents catastrophic misclassification. An estimated age of 17.5 years with a 95 percent prediction interval of ±3 years spans from 14.5 to 20.5 years, demonstrating why forensic practitioners must report prediction intervals and combine dental findings with skeletal indicators.
The authors acknowledged several important limitations that define the scope of their findings. The study cohort was drawn exclusively from Han Chinese individuals at a single hospital in Sichuan Province, necessitating future cross-population validation before applying these models across other ancestries. Furthermore, physiological variables influencing secondary dentin deposition, such as nocturnal bruxism, parafunctional habits, and minor occlusal wear, were not clinically quantified using formal wear indices. Demographically, adults aged 58 to 65 years were substantially underrepresented, comprising only 2.3 percent of the sample (106 individuals, including 42 males), which limits predictive confidence in older adults. Finally, the reliance on semi-automatic segmentation remains time-consuming and operator-dependent. The authors recommended integrating fully automated deep learning networks, such as U-Net or Mask R-CNN, and combining pulp chamber volume with root morphology and enamel thickness to develop robust, high-throughput multimodal age estimation frameworks for routine forensic casework.
Subject of Research: Adult dental age estimation using three-dimensional molar pulp chamber volume and machine learning
Article Title: Adult age estimation from first molar pulp chamber volume in a Han Chinese population using a machine learning approach
Article References: Ding, Y., Zhong, T., He, Y., Wang, W., Zhang, S., Zhang, X., Shi, W., & Jin, B. (2026). Adult age estimation from first molar pulp chamber volume in a Han Chinese population using a machine learning approach. Egyptian Journal of Forensic Sciences, 16(1), Article 50. https://doi.org/10.1186/s41935-026-00546-5
Image Credits: AI Generated
DOI: 10.1186/s41935-026-00546-5
Keywords: dental age estimation, pulp chamber volume, cone-beam computed tomography, first molar, XGBoost, Han Chinese population, forensic identification, secondary dentin deposition, cubic regression, three-dimensional segmentation, tooth 16, tooth 46, sexual dimorphism, mean absolute error, North Sichuan Medical College, Mimics software, machine learning algorithms
News Source: Teresa Odom. (November 13, 2026). Machine Learning on Molar Pulp Volume Estimates Adult Age With 3.14-Year Error. Scienmag.



