Automated Forensic Age Estimation through Deep Learning Staging of Third Molar Maturation: A Cross-Sectional Study in a Pondicherry Population

Aim: To develop and validate an automated deep learning pipeline for third-molar developmental staging using Demirjian’s eight-stage classification system and to estimate population-specific probabilities for legally relevant age thresholds.

Materials and Methods: A retrospective study was conducted using 785 digital panoramic radiographs from individuals aged 12–22 years in Puducherry, India. Mandibular third molars were manually staged according to Demirjian’s classification to generate reference annotations. A three-stage deep learning pipeline comprising YOLO for region-of-interest detection, U-Net for tooth segmentation, and a transfer-learning DenseNet121 convolutional neural network for developmental-stage classification was implemented. Model performance was evaluated using five-fold cross-validation. Chronological age estimates were derived from predicted developmental stages using  stage-to-age conversion method. Mean absolute error was calculated from held-out predictions using the absolute difference between predicted and observed chronological age. Legal-age probabilities were estimated separately by sex and developmental stage for thresholds of 16, 18, and 21 years. Ninety-five percent confidence intervals were calculated using [exact binomial/Wilson/bootstrap] methods.

Results: The automated model achieved an overall staging accuracy of 89.7%, and Stage H demonstrated the highest diagnostic performance (AUC = 0.96). The mean absolute error for chronological age estimation was 0.62 years. The analytical sample included males and females. Stage G corresponded to an age of ≥18 years in 69.2% of males; 95% CI: and 82.0% of females; 95% CI. All observed Stage H cases were aged ≥18 years, regardless of sex. Results for the ≥16-year and ≥21-year thresholds were [insert probabilities, subgroup sizes, and 95% confidence intervals].

Conclusion: Automated deep learning-based third-molar staging demonstrated good classification performance and may improve the reproducibility of forensic dental age assessment. The probabilistic framework provides population-specific estimates relevant to legal age thresholds, particularly the 18-year threshold. However, these probabilities should be interpreted as sample-based estimates rather than definitive proof of chronological age.