Comparative Analysis of Third Molar Segmentation Performance Across Dental Developmental Stages and the 18-Year Age Threshold Using Deep Learning Models

Background/Objectives: Third molar development is one of the most frequently used dental indicators in forensic age assessment because its maturation continues through adolescence and early adulthood. Manual staging on panoramic radiographs requires experience and may be affected by observer variability, especially in transitional developmental stages. This study evaluated the performance of segmentation-based deep learning models for automatic third molar localization and Demirjian-based developmental stage classification on panoramic radiographs. Methods: The study used a two-stage segmentation framework. In the first stage, third molar localization was performed using 737 panoramic radiographs and 2736 annotations for teeth 18, 28, 38, and 48. In the second stage, developmental stage classification was performed using 695 panoramic radiographs and 2533 annotations grouped as AB, CD, EF, and GH according to Demirjian developmental stages. A supplementary 18-year threshold segmentation analysis was added using 695 panoramic radiographs and 2573 labels divided into training, validation, and test sets at an 80:10:10 ratio. YOLO-based segmentation models were trained using AdamW optimization and learning-rate settings recorded in the training logs. Model performance was evaluated using TP, FP, FN, precision, recall, F1-score, accuracy, mAP, Dice coefficient, Jaccard index, and performance curves. For instance segmentation, true negatives were not calculated because the number of background non-objects is not finite or clinically meaningful. Results: The third molar localization model yielded TP = 256, FP = 4, and FN = 11 on the test set. The overall accuracy was 0.9446, mAP@0.5 was 0.980, mAP@0.5:0.95 was 0.875, Dice coefficient was 0.9064, and Jaccard index was 0.8696. In developmental stage classification, yolo11x-seg produced the most balanced segmentation profile, with an accuracy of 0.7272, Dice coefficient of 0.5783, and Jaccard index of 0.5529. In the supplementary 18-year threshold segmentation analysis, yolov8x-seg produced TP = 172, FP = 37, and FN = 32. Overall accuracy was 0.8152, precision was 0.8230, recall was 0.8431, F1-score was 0.8329, mAP@0.5 was 0.635, and mAP@0.5:0.95 was 0.563. Class-wise results showed better performance for the under-18 class than for the 18-years-and-older class. Conclusions: YOLO-based segmentation models can localize third molars on panoramic radiographs with high performance. Developmental stage classification is more difficult than anatomical localization because the radiographic boundaries between adjacent developmental stages are gradual rather than discrete. The added 18-year threshold analysis provides a clinically relevant age threshold experiment, but it also shows that class imbalance and weak segmentation of the 18-years-and-older class limit direct forensic use. Segmentation-based third molar analysis should therefore be interpreted as a visually auditable decision-support workflow rather than a stand-alone legal age determination tool.