This study aimed to develop and compare two YOLOv12-based deep learning models-object detection and pose estimation-for the automatic classification of orthodontically induced external apical root resorption (OIEARR) using panoramic radiographs. A total of 624 panoramic radiographs obtained from 312 patients aged 10-18 who underwent at least 12 months of fixed orthodontic treatment were retrospectively analyzed. Each maxillary central and lateral incisor was graded for OIEARR severity on a 4-point scale (Grade 0 to Grade 3) by two experienced orthodontists serving as the ground truth. Two YOLOv12-based models were trained: an object detection (OD) model for regional analysis and a pose estimation (PE) model for anatomical landmark localization. Both models were trained and validated on annotated panoramic images and evaluated using accuracy, precision, recall, specificity, F1-score, confusion matrix, and ROC-AUC. The PE model outperformed the OD model across all evaluation metrics, demonstrating superior performance in detecting OIEARR. Specifically, the PE model achieved a weighted F1-score of 0.88, compared to 0.60 for the OD model. It also showed higher accuracy (0.93 vs. 0.78), precision (0.88 vs. 0.64), and recall (0.88 vs. 0.59), confirming its robustness in root resorption classification. Particularly in Grade 1 and Grade 2 resorption categories, the PE model demonstrated markedly superior classification performance (F1 = 0.85 and 0.88, respectively), while maintaining excellent detection in Grade 3 cases (F1 = 0.95). Confusion matrix analysis revealed that most misclassifications occurred between neighboring grades. ROC-AUC values for the PE model were consistently high (0.90-0.99), indicating strong discriminative ability across all resorption stages. The YOLOv12x PE model offers a reliable and sensitive tool for detecting varying degrees of root resorption on panoramic radiographs. Its fine-grained anatomical localization capabilities provide an advantage for early diagnosis, making it a promising approach for clinical decision support in orthodontics.

BackgroundThe aim of the study was to evaluate automated digital diagnostic setup in bimaxillary dentoalveolar protrusion cases using two software packages and to compare them to manual digital setup.MethodologyPre-treatment intraoral scans of 14 patients whose treatment plans involved extraction of four first premolars were imported as Standard Tessellation Language files into dentOne® software (DIORCO co. ltd, Yongin, South Korea) and Ortho Simulation software (MEDIT Corp, Seoul, South Korea). Following tooth segmentation and selection of the teeth to be extracted, an automatic virtual setup was performed in each software. Moreover, manual virtual setups were performed by an orthodontist using dentOne® software. Dental arch changes and dental movements and the duration taken to perform the setups were evaluated and compared using the appropriate statistical tests.ResultsThe inter-canine, inter-premolar and inter-molar widths did not change significantly following manual virtual setup, while the arch length significantly decreased. The inter-premolar width, inter-molar width and arch length significantly decreased following both automated setups. The manual setup showed significantly greater lingual translation of maxillary and mandibular incisors compared to Ortho Simulation software (mean difference = 5.97 ± 1.10 mm and 7.02 ± 1.29 mm, respectively) and dentOne software (mean difference = 5.73 ± 0.96 mm and 6.95 ± 1.26 mm, respectively). The mesial translation of the maxillary and mandibular molars in Ortho simulation setup (8.35 ± 1.62 mm and 8.69 ± 1.91 mm, respectively) and dentOne setup (7.41 ± 1.28 mm and 7.74 ± 1.90 mm, respectively) was statistically significantly higher than that obtained using the manual setup (− 0.08 ± 0.27 mm, 0.03 ± 0.47 mm, respectively). All setups showed clinically significant lingual inclination of maxillary and mandibular incisors, with the manual setup exhibiting more lingual inclination than both automated setups. Ortho Simulation setup was the fastest method (4.14 ± 0.53 min), followed by dentOne automated setups (7.57 ± 0.94 min), then the manual setup (21.00 ± 1.66 min).ConclusionDespite being faster, the automated diagnostic setups for bimaxillary protrusion cases constricted the dental arch and did not manage the extraction spaces well, hence, simulating anchorage loss. These findings highlight the need for manual refinement of the automated setups.Supplementary InformationThe online version contains supplementary material available at 10.1186/s40510-026-00605-6.