Boundary-aware deep learning for automated segmentation of periodontal anatomical structures on panoramic radiographs

Accurate identification of periodontal anatomical structures on panoramic radiographs is a prerequisite for quantitative assessment of periodontal bone loss. However, automated segmentation of the alveolar bone level (BL), cementoenamel junction (CEJ), and root apex (RA) remains challenging because of complex morphology, low contrast, and indistinct boundaries. We developed and evaluated a boundary-aware deep learning framework for automated segmentation of these structures. A retrospective dataset of 311 panoramic radiographs from 311 patients was divided at the patient level into training (n = 217), validation (n = 49), and independent test (n = 45) sets. Expert-reviewed annotations were generated for BL, CEJ, and RA. Structure-specific models were trained using a LinkNet-style encoder–decoder architecture with a Boundary-Aware Module (BAM). For the final BAM-enabled models, image-level precision, recall, F1-score, and intersection over union (IoU) were calculated from raw predicted masks and averaged across 45 images; 95% confidence intervals (CIs) were estimated by patient-level bootstrap resampling. Models with and without BAM were descriptively compared using pooled metrics across the three structures. BL performed best, with an F1-score of 0.940 (95% CI, 0.919–0.955) and an IoU of 0.892 (95% CI, 0.863–0.914), followed by CEJ, with an F1-score of 0.854 (95% CI, 0.833–0.874) and an IoU of 0.752 (95% CI, 0.722–0.779). RA was more challenging, with an F1-score of 0.519 (95% CI, 0.476–0.557) and an IoU of 0.361 (95% CI, 0.326–0.393). In the pooled ablation analysis, models with BAM achieved higher precision (0.753 vs. 0.702), F1-score (0.846 vs. 0.822), and IoU (0.734 vs. 0.698), but lower recall (0.967 vs. 0.992), than models without BAM. The framework demonstrated the feasibility of automated periodontal structure segmentation on panoramic radiographs, with better performance for BL and CEJ than for RA. Pooled ablation results suggest a potential contribution of boundary-aware feature integration to overall segmentation performance. The framework provides a technical foundation for quantitative radiographic analysis; however, external validation, tooth-level landmark localization, and direct comparison with expert-derived bone loss measurements are required before clinical application.