ApexTransNet: an anatomy-aware multi-task framework for apical periodontitis segmentation and radiographic assessment on periapical dental radiographs

To develop and technically evaluate an anatomy-aware multi-task framework for pixel-level apical periodontitis (AP) lesion segmentation and case-level radiographic AP assessment on periapical radiographs. ApexTransNet combines a ResNet34 encoder, Transformer-ASPP context modeling, anatomy-guided fusion, and segmentation, localization, classification, and boundary heads. The patient-disjoint cohort comprised 600 radiographs from 600 patients (training/validation/test: 420/90/90). Each of the 400 AP-positive cases had four independent lesion masks; normal controls were represented by empty masks. Segmentation experiments used three random seeds, while case-level operating thresholds were selected from the validation set and locked before test evaluation. Across seeds 42, 2026, and 3407, ApexTransNet achieved Dice 0.702 ± 0.004, IoU 0.575 ± 0.006, precision 0.802 ± 0.005, recall 0.704 ± 0.004, and boundary F1 0.263 ± 0.005 on the 60 AP-positive test cases. The frozen case-level model achieved ROC AUC 0.988 (95% CI 0.965–0.999), average precision 0.995 (95% CI 0.984–1.000), and Brier score 0.067 on the 90-case test set. ApexTransNet produced both lesion-level and case-level radiographic outputs, with improved overlap and precision relative to the evaluated baselines, at the cost of lower recall, indicating a precision-recall trade-off rather than uniform superiority. The findings represent internal technical validation and require external evaluation before clinical use. Joint case-level assessment and lesion localization may support transparent radiographic review, but the AP-enriched single-system cohort and absence of external validation preclude claims of clinical diagnostic performance.