Comparative evaluation of YOLO models for classifying maxillary sinus and posterior tooth relationships on panoramic radiographs with CBCT validation
SCIENTIFIC REPORTS, cilt.1, sa.1, ss.1, 2026 (SCI-Expanded, Scopus)
- Yayın Türü: Makale / Tam Makale
- Cilt numarası: 1 Sayı: 1
- Basım Tarihi: 2026
- Doi Numarası: 10.1038/s41598-026-77298-8
- Dergi Adı: SCIENTIFIC REPORTS
- Derginin Tarandığı İndeksler: Academic Search Ultimate (EBSCO), Natural Science Collection (ProQuest), Biological Science Database (ProQuest), Biomedical Reference Collection: Corporate Edition (EBSCO), Health Research Premium Collection (ProQuest), Scopus, Science Citation Index Expanded (SCI-EXPANDED), BIOSIS, Chemical Abstracts Core, EMBASE, MEDLINE, Directory of Open Access Journals, Zoological Record
- Sayfa Sayıları: ss.1
- Karadeniz Teknik Üniversitesi Adresli: Evet
Özet
To develop and internally evaluate You Only Look Once (YOLO) deep learning models for automated classification of the anatomical relationship between the maxillary sinus floor and maxillary posterior tooth roots on panoramic radiographs, using cone-beam computed tomography (CBCT) as the reference standard. In this retrospective single-center study, 4,500 CBCT examinations acquired between 2023 and 2025 were screened. A total of 462 patients with paired panoramic radiographs and CBCT scans met the eligibility criteria, yielding 3,691 maxillary posterior tooth regions for analysis. Anatomical relationships were classified using CBCT as Type 1 (non-contact), Type 2 (cortical contact), or Type 3 (intrasinus penetration). The dataset was partitioned at the patient level into training (n = 370), validation (n = 46), and test (n = 46) cohorts. Model stability was assessed using 10-fold cross-validation within the training cohort. Final performance was evaluated on the held-out test set using precision, recall, F1-score, and mean average precision (mAP) at intersection over union (IoU) thresholds of 0.50 (mAP@0.5) and 0.50–0.95 (mAP@0.5:0.95), with bootstrap-derived 95% confidence intervals (CIs). YOLO26 yielded the highest observed test-set performance, with a precision of 0.94 (95% CI: 0.91–0.97), recall of 0.96 (95% CI: 0.93–0.98), F1-score of 0.95 (95% CI: 0.92–0.97), mAP@0.5 of 0.95 (95% CI: 0.92–0.97), and mAP@0.5:0.95 of 0.71 (95% CI: 0.67–0.75). Corresponding mAP@0.5 values were 0.91 for YOLOv11 and 0.90 for YOLOv12. Type 2 relationships represented the most challenging category across models, particularly at the boundary between cortical contact and intrasinus penetration. Within this single-center retrospective dataset, YOLO26 showed the highest internal test performance among the evaluated models for predicting CBCT-derived maxillary sinus–posterior tooth relationship categories from panoramic radiographs. These findings represent internal validation only and do not establish generalizable clinical performance. CBCT remains the reference standard for three-dimensional assessment. External multicenter validation is required before this approach can be considered for clinical use.