DOI: 10.29130/dubited.1906306 ISSN: 2148-2446

Automated Apical Periodontitis Detection on Panoramic Radiographs: Comparative Analysis of Modern Object Detectors and the Effect of Data Augmentation

Anday Duru, Fatih Karayürek
Show PDF Cite
Detecting apical periodontitis lesions on panoramic radiographs (PR) is challenging because lesions may be subtle and affected by anatomic superimposition, variable contrast, and restorative artifacts. This study compared five modern object detectors from the YOLOv8, YOLOv11, and RT-DETR families and quantified the effect of training data augmentation using a publicly available dataset. All final comparisons used a standardized input size of 1024×1024 pixels and batch size of 16. Each model–regime combination was evaluated over three independent random seeds, and performance was assessed on a fixed held-out test split comprising 393 images with 693 annotated lesions. Augmentation improved F1-score and AP-based detection performance across all architectures. Among augmented models, RT-DETR-l achieved the highest mean F1-score (0.685 ± 0.011; 95% CI: 0.657–0.714), while YOLOv11-l achieved the highest mean AP@0.50 (0.738 ± 0.018; 95% CI: 0.694–0.782) and AP@0.50:0.95 (0.352 ± 0.012; 95% CI: 0.323–0.380). The largest absolute F1 improvement was observed for YOLOv11-s (ΔF1 = +0.141). Two-sided paired comparisons with Holm correction showed significant augmentation-related F1 improvements for all five detectors. Region-wise failure analysis of augmented YOLOv11-l showed a higher false-negative rate in the maxilla (57.6%) than in the mandible (30.0%). These findings demonstrate that augmentation consistently improves PR-based apical lesion detection under repeated-seed evaluation and support the use of modern detectors as assistive second-reader tools, while final clinical decisions remain with the dentist.

More from our Archive