Comparative diagnostic accuracy of a You Only Look Once-11s deep learning model and human observers for detection of periapical lesions on intraoral periapical radiographs.

Abstract

Context: Interpretation of intraoral periapical radiographs (IOPAs) is influenced by image quality and observer-related factors. This study evaluated a lightweight You Only Look Once (YOLO)-11s model for automated detection of periapical lesions and compared its output with that of a human observer.

Aim: The study aimed to assess and compare the accuracy and diagnostic performance of YOLO-11s and human observers for detecting periapical lesions on IOPAs.

Materials and methods: Radiographic records from 500 patients were reviewed retrospectively, yielding 1330 annotated tooth-level images. The dataset was divided into training (80%), validation (10%), and test (10%) subsets. Images were normalized, contrast-enhanced, and resized to 640 × 640 pixels. YOLO-11s was trained for up to 200 epochs. Performance was assessed using precision, sensitivity (recall), F1-score, mean average precision (mAP), confusion matrix analysis, precision-recall curves, and inference time. Cohen's kappa was used to assess interobserver agreement.

Results: The YOLO-11s model demonstrated high diagnostic performance for periapical lesion detection, with a sensitivity of 89.4%, precision of 88.0%, and an F1-score of 88.4%. Out of 133 true tooth-level targets, 121 were detected, and 106 were correctly predicted as the periapical lesion (PLA) or healthy (PRA), giving an overall accuracy of 79.7%. The mAP50 and mAP50-95 values were 0.85 and 0.48, respectively, indicating good detection ability but moderate localization precision. However, no background region was correctly identified.

Conclusion: The YOLO-11s model demonstrated high sensitivity for automated detection of periapical lesions on IOPAs and may serve as a valuable adjunctive diagnostic tool in endodontic practice.