Automated detection of periapical lesions in pediatric panoramic radiographs using YOLOv7: a retrospective internal validation study

Abstract

This study aimed to assess the diagnostic capability of a YOLOv7 deep learning algorithm for the computerized detection of periapical lesions from pediatric panoramic radiographs. Its potential utility as a supportive diagnostic tool and accurate diagnosis in mixed dentition cases was further assessed by comparing the algorithm’s performance with the diagnoses made by dental students. In this study, a total of 408 panoramic radiographs were used, consisting of 333 original images and 75 images generated through feature-based preprocessing expansion. The YOLOv7 model was trained on 302 images, which included 227 original radiographs and 75 images specifically enhanced via grayscale conversion, noise reduction, and edge detection filters to emphasize structural pathological features. A relatively larger set was allocated for testing in order to enhance robustness despite the small sample size. The diagnostic capability of the algorithm and trainees was compared using accuracy, sensitivity, specificity, precision, F1 score, and error rate. YOLOv7 achieved higher diagnostic performance compared with the student group. Its sensitivity (76.1%) was also higher than that of the students (55.2%). The algorithm further demonstrated superior specificity (99.8% vs. 97.3%), precision (99.8% vs. 95.3%), and F1 score (86.4% vs. 69.9%). The findings suggest promising potential in the YOLOv7 algorithm’s performance for detecting periapical lesions in deciduous teeth on panoramic radiographs, compared with the diagnostic accuracy of the students.