Classification of apical periodontitis with YOLOv5 and YOLOv8 architectures

Abstract

To develop a deep learning (DL) model based on convolutional neural network (CNN) architecture, for assessing the periapical health/disease of teeth, that would support decision-making during endodontic treatment. The protocol for this retrospective study was approved by the IRB. Periapical digital radiographs were collected, de-identified and scored (1-5) according to the periapical index by Ørstavik et al., 1986. A final balanced dataset of 256 images were labelled using Roboflow Annotate software. Data were pre-processed before splitting into training (68%), validation (19.5%), and testing (12.5%) groups. DL models were trained using the You Only Look Once (YOLOv5) and (YOLOv8) CNN architectures using different model structures (n, s, m, l, x). Performance metrics included Precision, Recall, and mean average precision (mAP50) of the 5 classes. Furthermore, the confusion matrix across the 5 classes and the background were used in the statistical evaluation of the results. Ørstavik D, Kerekes K, Eriksen HM. The periapical index: a scoring system for radiographic assessment of apical periodontitis. Endod Dent Traumatol. 1986 Feb;2:20-34 The YOLOv5m, YOLOv5l and YOLOv8s architectures displayed the highest accuracy amongst all trained models achieving an mAP50 of (90%). The DL algorithms used showed a high success in inference of the PAI on unseen radiographs of different origins, and its use would offer better clinician- and patient-based outcomes.