The diagnosis of dental caries traditionally requires manual inspection of bitewing radiographs, followed by visual assessment and probing of the teeth with potential lesions. However, the use of artificial intelligence (AI), specifically deep learning, has the potential to assist in diagnosis by providing rapid and informative analysis of radiographic images.
In this study, a dataset of 13,887 bitewing radiographs from the HUNT4 Oral Health Study was used, each individually annotated by six different experts. These data were used to train three deep learning-based object detection architectures: RetinaNet (ResNet50), YOLOv5 (size M), and EfficientDet (sizes D0 and D1). For evaluation, a consensus set of 197 images, jointly annotated by the same six dentists, was used, applying a five-fold cross-validation scheme to assess the performance of the AI models.
The trained models showed an increase in average precision and F1 score, as well as a decrease in false negative rates compared to clinical dentists. In particular, the YOLOv5 model demonstrated the greatest improvement, reporting an average precision of 0.647, an average F1 score of 0.548, and an average false negative rate of 0.149. In contrast, the best human annotators in each of these metrics reported 0.299, 0.495, and 0.164, respectively.
These findings suggest that deep learning models have the potential to assist dental professionals in diagnosing caries. Nonetheless, the task remains challenging due to natural artifacts present in bitewing radiographic images.
Link: https://arxiv.org/abs/2310.00354
