A unified convolution neural network for dental caries classification

Abstract

Dental caries is one of the most common chronic diseases in the oral cavity. The early detection of initial dental caries is needed for treatment. It is problematic to diagnose the initial carious lesion, as known as enamel caries, due to the similarity of a tiny hole to human perception error. In this paper, we propose a unified convolution neural network to improve the diagnostic and treatment performance for dentists using classification from bitewing radiographs. We adapt the AlexNet and ResNet models to properly classify the dental caries dataset. The modified ResNet successfully achieves excellent binary-classification performance with accuracy of 86.67%, 87.78% and 82.78% of teeth with all conditions, teeth without dental restoration, and only teeth with dental restorations, respectively. For multilevel classification, our model has good performance with 5-class average accuracy of 80%. Remarkably, our adapted ResNet-18 has good performance with enamel caries and secondary caries with accuracy of 86.67% and 77.78%, respectively. Conversely, our ResNet-50 and ResNet-101 have contradictory low performance with enamel and secondary caries but high performance with sound teeth, dentin caries and teeth with restoration of 90%, 78.89% and 88.89%, respectively. The accuracies of our model are good enough that our model could support dentists to enhance diagnostic performance.