Development of a deep learning based prototype artificial intelligence system for the detection of dental caries in children

Abstract

Objective: To develop a prototype artificial intelligence image recognition system for detecting dental caries, especially those without cavities, in children. Methods: Seven hundred and twelve intraoral photos, which were taken by dental professionals using a digital camera from October 2013 to June 2020 in the Department of Pediatric Dentistry, Peking University School and Hospital of Stomatology, were collected from the children who received dental treatment under general anesthesia. The well-documented post-treatment electronic dental record of each child was identified as label standard to determine whether the teeth were carious and the type of caries types such as caries that had become cavities (caries with cavities), pit and fissure caries that had not become cavities (pit and fissure caries) and proximal caries which the marginal ridge enamel had not been destroyed (proximal caries). The various teeth and caries types were labeled by pediatric dentists using VoTT software (Windows 2.1.0, Microsoft, U S A). There were five labeled groups: pit and fissure caries, approximal caries, non-carious approximal surfaces, caries with cavities and teeth without caries (including intact fillings). Each group was randomly divided into training dataset, validation dataset and test dataset at a ratio of 6.4∶1.6∶2.0 by using random number table. After using the labeled training dataset for deep learning training, a deep learning-based artificial intelligence (AI) image recognition system for detecting dental caries was established, with the caries probability greater than 50.0% as the criterion for determining caries. Sensitivity and accuracy were used as indicators of recognition specificity. Results: Seven hundred and twelve single-jaw intraoral photographs were segmented and annotated into 953 pit and fissure caries, 1 002 approximal caries, 3 008 caries with cavities, 3 189 teeth without caries and 862 non-carious approximal surfaces, totaly 9 014 labels. The sensitivities and specificities of the test set were 96.0% and 97.0% for caries with cavities, 95.8% and 99.0% for pit and fissure caries and 88.1% and 97.1% for approximal caries. Conclusions: The current AI system developed based on deep learning of the intra-oral photos in the present study showed the ability to detect dental caries. Furthermore, the AI system could accurately verify different types of dental caries such as caries with cavities, pit and fissure caries and proximal caries.