Development and evaluation of deep learning for screening dental caries from oral photographs.

OBJECTIVES: To develop and evaluate the performance of a deep learning system based on convolutional neural network (ConvNet) to detect dental caries from oral photographs. METHODS: 3,932 oral photographs obtained from 625 volunteers with consumer cameras were included for the development and evaluation of the model. A deep ConvNet was developed by adapting from Single Shot MultiBox Detector. The hard negative mining algorithm was applied to automatically train the model. The model was evaluated for: (i) classification accuracy for telling the existence of dental caries from a photograph and (ii) localization accuracy for locations of predicted dental caries. RESULTS: The system exhibited a classification area under the curve (AUC) of 85.65% (95% confidence interval: 82.48% to 88.71%). The model also achieved an image-wise sensitivity of 81.90%, and a box-wise sensitivity of 64.60% at a high-sensitivity operating point. The hard negative mining algorithm significantly boosted both classification (p < .001) and localization (p < .001) performance of the model by reducing false-positive predictions. CONCLUSIONS: The deep learning model is promising to detect dental caries on oral photographs captured with consumer cameras. It can be useful for enabling the preliminary and cost-effective screening of dental caries among large populations.