Detecting caries lesions of different radiographic extension on bitewings using deep learning

Objectives: We aimed to apply deep learning to detect caries lesions of different radiographic extension on bi-tewings, hypothesizing it to be significantly more accurate than individual dentists. Methods: 3686 bitewing radiographs were assessed by four experienced dentists. Caries lesions were marked in a pixelwise fashion. The union of all pixels was defined as reference test. The data was divided into a training (3293), validation (252) and test dataset (141). We applied a convolutional neural network (U-Net) and used the Intersection-over-Union as validation metric. The performance of the trained neural network on the test dataset was compared against that of seven independent using tooth-level accuracy metrics. Stratification according to lesion depth (enamel lesions E1/2, dentin lesions into middle or inner third D2/3) was applied. Results: The neural network showed an accuracy of 0.80; dentists' mean accuracy was significantly lower at 0.71 (min-max: 0.61-0.78, p 0.05). The neural network was significantly more sensitive than dentists (0.75 versus 0.36 (0.19-0.65; p = 0.006), while its specificity was not significantly lower (0.83) than those of the dentists (0.91 (0.69-0.98; p 0.05); p > 0.05). The neural network showed robust sensitivities at or above 0.70 for both initial and advanced lesions. Dentists largely showed low sensitivities for initial lesions (all except one dentist showed sensitivities below 0.25), while those for advanced ones were between 0.40 and 0.75. Conclusions: To detect caries lesions on bitewing radiographs, a deep neural network was significantly more accurate than dentists. Clinical significance: Deep learning may assist dentists to detect especially initial caries lesions on bitewings. The impact of using such models on decision-making should be explored.