A deep learning model using convolutional neural networks for caries detection and recognition with endoscopes

BACKGROUND: Caries are common, especially in economically undeveloped countries with limited access to medical resources. Sometimes patient cannot even realize that they have oral problems until they feel obvious pain. Deep convolutional neural networks (CNNs) have been widely adopted for medical image analysis and management and have yielded some progress in stomatology while the endoscopes are cheap and easily used in daily life for families or other non-medical situations. Therefore, we created a deep learning model to detect and recognize caries using endoscopic images. METHODS: We used 194 images of non-caries and 1,059 images of permanent molar and premolar caries to build a classification and a segmentation model in patients of endoscope images from the Department of Stomatology of People’s Liberation Army General Hospital (PLAGH). A classification model combined with an end-to-end semantic segmentation model, DeepLabv3+ was used for segmenting the caries, then we evaluated with a 5-fold cross-validation protocol whereby each fold was used once. RESULTS: In the classification model, the mean area under the curve (AUC) [90% confidence interval (CI)] was 0.9897 (0.9821–0.9956) (P<0.01) In the segmentation model, the mean accuracy was 0.9843 (0.9820–0.9871), the recall was 0.6996 (0.6810–0.7194), the specificity was 0.9943 (0.9937–0.9954), the Dice coefficient was 0.7099 (0.6948–0.7343), and the intersection over union (IoU) was 0.5779 (0.5646–0.6006). CONCLUSIONS: We used a deep learning model to monitor caries and encourage their early diagnosis and treatment.