Diagnosis of interproximal caries lesions with deep convolutional neural network in digital bitewing radiographs.

OBJECTIVES: This study aimed to investigate the effectiveness of deep convolutional neural network (CNN) in the diagnosis of interproximal caries lesions in digital bitewing radiographs.

METHODS AND MATERIALS: A total of 1,000 digital bitewing radiographs were randomly selected from the database. Of these, 800 were augmented and annotated as "decay" by two experienced dentists using a labeling tool developed in Python programming language. The 800 radiographs were consisted of 11,521 approximal surfaces of which 1,847 were decayed (lesion prevalence for train data was 16.03%). A CNN model known as you only look once (YOLO) was modified and trained to detect caries lesions in bitewing radiographs. After using the other 200 radiographs to test the effectiveness of the proposed CNN model, the accuracy, sensitivity, specificity, positive predictive value (PPV), negative predictive value (NPV), and area under the receiver operating characteristic curve (AUC) were calculated.

RESULTS: The lesion prevalence for test data was 13.89%. The overall accuracy of the CNN model was 94.59% (94.19% for premolars, 94.97% for molars), sensitivity was 72.26% (75.51% for premolars, 68.71% for molars), specificity was 98.19% (97.43% for premolars, 98.91% for molars), PPV was 86.58% (83.61% for premolars, 90.44% for molars), and NPV was 95.64% (95.82% for premolars, 95.47% for molars). The overall AUC was measured as 87.19%.

CONCLUSIONS: The proposed CNN model showed good performance with high accuracy scores demonstrating that it could be used in the diagnosis of caries lesions in bitewing radiographs.

CLINICAL SIGNIFICANCE: Correct diagnosis of dental caries is essential for a correct treatment procedure. CNNs can assist dentists in diagnosing approximal caries lesions in bitewing radiographs.