Diagnosis of Interproximal Caries Lesions in Bitewing Radiographs Using a Deep Convolutional Neural Network-Based Software

The aim of this study was to evaluate the diagnostic reliability of a web-based artificial intelligence program for the detection of interproximal caries in bitewing radiographs. Three hundred bitewing radiographs of patients were subjected to the evaluation of a convolutional neural network. First, the images were visually evaluated by a previously trained and calibrated operator with radiodiagnosis experience. Then, ground truth was established and was clinically validated. For enamel caries, clinical assessment included a combination of clinical-visual and radiography evaluations. For dentin caries, clinical validation was performed by instrumentally accessing the cavity. Second, the images were uploaded and analyzed by the web-based software. Four different models were established to analyze its evaluations according to the confidence threshold (0–100%) offered by the program: model 1 (values >0% were considered positive and values of 0% were considered negative), model 2 (values ≥25% were considered positive and values <25% were considered negative), model 3 (values ≥50% were considered positive and values <50% were considered negative), and model 4 (values ≥75% were considered positive and values <75% were considered negative). The accuracy rate (A), sensitivity (S), specificity (E), positive predictive value (PPV), negative predictive value (NPV), positive likelihood ratio (PLR), negative likelihood ratio (NLR), and areas under receiver operating characteristic curves (AUC) were calculated for the four models of agreement with the software. Models showed the following results respectively: A = 70.8%, 82%, 85.6%, 86.1%; S = 87%, 69.8%, 57%, 41.6%; E = 66.3%, 85.4%, 93.7%, 98.5%; PPV = 42%, 57.2%, 71.6%, 88.6%; NPV = 94.8%, 91%, 88.6%, 85.8%; PLR = 2.58, 4.78, 9.05, 27.73; NLR = 0.2, 0.35, 0.46, 0.59; AUC = 0.767, 0.777, 0.753, 0.701. Findings in the present study suggest that the artificial intelligence web-based software provides a good diagnostic reliability on the detection of dental caries. Our study highlighted model 2 for showing the best results to differentiate between healthy teeth and decayed teeth.