Reliability of an Artificial Intelligence Software in the Detection of Approximal Caries Lesions Using Bitewing Radiographs.

INTRODUCTION

This study evaluates the reliability of an artificial intelligence (AI) software for detecting approximal caries lesions of different depth on bitewing radiographs.

METHODS

A total of 40 bitewing radiographs consisting of 288 teeth (576 approximal surfaces) were selected for analysis. Utilising the International Caries Detection and Assessment System radiographic scoring system, five dentists established a consensus on the assessment of all radiographs, serving as the gold standard of this study. These radiographs were later analysed using an AI software (Nostic software®), and the detection results were compared to the established ground truth.

RESULTS

The area under the curve (AUC), accuracy, sensitivity, specificity, positive predictive values, negative predictive values, and F1 scores were computed. A total of 246 surfaces were included for the detection of enamel lesions (D 1-2) while 341 surfaces were assed for dentinal lesions (D 3-4) and for both enamel and dentinal lesions (D 1-4). The accuracy (95% confidence interval) for detecting enamel lesions (D 1-2) was 0.78 (0.72-0.83), for dentinal lesions (D 3-4) was 0.85 (0.80-0.88) and for both enamel and dentinal lesions (D 1-4) was 0.77 (0.73-0.81). Correspondingly, the AUC (95% confidence interval) values for detecting enamel lesions (D 1-2), dentinal lesion (D 3-4) and both enamel and dentinal lesions (D 1-4) were 0.70 (0.65-0.76), 0.81 (0.75-0.87), 0.75 (0.71-0.80), respectively.

CONCLUSION

In conclusion, the performance of the AI software in detecting proximal caries lesions of varying depths on bitewing radiographs was found to be decent when compared to the gold standard. This AI software has the potential to serve as an effective tool to support diagnosing initial caries in bitewing images for dental practitioners.