Diagnostic Accuracy of Novel AI-Based Software in the Detection of Dental Caries on Bitewing and Intraoral Periapical Radiographs

Background: Dental caries detection using intraoral radiographs is essential for early diagnosis but may be affected by observer variability. Artificial intelligence (AI)-based systems have emerged as potential tools to improve diagnostic consistency. This study evaluated the diagnostic accuracy of a deep learning-based AI system (Better Diagnostics Caries Assist (BDCA) Version 1.0) for detecting dental caries on bitewing (BW) and intraoral periapical (IOPA) radiographs, and examined its performance across demographic, technical, and lesion-based subgroups. Methods: A retrospective validation study was conducted using anonymized digital BW and IOPA radiographs with expert-defined ground truth at the tooth-surface level. The AI software independently analyzed each image to identify carious lesions. Sensitivity, specificity, positive predictive value (PPV), and negative predictive value (NPV) were calculated with 95% confidence intervals. Generalized estimating equations were applied to adjust for the clustering of multiple surfaces per image. Subgroup analyses were performed by age, sex, digital sensor type, and lesion category. Results: The AI system demonstrated high diagnostic accuracy for both modalities. For BW radiographs, sensitivity was 0.892 and specificity was 0.995, while for IOPA radiographs sensitivity was 0.882 and specificity was 0.991. NPVs exceeded 0.99 for both modalities. Across age groups, BW sensitivity ranged from 0.881 to 0.901 and IOPA sensitivity from 0.854 to 0.906, with consistently high specificity (>0.989). Sex-based differences were minimal. Sensor-wise analysis showed sensitivity ranging from 0.833 to 0.933 for BW and 0.828 to 0.933 for IOPA, while specificity remained above 0.984 for all sensors. Detection performance was comparable for primary (sensitivity 0.883) and secondary caries (0.879), although PPV was slightly lower for secondary lesions. The lower AUC indicated reduced accuracy in lesion identification in the absence of BDCA v 1.0, the difference between BDCA v1.0 and Ground truth was 0.042 at 95% CI 0.030 to 0.055, and the difference was also statistically significant with (p ≤ 0.001). Conclusions: The evaluated AI system demonstrated excellent and consistent performance for detecting dental caries on both BW and IOPA radiographs across demographic groups, sensor technologies, and lesion types, supporting its potential role as a reliable decision support tool in dental radiographic interpretation.