Accuracy Assessment of Human and Artificial Intelligence-Assisted Bitewing Radiography and Near-Infrared Reflectance Imaging-Based Methods for Interproximal Caries Detection: A Histological Validation.

INTRODUCTION

This study compared the diagnostic accuracy of interproximal caries detection using intraoral bitewing radiographs, assessed by both human operators and an artificial intelligence (AI) program, a near-infrared reflectance imaging (NIRI) system with operator-conducted assessment, and histological evaluation as the reference.

METHODS

100 posterior teeth with or without caries were mounted on 10 typodonts. Initially, 180 surfaces were examined, but eight were excluded for different reasons. Therefore, 171 proximal surfaces were analyzed. NIRI imaging was performed using the iTero Element 5D®, and radiographs were analyzed by operators and an AI program, Denti.AI. Results were compared with histology, assessing sensitivity (Se), specificity (Sp), positive (PPV) and negative (NPV) predictive values, F1-score, areas under receiver operating characteristic curves (AUCs), and the Fleiss Kappa coefficient (k).

RESULTS

The statistical analysis results for each diagnostic test were as follows: examiner radiographic assessment (Se = 52%, Sp = 84.6%, PPV = 71.6%, NPV = 70.3%, F1-score = 60%, AUC = 0.684, k = 0.459); NIRI (Se = 37%, Sp = 98.9%, PPV = 96.4%, NPV = 67.8%, F1-score = 52%, AUC = 0.673, k = 0.475); AI-guided radiographic assessment (Se = 13.7%, Sp = 95.9%, PPV = 71%, NPV = 59.8%, F1-score = 23%, AUC = 0.548). McNemar's test showed a p < 0.05 for all diagnostic tests.

CONCLUSION

Both the operator-conducted NIRI system and examiner radiographic assessment demonstrated superior detection capabilities compared to the AI program. Among these methods, the examiner radiographic assessment yielded the best results, followed by the NIRI system, demonstrating its potential for clinical use.