Longitudinal assessment of an AI-based software for interproximal caries detection in bitewing radiographs

AI-based technologies are increasingly integrated into dental diagnostics, showing promising performance in supporting radiographic interpretation and pathology detection. However, few studies have studied the evolution of these tools over time. Thus, understanding their longitudinal development is crucial to ensure clinical reliability and use.

This longitudinal observational study reassessed the dataset evaluated in 2021 and reanalyzed it in 2025 using the latest version of an AI-Based software. A total of 300 digital bitewing radiographs were examined, with clinical validation serving as the reference standard. Carious lesions were classified following a modified ICDAS criteria. Diagnostic accuracy metrics (Sensitivity, Specificity, F1-score, and Area Under the Receiver Operating Characteristic Curve and McNemar's test) were calculated.

A total of 2646 interproximal surfaces were evaluated, of which 579 were carious and 2067 were sound. The latest version showed lower sensitivity (80 % vs. 87 %) but improved specificity (93 %), PPV (56 %), and AUC (0.794), with fewer error-free surfaces (65.7 %) and more bounding box and omission errors. McNemar's test for paired classifications found a significant difference between versions (p < 0.001).

The longitudinal assessment of the AI software reveals a clear shift toward higher specificity and reduced overdiagnosis, albeit with lower sensitivity and an increase in false negatives. Ongoing longitudinal evaluations are crucial for ensuring the safe and effective integration of AI tools in real-world dental settings.

AI-based diagnostic tools evolve with data exposure, requiring clinicians to account for changing performance over time. While these systems can sharpen detection decisions from bitewing radiographs, successful use depends on continuous, longitudinal performance monitoring, tracking accuracy, and adjusting thresholds as needed.