Radiological differentiation of original and generic dental implants using a multiclass AI model in comparison with human expertise: an in vitro study.

OBJECTIVES: The aim of the present study was therefore to evaluate whether a programmed artificial intelligence (AI) system can reliably differentiate between two original implant manufacturers and two corresponding generic implant systems. Diagnostic performance as well as the certainty of differentiation were directly compared with the assessments of human experts in order to determine the relative accuracy and reliability of AI in comparison with clinical expertise. METHODS: For the identification of four different implant types, a two-stage AI approach was implemented, consisting of an initial object segmentation followed by subsequent classification. A total of 1,384 standardised ex situ radiographs were used as the training dataset. Contour-based segmentation and a pretrained Mask R-CNN were employed for implant extraction. Classification was performed using a transfer-learned ResNet-50 model, which was validated on a separate test dataset. In addition, a comparative evaluation was conducted involving 61 dental students and dentists who assessed separate datasets comprising a total of 98 radiographs. Participants also reported the certainty of their decision as a percentage-based self-assessment. RESULTS: A total of 2,322 individual ratings from 61 assessors were analysed. The accuracy of the human participants was 86.60% and was therefore significantly lower than the error-free classification performance of the AI (p < 0.001). Implant type, clinical experience, and gender showed no significant influence on correct identification, whereas self-reported diagnostic confidence was a strong predictor (p < 0.001). CONCLUSIONS: The findings demonstrate that, under controlled conditions and including multiple implant systems, AI can achieve superior classification performance compared with human assessors in differentiating between nearly identical implants. Overall, the study suggests that AI-based implant identification software may represent a promising diagnostic support tool. CLINICAL SIGNIFICANCE: AI-supported implant identification may enhance clinical decision-making in situations with missing or unclear documentation by improving the reliability of implant recognition. This could reduce the risk of selecting incompatible components and streamline treatment workflows. However, given the controlled study conditions, clinical validation in real-world settings remains necessary before routine implementation.