AI-assisted radiographic identification of original vs. replica dental implants : comparing accuracy of human experts vs. probabilistic and deterministic AI

Abstract

Purpose In dental implantology, the application of artificial intelligence (AI) for the differentiation of various implant systems is gaining increasing importance. This study investigates the feasibility of distinguishing between two highly similar implant (original implant and its replica) systems using an automated, AI-based recognition software. Methods A dataset of 906 radiographic images was initially compiled, consisting of standardized ex situ recordings of both the original and the replica implants (with and without a cover screw in situ). Four deterministic AI-models and one probabilistic model were trained using different subsets of varying sizes of the dataset, including the full dataset and then evaluated against a designated test dataset. For comparison, 28 dental professionals also assessed the same test dataset. Results The accuracy of the deterministic model trained solely with 488 radiographs of implants with inserted cover screws was 0.579 (57.9%). The second and third models, trained with a greater number of radiographs without inserted cover screws, achieved accuracies exceeding 0.90 and, in some instances, even reached 1.00. The fourth deterministic model, as well as the probabilistic model, comprising 28 classifiers and trained on the complete dataset, classified the test dataset without error. The dental professionals achieved an overall accuracy of 0.8616 (86.16%) in their assessment of the test dataset. Conclusion This study suggests that AI-supported implant recognition software has the potential to offer valuable assistance in clinical practice for distinguishing between original and replica implants. Such differentiation can play a crucial role for prosthetic suprastructures and associated manufacturer warranties.