Assessment of AI software's diagnostic accuracy in identifying impacted teeth in panoramic radiographs.

Abstract

Background/objectives: Recently, advancements have been made in the application and development of artificial intelligence (AI) tools in dentistry. This study aims to assess the diagnostic accuracy of an AI-driven platform in identifying impacted teeth using panoramic radiographs.

Materials/methods: A total of four sets of 50 orthopantomograms were examined: one set featured impacted canines, another included impacted third molars, a third contained impacted incisors, premolars, and both first and second molars, and the final set had no impacted teeth. Two human observers and the Diagnocat™ 1.0 software independently evaluated the images. The level of agreement was measured using Cohen's Kappa, and calculations for sensitivity, specificity, positive predictive value (PPV), and negative predictive value (NPV), along with the corresponding 95% Confidence Intervals, were also conducted. The number of impacted teeth identified by both methods was compared using the Wilcoxon signed-rank test, and McNemar's tests were performed to identify any differences in the proportions of identified impacted teeth between the two methods. Analyses were carried out using IBM SPSS version 29.0.

Results: The evaluation of the AI software's diagnostic performance in recognizing impacted teeth compared with expert clinicians showed that Diagnocat performed exceptionally well in terms of specificity and positive predictive value (PPV), demonstrating a highly reliable identification of impacted teeth with no false positives. The sensitivity for identifying third molars was also good. However, there were significant limitations in sensitivity for other impacted teeth, suggesting that negative results might require further consideration. Cohen's Kappa indicated almost perfect agreement between Diagnocat™ and expert assessments for identifying impacted third molars, but only fair agreement for impacted canines and other teeth. Significant differences were observed in the average number and the proportions of impacted teeth detected by the two methods.

Limitations: Employing a retrospective design and convenience sampling may limit the study's generalizability and clinical relevance.

Conclusion: While the AI-based platform shows promise in detecting impacted third molars, it is still insufficient to replace human evaluation as the standard for assessing impacted teeth in panoramic radiographs.