Evaluation of automated tooth landmark localization on digital models

To evaluate agreement between the tooth landmark localization patterns of artificial intelligence (AI) and those from human examiners.

Three-dimensional (3D) digital dental model images were obtained from 284 participants. On a total of 5583 permanent teeth, six landmarks per tooth were manually identified and annotated using custom-made 3D annotation software. To develop an AI model capable of automatically identifying tooth landmarks, a deep-learning algorithm was applied to a training dataset consisting of 4519 teeth. To select the optimal AI model, datasets of 556 and 508 teeth were used as validation and test datasets, respectively. For intraexaminer and interexaminer reliability tests, 280 teeth from 10 participants were randomly selected, and two human examiners identified the same six landmarks on two separate occasions.

The mean error in tooth landmark localization of the AI model ranged from 0.01 mm to 1.68 mm. The intraclass correlation coefficient between the AI model and human examiner for all landmarks was excellent, ranging from 0.97 to 1.0. The landmark localization error from the AI model was smaller than human interexaminer differences for mesial and distal proximal points. However, errors for the cusp tip and facial axis points were greater in the AI model than the interexaminer differences.

AI exhibited localization accuracy for tooth landmarks comparable with that of human examiners for specific measurements related to tooth size. Nonetheless, its accuracy still needs improvement to match that of orthodontic clinicians in identifying cusp tips and facial axis points.