Performance of Chat Generative Pretrained Transformer-4.0 in determining labiolingual localization of maxillary impacted canine and presence of resorption in incisors through panoramic radiographs: A retrospective study

Abstract

Introduction: This study evaluates the diagnostic accuracy of Chat Generative Pretrained Transformer (ChatGPT)-4.0 in determining the labiolingual position of impacted maxillary canines and identifying resorptive changes in adjacent incisors using panoramic radiographs (PRs). Methods: A retrospective analysis was conducted on 105 patients with unilaterally impacted maxillary canine, including 25 patients with root resorption in adjacent incisors. To ensure accurate classification, PRs and cone-beam computed tomography images were independently evaluated by 3 orthodontists, serving as the reference standard for assessing ChatGPT-4.0's performance. Patients were categorized into 3 groups based on canine position: palatal (n = 49), midalveolar (n = 26), and labial (n = 30). For resorption evaluation, a balanced subset of 50 PRs was selected to maintain equal group sizes. Group 1 (with resorption, n = 25) included all available patients with resorption, whereas group 2 (without resorption, n = 25) was randomly selected from 80 patients without resorption. ChatGPT-4.0 analyzed the PRs to determine the labiolingual position of impacted canines and detect resorption in adjacent incisors. The results were recorded by the first researcher. ChatGPT-4.0's performance was evaluated using accuracy, precision, recall, and F1 score for both canine localization and resorption detection. Results: The model achieved an overall accuracy of 37.1% in canine localization, with the highest sensitivity (61.2%) and precision (48.4%) observed in palatal patients. However, its performance was considerably lower for midalveolar and labial positions. In detecting resorption, the model achieved an accuracy of 46.0%, performing better in identifying the absence of resorption compared with its presence. Conclusions: ChatGPT-4.0 demonstrated insufficient accuracy in determining the labiolingual position of impacted maxillary canines and detecting resorptive changes based on PRs, indicating its unsuitability for clinical applications. (Am J Orthod Dentofacial Orthop 2025;168:220-31)