[Artificial intelligence-assisted analysis of three-dimensional morphological development of the mandibular condyle...

Objective: To develop and evaluate artificial intelligence (AI) models for automated segmentation and three-dimensional (3D) landmark localization of the mandibular condyle on cone-beam CT(CBCT) images, and to preliminarily investigate the 3D morphological developmental characteristics of the condyle from mixed dentition to early permanent dentition based on AI-assisted analysis. Methods: CBCT data of 639 patients (6-14 years old) with oral diseases (non-temporomandibular joint-related) who underwent imaging at the Department of Medical Imaging, College and Hospital of Stomatology, Anhui Medical University, between December 2021 and June 2024 were retrospectively collected. Of these, 100 cases were assigned to the condylar morphological analysis set and 539 to the AI model development set (training set: 443 cases; validation set: 49 cases; test set: 47 cases). Two clinicians independently performed condylar segmentation and annotated 12 osseous landmarks on the development set, and the annotations that passed consistency evaluation served as the ground truth. The PointRend model (adaptive point sampling) and the PoseNet-3D model (supervised learning with 3D Gaussian heatmaps) were separately trained on the training and validation sets. On the test set, the performance of the PointRend model was evaluated using the Dice similarity coefficient (DSC), 95% Hausdorff distance (HD95), and recall rate, and that of the PoseNet-3D model was evaluated using the mean radial error (MRE) and successful detection rate (SDR). The two models were then applied to the morphological analysis set to obtain condylar morphological parameters, including the anteroposterior diameter, mediolateral diameter, and height, as well as their bilateral means and absolute differences. Generalized Procrustes analysis (GPA), Pearson correlation analysis, principal component analysis (PCA), and K-means clustering were used to evaluate age-related condylar morphological differences, major sources of variation, and natural groupings. Results: The PointRend model achieved a DSC of 93.67%, a HD95 of (0.63±0.19) mm, and a recall rate of 89.27%± 5.91%. The PoseNet-3D model achieved a MRE of (1.2±0.23) mm, with the SDR within 2.5 mm reaching 94.53%. The predominant trend in condylar morphology from 6 to 14 years of age was the enhancement of vertical features. K-means clustering partitioned the samples into three developmental stages (6-7, 8-10, and 11-14 years), and condylar height was the most discriminative growth parameter. Conclusions: The two AI models developed in this study enable automated segmentation and quantitative evaluation of three-dimensional condylar morphology, reveal the stage-specific developmental patterns and allometric growth of the condyle, and establish condylar growth reference values from mixed dentition to early permanent dentition.