Artificial intelligence-assisted analysis of tongue morphology and sagittal skeletal patterns: A multicenter study.

INTRODUCTION This study investigated the association between tongue morphology and sagittal skeletal patterns and assessed a deep learning model for automated tongue landmark detection and segmentation on lateral cephalometric radiographs (LCRs).

Methods

Pretreatment LCRs from 503 orthodontic patients (aged 6-40 years) were analyzed, categorized by dentition stage (mixed: 6-11; permanent: 12-40), and sagittal skeletal patterns. Tongue measurements were compared across subgroups. A multicenter dataset of 1179 LCRs trained a deep learning framework integrating U-Net for segmentation and RTMPose (real-time model for pose estimation) for landmark detection: tongue tip, dorsum of tongue, and base of the epiglottis. Model performance was assessed using mean radial error, success detection rate, mean intersection over union, and recall. External validation was performed on an independent hold-out dataset from 2 hospitals.

Results

In the permanent dentition group, subjects with Class Ⅲ malocclusion showed larger tongue areas than Class Ⅱ but not Class I; no significant differences in tongue area were observed in the mixed dentition group. The model achieved a mean radial error of 1.72 ± 1.78 mm, success detection rates of 75.65% (≤2 mm) and 91.02% (≤4 mm), a mean intersection over union of 0.907, and a recall of 0.952. Comparable performance was confirmed on external validation.

Conclusions

Sagittal skeletal patterns were associated with sagittal tongue area, highlighting the importance of tongue assessment during growth. The proposed deep learning model achieved accurate tongue landmark detection and segmentation, supporting automated cephalometric tongue analysis in orthodontics.