Related Experiment Video
Updated: May 28, 2026
Measuring Maxillary Posterior Tooth Movement: A Model Assessment using Palatal and Dental Superimposition
Published on: February 23, 2024
Clinically oriented deep learning system integrating linear and morphological assessment for external orthodontic
Dan Yang1, Shuangjiang Yu2, Yue Zhao2
1Department of Orthodontics, The Affiliated Stomatological Hospital of Chongqing Medical University, Chongqing, China; Chongqing Key Laboratory of Oral Diseases, Chongqing, China; Chongqing Municipal Key Laboratory of Oral Biomedical Engineering of Higher Education, Chongqing, China.
Introduction:
Manual linear assessment, a classic method for evaluating orthodontic external root resorption (OERR), has limitations: unreliable dento-osseous junction identification and operator variability, time-consuming measurements, and inability to capture 3-dimensional (3D) morphologic changes.
Methods:
We developed OERR-Net, a deep learning system for objective, real-time OERR linear assessment and 3D visualization using pre and postorthodontic treatment cone-beam computed tomography scans. First, leveraging Transformer architecture, inspired by ChatGPT (OpenAI, San Francisco, Calif), the Swin-UNETR model was adopted for apex-aware tooth segmentation and 3D reconstruction. Second, a novel algorithm (ToothLM) was proposed for automatic tooth length measurement. Third, the system achieved simultaneous grading and 3D morphologic visualization. An end-to-end validation workflow was established, covering segmentation to grading, with Swin-UNETR's superiority demonstrated through qualitative, saliency, and quantitative analyses. Length accuracy was validated via difference and the Bland-Altman analyses, and grading performance was compared with orthodontists' evaluations.
Results:
The study included 100 paired cone-beam computed tomography scans (1560 incisors). First, Swin-UNETR outperformed U-Net and UNETR, achieving the highest agreement with the ground truth (dice similarity score = 90.98%). Second, ToothLM showed excellent agreement with expert manual measurements (intraclass correlation coefficient = 0.999). Finally, OERR-Net achieved superior grading accuracy (maxillary incisors: 97.37% vs 74.34%; mandibular incisors: 96.82% vs 94.27%) than the subjective assessments by orthodontists, captured subtle morphologic changes, and reduced subjective assessment time by 50%, enhancing efficiency and accuracy.
Conclusions:
The proposed automatic OERR assessment system aligns with classic practices, clarifies resorption patterns, and helps treatment selection based on severity. Current validation is single-center; broader applicability requires future validation.
