Abstract
Background: The assessment of orthodontic treatment outcomes traditionally relies on clinical examination, two-dimensional radiographs, cephalometric measurements, and manual evaluation of dental casts or digital models. Although these methods are widely used, they are associated with operator-dependent variability, time-consuming workflows, and limitations in accurately evaluating complex three-dimensional dentofacial changes. Recent advances in artificial intelligence (AI), deep learning, computer vision, and three-dimensional image processing offer opportunities for automated and reproducible assessment of orthodontic treatment outcomes.
Aim: To develop and evaluate a multimodal artificial intelligence-based model integrating three-dimensional (3D) digital dental models and cone-beam computed tomography (CBCT) imaging for automated assessment of orthodontic treatment outcomes.
Materials and Methods: A retrospective multicentre dataset comprising 240 orthodontic patients was constructed for this methodological study. Pretreatment and post-treatment 3D digital dental models and clinically indicated CBCT datasets were considered. The proposed AI framework incorporated convolutional neural networks, a 3D U-Net segmentation architecture, point-cloud processing, and a multimodal feature-fusion network. Automated assessment included dental alignment, arch form, intercanine and intermolar widths, overjet, overbite, transverse discrepancies, root positioning, root parallelism, and skeletal relationships. Expert orthodontist assessments were considered the reference standard. The dataset was divided at the patient level into training, internal testing, and external testing cohorts. Model performance was evaluated using Dice similarity coefficient (DSC), mean absolute error (MAE), intraclass correlation coefficient (ICC), sensitivity, specificity, F1-score, and area under the receiver operating characteristic curve (AUC).
Results: The proposed multimodal model demonstrated an overall AUC of 0.94 (95% CI: 0.87–0.98) for classifying orthodontic treatment outcomes as satisfactory or unsatisfactory. The corresponding sensitivity and specificity were 91.3% and 89.7%, respectively. The mean absolute error for quantitative dental measurements was 0.42 ± 0.18 mm, while the mean radial error for CBCT-derived anatomical landmark detection was 1.12 ± 0.46 mm. The DSC for automated tooth segmentation was 0.96 ± 0.02. The multimodal model demonstrated better diagnostic performance than digital-model-only and CBCT-only models. The mean assessment time was 2.8 minutes compared with 24.6 minutes for conventional manual assessment.
Conclusion: The proposed multimodal AI framework demonstrates the potential to provide an integrated, efficient, and reproducible approach to orthodontic treatment outcome assessment. Combining 3D digital models with CBCT-derived anatomical information may improve the comprehensiveness of automated evaluation. However, the findings require confirmation through actual multicentre clinical datasets, independent external testing, and prospective clinical validation before clinical implementation.
Keywords: Artificial Intelligence, Orthodontics, Deep Learning, Three-Dimensional Digital Models, Cone-Beam Computed Tomography, Treatment Outcome Assessment, Tooth Segmentation, Multimodal Learning, Automated Orthodontic Evaluation.
References