Deep Learning Based on Swin-Transformer and 3D U-Net: Implant Three-Dimensional Position Planning

Clinical Snapshot

Key Findings

• P Value: Not reported

• Effect Size: Dice similarity coefficient 0.87–0.91 for dental segmentation; lingual concavity classification accuracy 0.92–0.97; mandibular nerve canal safety margin 3.20–3.89 mm; cervical buccal/lingual bone width 4.94–5.78 mm

• Primary Outcome: Automated 3D implant position planning with anatomical segmentation and lingual concavity classification from CBCT images

• Nnt Or Sensitivity: Sensitivity and specificity for lingual concavity classification not reported; angular and linear implant position deviation metrics not reported; positive/negative predictive values absent

• Confidence Interval: Not reported

Clinical Application

Technically feasible as a pre-surgical planning adjunct integrated into CBCT software workflows. Clinical implementation would require prospective validation, regulatory approval, integration with existing implant planning platforms, and clinician training to appropriately interpret and override model outputs. Workflow integration into busy general dental or specialist oral surgery practice has not been evaluated. In Australia, dental implants are not covered under the Medicare Benefits Schedule and represent a significant out-of-pocket cost for patients. CBCT imaging for implant planning is available through specialist oral and maxillofacial surgery, periodontology, and prosthodontics practices, as well as some general dental practices, and is subject to ARPANSA radiation guidelines. The TGA would need to classify and approve any AI-based implant planning software as a medical device (likely Class IIa or IIb) before clinical deployment. The RACGP and Australian Dental Association (ADA) do not currently have specific guidelines for AI-assisted implant planning. The tool's potential to assist general dentists in identifying high-risk anatomy (lingual concavity, nerve proximity) before referral has relevance in rural and regional Australian settings where specialist access is limited. External validation in Australian patient populations would be essential given potential morphological differences. Adult patients with single-tooth posterior mandibular edentulism being evaluated for endosseous dental implant placement, particularly where mandibular lingual concavity or inferior alveolar nerve proximity poses surgical risk

Abstract

INTRODUCTION AND AIMS: This study aimed to devise a deep learning-based model for the automated identification of anatomical mandibular lingual concavities in the posterior mandible and the prediction of biologically guided three-dimensional implant positions. METHODS: Cone-beam computed tomography (CBCT) images with single-tooth posterior mandibular edentulism were included. A deep learning framework was constructed utilizing 3D U-Net and Swin-Transformer. This framework was designed to perform automated segmentation of teeth, mandible, and the mandibular nerve canal; to classify morphological types of lingual concavities; and to identify implant key points with coordinate prediction. Model performance was assessed via five-fold cross-validation. RESULTS: The proposed model achieved Dice similarity coefficients ranging from 0.87 to 0.91 for dental segmentation. In the classification of mandibular lingual concavities, an accuracy of 0.92 to 0.97 was attained. Regarding the prediction of three-dimensional implant positions, the automatically generated plans maintained a safety margin of 3.20 to 3.89 mm from the mandibular nerve canal. Furthermore, sufficient bone volume was preserved at both cervical and apical implant levels, with buccal/lingual cervical bone widths averaging 4.94 to 5.78 mm. CONCLUSION: The deep learning model in this experiment performed well across different views and learning tasks in a retrospective internal validation setting. Furthermore, it demonstrated the ability to accurately identify relevant anatomical structures, and the predicted three-dimensional implant positions showed clinically acceptable safety margins in the internal validation cohort. CLINICAL RELEVANCE: This model provides a preliminary AI-assisted framework for identifying critical mandibular anatomical structures and generating biologically guided preliminary implant position suggestions, thereby helping to mitigate intraoperative complications and reduce the risk of postoperative mechanical and biological complications in preclinical evaluation.

References

1. 1.Shen, J., Yang, X., Zhang, J., Lu, X., Chen, G., Liu, Y., Ye, B., & Ma, M. (2026). Deep learning based on Swin-Transformer and 3D U-Net: Implant three-dimensional position planning. International Dental Journal. https://doi.org/10.1016/j.identj.2026.109690