Fully automated deep learning for the identification and classification of dental implant brands from multimodal radiographic images.
Authors
Affiliations (7)
Affiliations (7)
- General Dentistry Department, The Stomatology Hospital Affiliated of Suzhou Vocational Health College, Suzhou, China.
- Medical School of Nanjing University, Nanjing, China.
- The Stomatology Hospital Affiliated of Suzhou Vocational Health College, Suzhou, China.
- Department of Oral Implantology, The Stomatology Hospital Affiliated of Suzhou Vocational Health College, Suzhou, China.
- Suzhou Vocational Health College, Suzhou, China.
- Outpatient Operating Theatre, The Stomatology Hospital Affiliated of Suzhou Vocational Health College, Suzhou, China.
- Department of Oral Implantology, The Stomatology Hospital Affiliated of Suzhou Vocational Health College, Suzhou, China. [email protected].
Abstract
Accurate radiographic identification of dental implant systems is essential for effective clinical management; however, manual assessment remains time-consuming and susceptible to diagnostic error. Furthermore, conventional computational approaches are constrained by their dependency on manually annotated regions of interest. To overcome these limitations, this study introduces DentalImplantNet, an end-to-end automated deep learning architecture engineered for robust, high-precision implant classification across heterogeneous and radiographically noisy clinical settings. A multi-modal dataset was constructed from 2,714 panoramic and periapical radiographs alongside high-resolution Bruker micro-CT scans of 64 standalone implant fixtures. An automated pipeline utilizing RT-DETR for real-time localization and ROI extraction was implemented to eliminate manual intervention. To enhance robustness, a Causal-Informed Generalization Framework based on Backdoor Adjustment was utilized to dissociate invariant morphological features from non-causal radiographic noise. Model performance was benchmarked against ResNet-50, EfficientNet-B2, and 12 dental professionals who were divided into three levels of expertise. DentalImplantNet achieved an overall accuracy of 95.29% and a micro-average AUC of 99.6%, significantly outperforming ResNet-50 and EfficientNet-B2, which recorded accuracies of 92.4% and 88.2%, respectively. The model demonstrated high diagnostic stability with an overall F1-score of 94.2% and achieved perfect precision for the Straumann and Osstem categories. In clinical validation, the artificial intelligence (AI) performance substantially surpassed senior implant specialists whose overall and specifically for periapical radiographs did not exceed average numbers of 58% and 53%, respectively. Grad-CAM visualization confirmed that the model prioritized sub-millimeter morphological signatures over transient radiographic artifacts, ensuring high interpretability. The integration of causal data synthesis with automated deep learning architectures provided a scalable and robust solution for implant identification, effectively minimizing human error in clinical practice. Automated deep learning models have demonstrated significant potential in assisting dentists with the precise identification of dental implant brands on panoramic and periapical radiographs, thereby facilitating more effective, efficient, and safer clinical management.