Advanced deep learning techniques for recognition of dental implants
Background: Dental implants are the most accepted prosthetic alternative for missing teeth. With growing demands, several manufacturers have entered the market and produce a variety of implant brands creating a challenge for clinicians to identify the implant when the necessity arises. Currently, ra...
Saved in:
| Main Authors: | , , , |
|---|---|
| Format: | Article |
| Language: | English |
| Published: |
Elsevier
2025-03-01
|
| Series: | Journal of Oral Biology and Craniofacial Research |
| Subjects: | |
| Online Access: | http://www.sciencedirect.com/science/article/pii/S2212426825000181 |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | Background: Dental implants are the most accepted prosthetic alternative for missing teeth. With growing demands, several manufacturers have entered the market and produce a variety of implant brands creating a challenge for clinicians to identify the implant when the necessity arises. Currently, radiographs are the only tools for implant identification which is inherently a complex process, hence the need for implant identification technique. Artificial intelligence capable of analysing images in a radiograph and predicting implant type is an efficient tool. The study evaluated an advanced deep learning technique, DEtection TRanformer for implant identification. Methods: A transformer-based deep learning technique, DEtection TRanformer was trained to identify implants in radiographs. A dataset of 1138 images consisting of five implant types captured from periapical and panoramic radiographs was chosen for the study. After augmentation, a dataset of 1744 images was secured and then split into training, validation and test datasets for the model. The model was trained and evaluated for its performance. Results: The model achieved an overall precision of 0.83 and a recall score of 0.89. The model achieved an F1-score of 0.82 indicating a strong balance between recall and precision. The Precision-Recall Curve, with an AUC of 0.96, showed that the model performed well across various thresholds. The training and validation graphs showed a consistent decrease in the loss functions across classes. Conclusion: The model showed high performance on the training data, though it faced challenges with unseen validation data. High precision, recall and F1 score indicate the model's potential for implant identification. Optimizing this model for a balance between accuracy and efficiency will be necessary for real-time medical imaging applications. |
|---|---|
| ISSN: | 2212-4268 |