Deep learning-based object detection of dental implant systems in panoramic and periapical radiographs

Clinical Snapshot

Key Findings

• P Value: Not reported

• Effect Size: Overall mAP: 98.3%; Mean precision: 93%; Mean recall: 86%; Mean F1-score: 89%. Osstem: precision 99%, recall 95%. Nobel: precision not specified, recall 72.7%

• Primary Outcome: Mean average precision (mAP) for automated identification of 7 dental implant systems on panoramic and periapical radiographs using YOLOv10

• Nnt Or Sensitivity: Recall (sensitivity analogue): overall 86%; Osstem 95%; Nobel 72.7%. Specificity not reported. Positive/negative likelihood ratios not calculable from reported data

• Confidence Interval: Not reported for any metric

Clinical Application

Technical feasibility is demonstrated in a controlled research environment. Clinical deployment would require integration with practice management or PACS systems, regulatory approval, prospective clinical validation, and clinician training. Real-time detection capability (YOLO architecture) is theoretically compatible with chairside use. However, the model is not yet validated for clinical deployment and should not replace expert clinical judgement Osstem, Nobel Biocare, and MIS implants are among those used in Australian prosthodontic practice. The TGA would require regulatory approval as a Class IIb or Class III medical device software (SaMD) under the Therapeutic Goods (Medical Devices) Regulations 2002 before clinical deployment. RACGP and the Australian Dental Association (ADA) have not yet issued specific guidelines on AI-assisted implant identification. PBS does not currently fund AI diagnostic tools in dentistry. Australian dental practices would need to consider medicolegal liability for AI-assisted diagnoses. Multi-centre Australian validation including diverse implant brands used locally (e.g., Straumann, Zimmer Biomet) would be required before adoption Patients presenting with existing dental implants of unknown system, particularly where clinical documentation is unavailable. Most applicable where Osstem, Dentium, or Adin implants predominate; less reliable for Nobel implants in current form

Abstract

STATEMENT OF PROBLEM: The manual identification of dental implant systems on radiographs is time-consuming, operator-dependent, and prone to diagnostic inaccuracies, particularly for patients where clinical documentation is lacking. The increasing variety of implant designs further complicates identification in prosthetic and surgical practice. PURPOSE: The purpose of this study was to develop and evaluate a deep learning-based model for the automated identification of 7 implant systems (Adin, Dentium, Dionavi, Make It Simple (MIS), Nobel, Noris, and Osstem) using panoramic radiographs and periapical radiographs in an effort to enhance diagnostic efficiency and support clinical decision-making in prosthodontic care. MATERIAL AND METHODS: A total of 4677 anonymized radiographic images with 8189 implants were curated and annotated using Roboflow with bounding boxes outlining fixture components. The preprocessing involved normalization, resizing to 640×640 pixels, and geometric augmentation (rotation, cropping, and blurring) to handle class imbalances. You Only Look Once (YOLO) v10 architecture, implemented with PyTorch, using CSPDarknet and PANet for multiscale feature fusion, was used to optimize real-time detection. Transfer learning used pretrained weights, with training for over 500 epochs (batch size: 32) on NVIDIA T4 GPUs. Data partitioning involved an 80:10:10 ratio (training: validation: testing), with performance evaluated using precision, recall, F1-score, and mean average precision (mAP). RESULTS: The model achieved a mAP of 98.3%, with mean precision, recall, and F1-score values of 93%, 86%, and 89%, respectively. Osstem implants demonstrated maximum discriminability (99% precision, 95% recall). In contrast, Nobel implants exhibited low recall (72.7%), attributed to the sparsity of the dataset (564 samples for Nobel compared with 2320 for Osstem) and similar radiopacity patterns. CONCLUSIONS: The YOLOv10 model demonstrated good performance in identifying dental implants, showing clinical promise for minimizing prosthetic mismatches. Subject to ethics and regulatory approvals, additional improvements involving 3-dimensional imaging and heterogeneous datasets may add precision and validate artificial intelligence as an evidence-based advance in implant dentistry.

References

1. 1.Pachaiappan, B., Shafiullah, R. S., Balaraman, G., Bhuvaneshwar, Y. P., & Sadasivan, S. (2026). Deep learning-based object detection of dental implant systems in panoramic and periapical radiographs. The Journal of Prosthetic Dentistry. https://doi.org/10.1016/j.prosdent.2026.01.029