95 % F-score in teeth recognition on periapical X-rays using Faster R-CNN

A Relation-Based Framework for Effective Teeth Recognition on Dental Periapical X-Rays

Authors: Kailai Zhang, Hu Chen, Peijun Lyu, Ji Wu

---

Summary This study from Tsinghua University and Peking University School and Hospital of Stomatology (Beijing, China) introduces a relation-based deep learning framework for automated teeth recognition in periapical radiographs. Traditional CNN-based methods often struggle with small datasets containing many tooth positions (32 labels). To address this, the authors embedded prior dental knowledge into the model architecture. The proposed system combines three key innovations:

1. Label Reconstruction Technique – decomposes the 32-class problem into smaller subtasks (tooth presence, upper/lower jaw, tooth 1–8).
2. Proposal Correlation Module – models spatial relationships between adjacent teeth using an attention-like mechanism.
3. Teeth Sequence Refinement Module – corrects classification results based on anatomical logic. The dataset included 1,250 periapical X-rays (4,336 teeth), annotated by dentists using the FDI numbering system.
   Using a multi-task CNN based on Faster R-CNN (ResNet backbone) and 10-fold cross-validation, the system achieved:

• Precision: 0.951
• Recall: 0.955
• F-score: 0.953

These results outperform conventional detectors (Fast R-CNN, Faster R-CNN, R-FCN), confirming the benefits of integrating domain-specific dental knowledge into the network. The framework offers reliable support for both clinical diagnostics and forensic dental identification, particularly when analyzing partial dentitions or isolated teeth.

DOI: https://doi.org/10.1016/j.compmedimag.2021.102022

---

Key Words

Periapical Radiography, Teeth Recognition, Multi-Task CNN, Label Reconstruction, Proposal Correlation Module, Sequence Refinement, Deep Learning, Dental AI, Computer Vision, Oral Radiology

Extracted Data

• Year: 2022
• Modality: Dental Periapical Radiography
• Dataset: 1,250 images (4,336 teeth)
• Dataset Split: 10-fold cross-validation (1125 training / 125 test)
• Network Architecture: Multi-task CNN based on Faster R-CNN (ResNet backbone)
• Metrics: Precision 0.951 | Recall 0.955 | F1 0.953
• AP – Strategy: Professional bounding-box annotation using FDI system
• AP – Professional Qty: No Information
• AP – Supervisor Presence: No Information
• AP – Experience Level: No Information
• AP – Expertise Area: General Dentists
• AP – Tool or System: No Information
• Task: Object Detection (bounding box)
• Project Objective: To develop a relation-based deep learning framework for accurate teeth recognition in periapical radiographs by embedding prior anatomical knowledge into CNN structure.

---

Clinical Relevance

• Clinical importance: Periapical radiographs show small and localized regions, which limits contextual information and makes tooth identification more difficult for both clinicians and AI models.
• Innovation: The model integrates anatomical and positional relationships between teeth through correlation and refinement modules, achieving high performance even with limited training data.
• Practical impact: This framework can support automated diagnostic and forensic workflows, reducing manual effort in tooth labeling and improving consistency in dental radiographic analysis.

Related Video

Watch on YouTube