dental-fdi-detection

🦷 Dental FDI Detection

Transformer-based tooth detection for panoramic dental X-rays

Evaluation release trained on 4,000+ panoramic radiographs, curated for clinical diversity.
CPU-optimized, clinic-ready design for integration into dental AI systems.

🦷 Live Demo • 📧 Professional Services

What this is

This repository contains a production-targeted tooth detection model designed for dental AI pipelines.

Key characteristics

• Architecture: Transformer-based object detector
• Output: 32-tooth FDI proposals (bounding boxes + confidence)
• Inference: ~20 ms GPU / ~300 ms CPU
• Deployment: ONNX, CPU-first, on-prem ready

This model represents the detection stage of a larger clinical workflow and is optimized for integration, not standalone diagnostic use.

Scope: Permanent dentition (adults, 18+). Primary/deciduous teeth not supported.

Why this design works

Production medical AI requires different trade-offs than research:

Architecture: Decoupled proposal + validation stages
Benefit: Each stage optimizable, testable, replaceable independently

This is systems engineering, not model tuning.

Dataset & training

• 4,000+ panoramic radiographs
• Population: Adult permanent dentition (18+ years, 32 teeth)
• Exclusions: Primary teeth, mixed dentition, pediatric cases — significantly different topology
• Engineered for clinical diversity and edge-case stability:
  • Variable dentition states (full, partial, edentulous)
  • Metal artifacts and implants
  • Exposure and positioning variance
• Quality-controlled FDI annotations

Training objective: Recall stability under artifact-heavy conditions, not benchmark maximization.

Scale note: This repository contains a reference training run.
Production deployments use expanded, institution-specific datasets.

Performance characteristics

• CPU inference: ≤350 ms on reference hardware (Intel Xeon E5-2680 v4)
• GPU inference: ≤22 ms on NVIDIA T4
• Recall behavior: Stable tooth proposal under heavy artifacts, partial dentition, and missing molars

Exact numbers vary by hardware.
The system prioritizes predictable upper bounds, not best-case benchmarks.

Distribution Scope

Included:

✅ ONNX detection weights for 32-tooth FDI proposals
✅ Inference pipeline (CPU/GPU)
✅ Clinical preprocessing utilities

Proprietary Modules (Decoupled):

🔒 Anatomical consistency engine — available under commercial license
🔒 Training infrastructure and dataset curation tools
🔒 Regulatory documentation (ISO 13485 / CE Mark support)

Why this matters

Benchmarks optimize average accuracy.
Clinics need worst-case reliability.

This system prioritizes:

• Recall over precision (missing teeth vs mislabeling)
• Auditability over automation (separated stages)
• CPU deployment over GPU clusters

Engineered for deployment, not leaderboard optimization.

Professional use

This repository demonstrates a single component of a deployable dental AI system.

In practice, production deployments typically require:

• Integration with clinical imaging workflows (PACS / DICOM)
• Adaptation to site-specific acquisition protocols
• Validation under local artifact and population distributions
• Alignment with applicable regulatory frameworks

The goal of this repository is to expose a robust detection stage that can be integrated into such systems, not to present a complete end-to-end solution.

If you are working on dental imaging workflows and need a system that functions reliably in real clinical conditions rather than benchmark settings, feel free to reach out for a technical discussion.

📧 [email protected]

License

Proprietary — Non-Commercial Research License

Allowed

• Academic research and evaluation
• Internal testing and validation
• Educational use

Prohibited

• Commercial deployment without agreement
• Clinical diagnostic use
• Model redistribution or derivative works

Commercial licensing: [email protected]

See LICENSE and NOTICE for details.

Citation

@misc{bychkov2026dentalfdi,
  author       = {Bychkov, Anton},
  title        = {Dental FDI Detection: Transformer-Based Tooth Proposals},
  year         = {2026},
  publisher    = {Hugging Face},
  howpublished = {\url{https://huggingface.co/abychkov/dental-fdi-detection}},
  note         = {Contact: [email protected]}
}