OralGuard — Uncertainty-Aware Dental Pathology Detection
Trained model weights for the OralGuard pipeline. Detects caries, deep caries, periapical lesions, and impacted teeth on panoramic dental X-rays.
Model Files
- oralguard_det_best.pt — YOLOv8m detector (mAP@50: 0.548)
- classifier_best.pt — ResNet50 multi-label classifier (F1: 0.564)
Dataset
DENTEX Challenge 2023 (MICCAI) — 678 annotated panoramic X-rays across 4 pathology classes with FDI tooth notation.
Architecture
- Tooth detection: YOLOv8m
- FDI notation mapping: Custom coordinate-to-tooth mapper
- Pathology classification: ResNet50 + MC Dropout (T=30)
- Uncertainty quantification: Predictive entropy
- Explainability: GradCAM++
- Active learning: Entropy-based uncertainty sampling
Performance
| Class | mAP@50 |
|---|---|
| Caries | 0.544 |
| Deep Caries | 0.431 |
| Periapical Lesion | 0.263 |
| Impacted Tooth | 0.955 |
| Overall | 0.548 |
Limitations
The classifier flags 100% of predictions as uncertain due to class imbalance in the training data (only 128 periapical lesion examples). This reflects genuine model uncertainty and is the intended behaviour of the uncertainty mechanism. Not validated for clinical use.
Author
Dr. Enosh A. Paulson BDS (RGUHS) | PGDMI Candidate, IIHMR Bangalore GitHub: https://github.com/enosh729-design/Oral_guard HuggingFace: https://huggingface.co/Enosh729
Citation
If you use these weights, please cite the DENTEX 2023 dataset: DENTEX: Dental Enumeration and Diagnosis on Panoramic X-rays Challenge, MICCAI 2023.