DentalScan AI — Model Weights
Trained weights for a dual-model dental OPG (panoramic X-ray) analysis pipeline: caries detection and mandibular nerve canal segmentation. Full source code and app: github.com/syahmik18/Dental-OPG-AI
Files
| File | Model | Task |
|---|---|---|
attention_unet_best.pth |
Attention U-Net | Mandibular nerve canal segmentation |
best_caries_yolo26.pt |
YOLO26m | Caries (cavity) detection |
Performance
Attention U-Net (nerve canal segmentation)
- Dice: 0.5335
- IoU: 0.3693
- Recall: 0.6726
YOLO26m (caries detection)
- mAP50: 0.4705
- Precision: 0.6498
Training data
445 panoramic dental X-ray (OPG) images, manually annotated with LabelMe — segmentation masks for the nerve canal model, bounding boxes for the caries detection model. Trained on Google Colab and Kaggle (T4×2 GPUs).
How to use
Download the weights and load them with the inference scripts from the source repo:
from huggingface_hub import hf_hub_download
unet_weights = hf_hub_download(repo_id="syahmik18/dentalscan-ai", filename="attention_unet_best.pth")
yolo_weights = hf_hub_download(repo_id="syahmik18/dentalscan-ai", filename="best_caries_yolo26.pt")
Then pass these paths into unet_inference_v2.py / yolo_inference_v2.py from the main repo, or run the full pipeline via final_integrated_detection_v3.py.
Intended use & limitations
This is a research/portfolio project (Final Year Project), not a clinical diagnostic tool. Predictions should not be used to make real treatment decisions without review by a qualified dental professional. Performance metrics reflect a relatively small training set (445 images) and should be interpreted accordingly.
Citation / Acknowledgements
Developed as a Final Year Project in Computer Science, supervised by Dr. Shakirah Binti Hashim.