DentalVision AI - EfficientNetB3 (v2.0.0)
Multi-label classifier for six findings on dental OPG X-rays (EfficientNetB3 + DenseNet121 + ConvNeXt-Tiny ensemble). Educational and research prototype. Not a diagnostic tool. Predictions must be reviewed by qualified dental professionals.
- Output: 6 independent sigmoid probabilities in this order: 0=Healthy Teeth, 1=Caries, 2=Impacted Teeth, 3=Infection, 4=Fractured Teeth, 5=BDC/BDR. "Healthy Teeth" means the image contains healthy teeth.
- Input: RGB, 288x576 (H x W), bilinear resize, ImageNet mean/std normalisation
- Training: ImageNet transfer learning, BCE with positive class weights, AdamW, cosine schedule, early stopping on validation macro-AUC
- Data: Kaggle
imtkaggleteam/dental-opg-xray-dataset(not redistributed): 230 unique radiographs after removing byte-identical copies stored under several class folders.
Results
| Evaluation | Macro AUC | mAP | Macro F1 |
|---|---|---|---|
| Fold-0 test (n=37) | 0.695 | 0.599 | 0.508 |
| Pooled 6-fold out-of-fold (230 images) | 0.661 | 0.544 | 0.503 |
Limitations
Very small dataset, rare findings, single source, no patient IDs, no external validation.