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.
