Abstract
Purpose: This study evaluated the effectiveness of hybrid deep learning and vision models in classifying and segmenting impacted molar teeth from full-mouth panoramic radiographs. Methods: The dataset of 693 panoramic radiographs was split into training (80 %), validation (10 %), and test (10 %) sets. Random oversampling was applied to address class imbalance. CNN, CLIP, Vision Transformer (ViT), CNN with ViT embedding were used to classify the radiographs, and a U-Net model was applied to segment the impacted teeth in the dataset. Precision, recall, F1 score, accuracy, AUC, and IoU were used to assess performance. Results: The CNN with ViT embedding achieved the highest overall accuracy (0.77) with a precision of 0.43 in identifying panoramic images with impacted molar teeth. ViT alone demonstrated the highest AUC (0.85) with the fewest misclassifications. CNN performed best in identifying radiographs with no impaction, achieving a precision of 0.84 and recall of 0.86. The U-Net segmentation model produced a class-wise IoU of 14.2497 for the background and 1.1050 for impacted teeth. CLIP-based models exhibited varied performance. Conclusions: A hybrid CNN with ViT embedding was able to classify impacted molar teeth from panoramic radiographs. U-Net struggled to segment impacted molar teeth.