OralHybridNet: Multi-Label Dental Restorations, Dental Prosthesis and Endodontic Treatments Classification in Panoramic Radiographs via Adaptive Augmentation and Hierarchical Feature Fusion

Automated prediction of dental conditions in panoramic OPG radiographs is challenging due to class imbalance, rare conditions, and complex anatomy. This study introduces OralHybridNet, a deep learning framework combining hierarchical CNNs (CustomDentalNet) with dual attention mechanisms (OralNetXPlus). Using a new multi-national dataset of 947 annotated OPGs from Pakistan and Thailand, the study applies adaptive augmentation (Elastic Transformations, gamma correction) to balance classes across seven labels (e.g., caries, implants, crowns). A Hybrid Feature Selection algorithm reduces 1,208 features to 300, improving efficiency. OralHybridNet achieves 96% accuracy, 97.6% precision, and 0.99 AUC-ROC, outperforming ResNet50 baselines. It uses spatial and channel-wise attention for early lesion detection and multi-label partitioning to avoid data leakage. The curated dataset is publicly released to support reproducibility. Limitations remain in rare-class generalization and external validation.