A Dual-Stream Deep Learning Framework for Multi-Label Classification of Dental Restorations in Panoramic Radiographs with Explainable AI Integration

Panoramic dental radiographs provide comprehensive views of dentition, but manual interpretation is time-consuming and prone to variability. Conventional convolutional neural networks often fail to detect fine features and rare pathologies, limiting clinical use. This study introduces a hybrid dual-stream deep learning framework combining Dense Pyramid-Net and Multi-Scale Hierarchical Attention Network for enhanced feature extraction. Using a curated panoramic dataset and an external public dataset (augmented and stratified 80:10:10), the framework applies ML-SPFS for optimized feature selection, followed by serial feature fusion and multi-label classification of seven dental conditions. Explainable AI was incorporated to ensure interpretability. The proposed method achieved 97.4% accuracy, outperforming state-of-the-art approaches. External validation confirmed generalization, while ablation studies highlighted the role of adaptive augmentation, hierarchical fusion, and optimized feature selection. The model also demonstrated high precision for rare classes and reduced inference time, supporting real-time clinical use. This work represents the first integration of MSHA-Net and Dense Pyramid-Net with ML-SPFS, providing a scalable AI framework that improves diagnostic reliability and supports evidence-based treatment planning in dentistry.