OralHybridNet: A Deep Learning Framework for Multi-Label Classification of Dental Restorations and Prostheses in Panoramic Radiographs

Abstract

Automated prediction of dental conditions in Orthopantomogram (OPG) panoramic radiographs faces significant challenges due to class imbalance, rare pathologies, and complex anatomical structures. This study proposes OralHybridNet, a novel hybrid deep learning framework integrating hierarchical convolutional neural networks, such as CustomDentalNet integrate dual-attention mechanisms and OralNetXPlus. A multinational dataset comprising 2047 clinician-annotated panoramic radiographs spanning 7 diagnostic labels was used. An adaptive augmentation protocol combining Elastic Transformations and gamma correction mitigated class imbalance. A Hybrid Feature Selection (HFS) algorithm condensed 1208-dimensional embeddings into a discriminative 300-feature subset. Evaluated against ResNet50 baselines, OralHybridNet achieved 96.0% accuracy, 97.6% precision, and 0.993 AUC-ROC. The KNN Fine classifier on fused features yielded the highest performance, with real-time capability (9 ms inference time) using a neural network classifier. The proposed framework demonstrates a promising proof-of-concept framework for automated multi-label dental restoration classification.