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

Description

Automated detection of dental restorations and prosthetic treatments in Orthopantomogram (OPG) panoramic radiographs remains challenging due to substantial class imbalance, the presence of rare pathological findings, and the complex anatomical structures inherent in dental imaging. Conventional deep learning models often struggle to generalize across heterogeneous clinical datasets, particularly when multiple restoration types coexist within a single radiograph. This study aimed to develop and evaluate OralHybridNet, a hybrid deep learning framework designed to enhance multi-label classification performance for dental restorations and prostheses in panoramic radiographs. OralHybridNet integrates hierarchical convolutional neural network architectures, combining CustomDentalNet with dual-attention mechanisms and the OralNetXPlus feature extraction module. A multinational dataset consisting of 2,047 clinician-annotated panoramic radiographs was compiled, covering seven diagnostic labels representing common dental restorations and prosthetic structures. To address class imbalance and improve model generalization, an adaptive augmentation pipeline incorporating elastic transformations and gamma correction was implemented. Feature embeddings generated by the network were further refined using a Hybrid Feature Selection (HFS) algorithm, which reduced a high-dimensional 1,208-feature representation to an optimized subset of 300 discriminative features for downstream classification. The proposed framework demonstrated superior performance compared with conventional deep learning baselines. Relative to ResNet50, OralHybridNet achieved an overall accuracy of 96.0%, precision of 97.6%, and an AUC-ROC of 0.993. Among evaluated classifiers, the K-Nearest Neighbor (KNN) Fine classifier applied to fused feature embeddings produced the highest predictive performance. Additionally, the framework demonstrated efficient computational performance with real-time inference capability (~9 ms per image).

Steps to reproduce

Given in the READ ME file.

Institutions

• King Faisal University

  Eastern, Al-Hasa
• Chulalongkorn University

  Bangkok, Bangkok

Categories

Licence