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

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).在全景曲面体层X光片（Orthopantomogram, OPG）中自动检测牙科修复体与修复治疗方案仍颇具挑战，这是由于数据集存在严重的类别不平衡问题、罕见病理征象的干扰，以及牙科影像固有的复杂解剖结构。传统深度学习模型往往难以在异质性临床数据集上实现良好泛化，尤其是当单张X光片中同时存在多种修复体类型时。本研究旨在开发并评估OralHybridNet——一款专为提升全景X光片中牙科修复体与修复装置多标签分类性能而设计的混合深度学习框架。OralHybridNet整合了层级卷积神经网络架构，将CustomDentalNet与双注意力机制以及OralNetXPlus特征提取模块相结合。本研究构建了一个跨国数据集，包含2047张由临床医师标注的全景X光片，涵盖7种代表常见牙科修复体与修复结构的诊断标签。为解决类别不平衡问题并提升模型泛化能力，本研究采用了融合弹性变换与伽马校正的自适应数据增强流程。网络生成的特征嵌入还通过混合特征选择（Hybrid Feature Selection, HFS）算法进行了进一步优化，将1208维的高维特征表示精简为300个具备区分度的最优特征子集，用于后续分类任务。所提出的框架相较传统深度学习基准模型展现出更优的性能。相较于ResNet50，OralHybridNet的整体准确率达96.0%，精确率为97.6%，受试者工作特征曲线下面积（AUC-ROC）为0.993。在所有评估的分类器中，应用于融合特征嵌入的K近邻（K-Nearest Neighbor, KNN）精细分类器取得了最高的预测性能。此外，该框架还展现出高效的计算性能，具备实时推理能力（单张图像推理耗时约9毫秒）。