Multi-Label Classification and Detection of Third Molar Impaction in Panoramic Radiographs
Description
Dental impaction, particularly of mandibular third molars, presents significant diagnostic and surgical challenges due to its association with complications such as inferior alveolar nerve injury, pericoronitis, and odontogenic cyst formation. Existing artificial intelligence (AI)-based diagnostic tools are constrained by binary classification schemes, class imbalance, and computational over- head, limiting their clinical utility. This study introduces a novel hybrid framework that integrates a Genetic Algorithm (GA)-optimized YOLOv10 for real-time detection and employs ResNet50 and InceptionNetV3 deep learning architectures for multi-label classification of third molar impactions. Our contributions include: (1) a lightweight YOLOv11n model enhanced with Multi-Head Self- Attention (MHSA) for precise angular subtype differentiation; (2) generative adversarial network (GAN)-based augmentation achieving a 285× oversampling rate for underrepresented classes such as inverted impactions, addressing dataset imbalance; and (3) standardized performance evaluation using mean Average Precision at IoU thresholds (mAP@0.5:0.95) and Receiver Operating Characteristic Area Under Curve (ROC AUC). The framework was validated on a multinational dataset comprising 1,796 expert-annotated orthopantomograms (κ = 0.92), achieving a classification accuracy of 97.56% with the ResNet50- AdaBoost model. The YOLOv11n detector yielded a mAP@0.5:0.95 of 85.7% while reducing computational complexity by 31% (19.7 GFLOPs vs. 28.4 GFLOPs) compared to conventional architectures. Ablation studies demonstrated that PANet improved mAP by +5.1% for angular subtype detection, while attention modules significantly enhanced model precision. This framework enables fully automated, real-time surgical risk stratification across six angulation-based impaction classes. It demonstrates best diagnostic performance and computational efficiency relative to state-of-the-art models.
Files
Steps to reproduce
Given in the READ ME file.