A Dual-Stream Deep Learning Framework for Multi-Label Classification of Dental Conditions in Panoramic Radiographs With Explainable AI Integration.
Authors
Affiliations (4)
Affiliations (4)
- Department of Prosthodontics and Dental Implantology, College of Dentistry, King Faisal University, Al Ahsa, Saudi Arabia. Electronic address: [email protected].
- Department of Prosthodontics and Dental Implantology, College of Dentistry, King Faisal University, Al Ahsa, Saudi Arabia.
- Center of Excellence in Precision Medicine and Digital Health; Geriatric Dentistry and Special Patient Care International Program, Department of Physiology, Faculty of Dentistry, Chulalongkorn University, Bangkok, Thailand.
- Dental Public Health Department, College of Dentistry, King Abdulaziz University, Jeddah, Saudi Arabia.
Abstract
Panoramic radiographs are routinely used in dental practise for diagnosis, treatment planning and follow-up assessment. However, manual interpretation of multiple dental restorations is time-consuming and subject to inter-observer variability, highlighting the need for reliable artificial intelligence (AI)-assisted diagnostic systems. To develop and validate a clinically interpretable deep learning framework for automated multi-label classification of dental restorations in panoramic radiographs. A novel multi-label dental panoramic radiograph dataset comprising 1247 clinician-annotated orthopantomograms (OPGs) was developed for model training and internal validation. Cross-dataset evaluation was performed using an independent dataset containing 11,500 panoramic radiographs. Following stratified image level dataset partitioning (80:10:10), data augmentation was applied to the training sets to prevent data leakage. The proposed framework integrates a Multi-Scale Hierarchical Attention Network (MSHA-Net) and Dense Pyramid Net for complementary feature extraction. Optimised feature fusion was achieved using a Multi-Label Sparsity-Promoting Feature Selection (ML-SPFS) strategy, followed by multi-label classification of 7 dental restoration categories. The proposed framework achieved an accuracy of 97.40%, a Macro F1-score of 99.40% and a Hamming Loss of 0.006 on the internal test set. During cross-dataset evaluation, it maintained strong generalisability, achieving 93.90% accuracy and a Macro F1-score of 98.70%. Ablation analysis demonstrated the superiority of the fused feature representation over the individual network architectures, while statistical analyses confirmed significant performance improvements compared with baseline models (P < .001). The proposed dual-stream framework enables accurate and interpretable multi-label classification of dental restorations from panoramic radiographs and demonstrates robust generalisation across independent datasets. These findings support its potential as an AI-assisted clinical decision-support tool for dental radiographic interpretation, although prospective multi-centre validation across diverse imaging systems is warranted before routine clinical deployment.