Integrating Graph Neural Networks and Vision Transformers for Oral Health Diagnostics in Predicting Overhanging Restoration Classification from IOPA Images

Abstract

Introduction: Dental restoration quality assessment through radiographic imaging is crucial for evaluating the long-term success of dental treatments. Intraoral periapical radiographs (IOPA) are diagnostic tools that provide detailed information about dental restorations, surrounding tissues, and potential complications. Predicting overhanging tooth restorations is crucial in dentistry, especially from a periodontal perspective, as it can lead to plaque accumulation, periodontal health issues, functional issues, and aesthetic concerns. Hybrid graph and transformer architecture fusion approaches detect dental overhanging restorations, capturing spatial relationships and topology. This study proposes a novel GNN-Transformer architecture for dental overhanging restoration classification tasks. The study introduces a novel method that integrates Graph Neural Networks and Transformer architectures to enhance the precision of dental overhanging restoration classification.
Methods: The study uses 50 IOPA images from various online databases to analyze dental radiography. The images are segmented and annotated for overhanging restorations and classified as normal or overhanging. The data is split into training and test data, with 80 percent training data and 20 percent test data subjected to deep learning architecture. The model uses advanced image preprocessing techniques like the CLAHE method to enhance dental features' visibility while managing noise levels. The Graph Neural Network (GNN) architecture is used to refine and enrich feature representations, with three convolutional layers and a transformer architecture to improve the model's ability to understand intricate patterns. The classification approach integrates advanced pooling techniques with neural network layers, generating a holistic representation of dental anatomy. The architecture employs dropout techniques to mitigate overfitting and ensure strong generalization to novel instances.
Results: The study assessed the effectiveness of hybrid graph transformers in predicting overhanging restorations. The model correctly identified 71% of normal cases, balancing precision and recall. However, it missed 67% of true overhanging cases. The F1-score was 0.80, indicating its potential for dental imaging analysis.
Conclusion: The study explores using hybrid graph transformers to detect overhanging restorations in dental images, showing promising accuracy and precision. However, limitations like class imbalance and overfitting need to be addressed. Future improvements include expanding the dataset, refining annotation practices, and integrating clinical features. This could improve patient outcomes and treatment planning, necessitating ongoing refinement and validation.