Privacy-aware and interpretable deep learning framework for dental caries classification

Abstract

Dental caries remains one of the most prevalent and persistent chronic diseases globally, affecting individuals across all age groups and posing a significant burden on public health systems. Early detection is critical to prevent the progression of tooth decay, reduce treatment complexity, and improve long-term oral health outcomes. In response to these clinical demands, this study presents a comprehensive, privacy-aware, and interpretable deep learning framework for the automated classification of dental caries from X-ray images. The approach addresses the issues of class imbalance, low Resolution image and privacy preserved patient's medical images.The framework is structured into three progressive phases that incorporate supervised learning through Convolutional Neural Networks (CNN), ResNet-18, and DenseNet; unsupervised clustering using Principal Component Analysis (PCA); and a decentralized federated learning strategy to ensure secure model training across distributed datasets. The experimental dataset consists of 957 labelled dental radiographs, including 174 healthy and 783 carious cases, emphasizing the issue of class imbalance. Initial baseline models achieved an accuracy of 84 %, which improved to 96 % following strategic data augmentation and class balancing interventions. PCA-based clustering visualizations revealed well-separated clusters (Silhouette Score: 0.6660), confirming the discriminative power of the selected features. Meanwhile, the federated learning implementation preserved data confidentiality without sacrificing performance, reinforcing the model's suitability for real-world clinical deployment. Collectively, these findings validate the framework's robustness, interpretability, and adaptability, offering a scalable and ethically aligned solution for AI-driven dental diagnostics in modern healthcare systems.