Hybrid-ResNet-C-Deep: An Explainable Framework for Intelligent Dental Caries Localization and Diagnosis Decision

Dental caries is one of the most prevalent oral diseases in the world. The early detection of dental caries is essential to avoid tooth loss and the complications caused by the disease. Panoramic dental radiographs are routinely used for dental caries screening, but detecting subtle signs of dental caries on these radiographs can be challenging. To improve the detection of dental caries on panoramic dental radiographs, a deep learning framework named Hybrid-ResNet-C-Deep is proposed. This deep learning framework employs residual blocks with attention mechanisms to enhance the localization of dental caries. Additionally, the framework uses Grad-CAM to provide clinicians with explainable predictions from the model. The model was trained on 4,000 panoramic dental radiographs from the OdontoAI database and additional dental caries images with annotations from dental clinicians. The model was evaluated using cross-validation and stratified sampling of the database. The proposed framework achieved an accuracy of 95.0%, a Dice score of 90.1%, an Intersection over Union score of 84.3%, a sensitivity of 93.5%, and a specificity of 96.2% on the test data. The deep learning framework can be utilized as an intelligent decision support system for dental clinicians to improve the diagnosis of dental caries. Further validation from different dental clinics is required before the framework can be deployed in dental clinics for dental caries diagnosis.