Purpose: This study aimed to develop a deep learning framework for automated dental caries detection by integrating ConvNeXt with a Feature Pyramid Network (FPN) and Faster Region-based Convolutional Neural Network (Faster R-CNN).
Materials and methods: Two heterogeneous datasets were used: the CariesXrays dataset, containing 6,000 panoramic radiographs, which was divided into training (3,840), validation (960), and test (1,200) sets, and an annotated intraoral image dataset with 6,313 images, divided into training (4,040), validation (1,010), and test (1,263) sets. Grad-CAM-based explainable AI techniques were employed to enhance interpretability. Model performance was evaluated using AP, AP50, AP75, mAP@0.5, precision, and recall.
Results: On the CariesXrays dataset, the model achieved an AP of 72.14%, with AP50 and AP75 values of 97.78% and 86.07%, respectively. On the intraoral dataset, the model achieved an mAP@0.5 of 97.40%, a precision of 97.35%, and a recall of 98.06%.
Conclusion: The proposed framework showed robust performance in dental caries detection across different imaging modalities. Nevertheless, the absence of clinical validation and direct comparison with clinicians highlights the need for multicentric studies and real-world deployment to establish clinical reliability.
See More
See Less
