<![CDATA[Early detection of dental caries presents a significant challenge, particularly in regions with limited access to healthcare services. While many AI models focus on binary classification, real-world applications must handle irrelevant inputs to be robust. This study develops and evaluates a web-based system using a Convolutional Neural Network (CNN) for a three-class dental image classification task: 'Caries', 'No Caries', and 'Not a Tooth'. The method employs transfer learning with the MobileNetV3 Small architecture, trained on a custom augmented dataset of 5,811 images. The model was implemented into an accessible web application using the Flask framework and OpenCV library, supporting both image upload and real-time detection. On the test set, the model achieved an overall accuracy of 93%. It demonstrated exceptional performance in rejecting irrelevant images and high reliability in identifying caries. This study presents a practical and robust tool for initial dental screening, highlighting the importance of a dedicated 'non-target' class for building trustworthy real-world AI applications in tele-dentistry.]]>
Copyrights © 2026
