DCDNet: A Deep Learning Framework for Automated Detection and Localization of Dental Caries Using Oral Imagery

Abstract

Dental caries is known to be one of the most common diseases seen, and it also represents a public health problem worldwide that needs timely and accurate diagnosis for successful treatment. Faster R-CNN, YOLOv3, SSD, and RetinaNet are some of the prevalent deep learning models that have demonstrated significant success in varied fields of medical image analysis; however, they mainly encounter difficulty when balancing the accuracy of classification and localization, which is crucial, mostly when analyzing subtle patterns in dental radiographs. Such models consistently suffer from inconsistent feature extraction, overlapping bounding boxes, and poorly generalizing between different datasets. Acknowledging these shortcomings, this work presents DCDNet, an innovative deep-learning architecture capable of discovering dental caries using the UFBA UESC Dental Image Dataset. The DCDNet features multi-scale feature extraction to generate more salient features and residual connections to improve the backpropagation, and it incorporates NMS to further reduce the repeated bounding box predictions. Data augmentations are also applied extensively to improve the model's generalization ability. Experimental results show that DCDNet achieves state-of-the-art results of 97.23% precision, 97.02% recall, 97.12% F1 score, and accuracy of 97.61%. The framework should significantly reduce the rate of false positives and false negatives in the process, increasing reliability in clinical scientific use. The superior performance of DCDNet makes it a potentially powerful automated dental diagnostic tool, facilitating early detection and more efficient treatment planning. These findings promote the development of AI-based solutions for dental radiography analysis and serve as a baseline for future studies.