<![CDATA[Dental caries is a common oral health condition that requires early diagnosis and identification for effective intervention. Existing deep models, such as Faster R-CNN, YOLOv3, SSD, or RetinaNet, exhibit great effectiveness in generic medical imaging; however, they struggle to precisely and explicitly handle localization in complex dental radiographs. In this paper, we propose DCDNet, a convolutional neural network architecture specifically designed for the detection and segmentation of dental caries in oral X-ray images. However, such deep learning methods currently lack strong generalization due to imbalanced training data, limited lesion-localization ability, and noninterpretable features, which hamper their utility for large-scale clinical evaluation. In addition, most models overlook the severity distinction between classes, which is less ideal for the entire diagnosis and treatment planning process. DCDNet was trained and tested on the UFBA UESC Dental Image Dataset, which comprises over 1,500 labeled grayscale dental radiographic images. The proposed network incorporates multiscale feature extraction, residual connections, and non-maximum suppression (NMS) for more accurate classification and bounding box prediction. Data augmentation techniques were used to increase generalization. The model was evaluated based on accuracy, precision, recall, and F1-score, and compared with ResNet50, VGG16, AlexNet, Faster R-CNN, YOLOv3, SSD, and RetinaNet in terms of accuracy. DCDNet achieved excellent performance in all its performance indices, with precision at 97.23%, recall at 97.02%, F1-score at 97.12%, and overall accuracy at 97.61%. Experiments demonstrate that the proposed DCDNet surpasses all the baselines and state-of-the-art methods by a significant margin. Ablation experiments validated the importance of residual connections, NMS, and data augmentation for performance improvement. DCDNet represents a significant step toward automatic dental diagnosis, having successfully detected and localized carious lesions in X-ray images. Its design overcomes the drawbacks of previous models and is a ready option for integration into clinical routine.]]>
