Dental caries, a prevalent disease with significant health and economic consequences, goes undiagnosed during the early phases of its progression since conventional diagnostic methods like visual inspection and radiography possess low sensitivity as well as inter-observer consensus. This review discusses the use of deep learning (DL) for the automatic detection and grading of caries, comparing systematically different imaging modalities, such as bitewing and periapical radiography, intraoral photography, optical coherence tomography (OCT), cone-beam computed tomography (CBCT), and laser fluorescence, and their implications in caries diagnosis. It emphasizes how DL models, especially convolutional neural networks (CNNs), Transformers, and U-Net architectures, perform well in classification, detection, and segmentation tasks with expert-level performance and quantitation of lesions. They facilitate diverse clinical applications such as tele dentistry and personalized treatment planning and are advancing with multimodal data fusion, explainable AI, and real-time processing. However, there are still challenges regarding limited annotated datasets, model generalizability, computational requirements, and clinical interpretability. The review aims to promote clinical translation by summarizing recent advances, comparing methodologies, and pointing out future directions for intelligent oral healthcare.
Kejun Liu (Mon,) studied this question.