Diagnosis of Dental Caries in OCT Images Based on Deep Learning

Optical Coherence Tomography (OCT), characterized by its non-invasive nature and high resolution, enables non-destructive, cross-sectional imaging of oral tissues. Deep learning (DL) has prominently contributed to dental medicine, specifically in tasks such as image recognition (IR), classification (CL), segmentation (SG), and quantification (QT), owing to its robust feature learning abilities and outstanding portability. Presently, prevalent clinical techniques employed in the identification of dental caries encompass visual examination and X-ray radiography. Nevertheless, these methodologies fall short in achieving precise identification of incipient enamel caries or visualization of minuscule structural alterations within dental structures. OCT has the capability to detect minute demineralized regions both on the surface and within the internal structure of teeth, thus overcoming the limitations associated with alternative optical detection methods. Initially, we employed a swept-source Optical Coherence Tomography (SS-OCT) system to conduct imaging and preliminary analysis on demineralized dental caries samples. Manual annotation of regions and relevant information associated with dental caries within OCT images to generate a usable dataset. Subsequently, we trained and tested the dataset using three distinct object detection models: YOLOv5, Faster R-CNN, and RetinaNet. After training, the mean average precision achieved was 86%, 86%, and 75% respectively. The experimental results offer potential to expedite diagnosis time for clinicians, thereby serving as a foundation for supplementary diagnostic support by aiding medical professionals in making informed decisions.