Cavit: Early stage dental caries detection from smartphone-image using vision transformer

Caries detection is a routine clinical task in dental practice. If caries are detected at an early stage, non-invasive or micro-invasive treatment such as fillings and a root canal can be effective and thereby invasive treatment and therapies such as gum surgery and dental implants can be avoided. Invasive treatments are expensive and inappropriate for patients with low blood cell counts, cardiac problems and other health issues. Consequently, early caries detection is critical in dentistry. Caries are typically identified through a visual tactile examination in support of radiographic imaging. Fluorescence imaging, cone beam computed tomography or optical coherence tomography are also used. However, these procedures are time-consuming and expensive and require a physical examination of the patient. Moreover, the COVID-19 lessons taught us that such diagnoses should be avoided to prevent contagious diseases. Existing automated caries detection methods fail to achieve sufficient accuracy. Therefore, in this paper, we propose a highly accurate automatic system to detect early caries without any face-to-face interaction with the patient. This system is economical, rapid and easy to use. The proposed system uses a smartphone to capture teeth images and then relies on a vision transformer (ViT) to classify the images as advanced, early or no caries. Finally, the caries are segmented using a U-Net network. The proposed method outperformed the existing methods and achieved a sensitivity of 95%, 91% and 100% for the no caries, early caries and advanced caries classes when tested on a dataset of 300 images, developed for this study.