Automatic teeth segmentation of panoramic dental radiographs using multi-frequency processing

Abstract

Panoramic dental radiographs have been widely used for mass examinations due to the fast speed and lower radiation exposure compared to full-mouth intraoral radiographs. Panoramic radiographs provide a reference for the potential position and orientation of all teeth and alveolar bone, thus provide good information for identifying individual teeth or distinguishing regions. Teeth segmentation of panoramic dental radiographs is critical for patients’ treatment planning and treatment, such as implant planning and surgery. Accurate teeth segmentation requires correct boundaries of each tooth and those prerequisite boundaries are manually obtained from dentists or experts. However, this process is time-consuming and exhausting, so fully automatic instance teeth segmentation, which can be done efficiently without labeled ground truth images, is desirable. First, panoramic dental radiographs from three different sources were pre-processed by multi-frequency processing to obtain the image information of each frequency. Next, each image was normalized according to the reference dataset to minimize image discrepancies. Reference panoramic dental radiographs with corresponding manual teeth masks were trained in U-Net to test panoramic dental radiographs from three different sources. The proposed normalization method provided a solution to image texture discrepancies between panoramic dental radiographs from different sources to generate accurate instance teeth segmentation masks. This study can further be used for teeth classification where accurate instance teeth segmentation is also required.