Automatic Tooth Segmentation and 3D Reconstruction from Panoramic and Lateral Radiographs

Abstract

The panoramic and lateral radiographs are commonly used in orthodontic dentistry to acquire patient-specific tooth morphology for diagnosing and treatment planning. Considering the variational dentition configurations and image blurs caused by device-specific artefacts and structure overlapping, the robust tooth segmentation and 3D reconstruction from radiographs remain a challenging issue. We propose a deformable exemplar-based conditional random fields (CRF) model for tooth segmentation and 3D shape estimation from the panoramic and lateral radiographs. The shared tooth foreground in the lateral and the panoramic radiographs are utilized for consistent labeling. The 3D deformable exemplars are introduced to provide a regularization of tooth contours to improve the tooth parsing in noisy and ambiguous radiographs. We introduce an alternating optimization scheme to solve the discrete superpixel labels and the continuous deformation of the 3D exemplars simultaneously. Extensive experiments on clinically obtained radiographs demonstrate that the proposed approach is effective and efficient for both tooth segmentation and 3D shape estimation from the panoramic and lateral radiographs.