This work addresses the problem of teeth segmentation in panoramic dental images, specifically the detection of the adult-stage mandibular teeth. Random Forest Regression-Voting Constrained Local Models (RFRV-CLM) were used to train individual teeth shape models. In order to fully automatically initialise the teeth shape search process, RFRV-CLM model was trained from a set of mandible and teeth key landmark points, and the predicted keypoints were used to estimate the initial pose of each tooth shape. Furthermore, a method to detect present/missing teeth has been proposed, based on the quality of each tooth shape segmentation. The results of this two-step approach, evaluated using a set of 346 annotated images, show good performance in present/missing teeth detection and state-of-the-art accuracy for both correct teeth localisation and shape segmentation, with an average median point-to-curve error of 0.2mm.
Keywords: teeth segmentation, panoramic dental images, random forest regression-voting, machine learning