Segmentation automatique des dents en imagerie maxillo-faciale Cone Beam CT

Abstract

Surgical treatment planning and virtual simulation in fields of dentistry such as orthodontics require personalized three-dimensional models of the patient’s dental anatomy. Such 3D representation, including teeth models, provide an understanding of dentomaxillofacial relationships and support clinicians in selecting and optimizing treatment plans. These models are obtained from 3D imaging modalities. Thus, segmentation remains an essential step in acquiring precise and reliable patient-specific models. However, in this context, purely automated segmentation remains a challenge : the complex morphology of teeth and the lack of contrast accuracy of Cone Beam CT being partially responsible. Therefore, this work aims to develop an automatic individual tooth segmentation method from maxillofacial Cone Beam CT images. The segmentation process is divided in two main steps. First, sub-regions that roughly confine each tooth are extracted from the original 3D volume. Automatic identification of maxillofacial anatomical landmarks, such as dental arch curve and teeth separation planes, helps define the dental sub-volumes. The sub-volumes are then used independently in a pulp detection algorithm based on morphological reconstruction. Second, dental pulp shape acts as a starting point for a gradient-based level-set algorithm that outlines the outer shape on significant planes using the tooth’s main axis. These contours help define the inputs of a random walk segmentation algorithm with texture-based constraints. The random walk based algorithm provides prepersonalized 3D models of each tooth. The surface meshes are then iteratively refined using laplacian mesh optimization, resulting in precise and reliable 3D models of the patient’s dental anatomy. The quality of the automatic segmentation is quantitatively assessed using a database of 88 manual segmentations. The method’s performance can be globally summarized with the following metrics : a Dice index (DICE) of 95.20±1.07 %, a relative volume difference (RVD) of 2.57±3.21 % and an average symmetrical-surface distance of 0.16±0.04 mm. The results of this study suggest that our method can provide multi-organs segmentations both accurately and automatically from a patient’s maxillofacial Cone Beam CT scan.

Date	31 May 2019
Original language	French
Awarding Institution	École de technologie supérieure
Supervisor	Jacques A. de Guise (Supervisor) & Matthieu Schmittbuhl (Co-supervisor)