Segmentation of the Mandibular Canal in Cone-Beam CT Data

Abstract

Accurate information about the location of the mandibular canal is essential in case of dental implant surgery. The goal of our research is to find an automatic method which can segment the mandibular canal in Cone-beam CT (CBCT). Mandibular canal segmentation methods in literature using a priori shape information are, the 2D active appearance model of Rueda et al., and 3D active shape model (ASM) of Kainmueller et al. The mean distance to manual annotation of the mandibular canal of the method of Kainmueller is around 1.1mm. The best method in literature is Kim et al. with an average distance of 0.7mm. We develop and evaluate five methods for mandibular canal localization. The methods, Lukas Kanade tracking (LK), B-spline registration, demon registration, 3D active shape model (ASM), and active appearance model (AAM). The ASM and AAM need corresponding points between the mandibles in the training data. We develop and evaluate two methods to find corresponding points, minimum description length (MDL) and the second shape context (SC) based registration. To improve the quality of the CBCT scans we introduce a rotational invariant edge preserving optimized anisotropic diffusion filter. We evaluate the performance on 13 CBCT scans. The registration methods have an average distance to expert annotation of the canal of more than 4mm, LK tracking a distance of 3mm, AAM and ASM a distance of respectively 2.0mm and 2.3mm. The MDL method does not improve point correspondences found by the SC method, and the pre-filtering with the introduced diffusion filter does not improve the ASM result. By using location based intensity weights we improve the AAM results, to an average distance of 1.88mm. The relatively large error is due to a low number of training datasets, and low CBCT scan quality.